In February I posted a simple one... Split AI users into tiers. A small group of super users consume most tokens. Ask how much more each tier consumes once it runs agents, and how many adopt and multiply.

It said total token demand would reach 1.9x to 2.9x within 12 months. Every disclosed series I can find passed the top of that range in 3-7 months:

• Anthropic run-rate revenue: $14B in February to a reported $65B by end of July. That is 4.6x in 5.5 months.

• OpenRouter weekly tokens: about 7 trillion to 56 trillion from January to July. That is 8x in 7 months.

• Google tokens, all surfaces: 480 trillion to 3.2 quadrillion a month. That is 6.7x in 12 months.

Subscribe now

So structure was right and the numbers too low. The super users’ share was bigger than I assumed, and I left out a tier entirely.

Two lessons... framework is worth using again, and when I have been wrong on the inputs, I have been wrong by being too cautious. There are some who are bearish AI because they see it as worthless, while others are concerned about timing of demand/ROI. If you see it as worthless, then you should stop reading here. If you fall in the latter camp, then what you really should care about is the timing of agentic adoption. If you have better inputs or anecdotes etc, please share!

Same framework

The tier I left out was the agent itself. On OpenRouter, agents passed humans in token use on Feb 6, four days before my post. Agent tokens then rose 14x in 6 months and human tokens rose 2.8x.

Every AI demand forecast so far has carried a hidden cap: human attention. A person types, reads, thinks and types again. Jellyfish’s developer data shows 84% of users top out at one or two agents at a time.

Agents remove that cap. Google’s AI infrastructure chief, Amin Vahdat, said this month that with agents the gap between model calls goes from seconds toward milliseconds. Meta’s Muse and OpenAI’s Dots, both launched in September, give each agent its own cloud computer.

So I rebuilt the February table with a Tier 0 on top - the always-on agent.

Here is what the framework says if adoption follows the table:

• At the 12-month adoption rates: token demand of 2.2x-3.5x today.

• At the 24-month rates: 4.5x to 12x.

• Tier 0 share of all tokens: from roughly 15% today to more than half.

I hold the scale with far more confidence than the dates. The table has a clock on it because a model needs one. The clock is the least reliable part. The 15% starting share is also my assumption, not a disclosed number.

For reference, Goldman Sachs forecasts 24x by 2030, which is 2.2x a year. That is my low case. Google’s actual trailing pace is 6.7x.

What has to happen

This is not a claim about what is happening today. It is the list of things that have to happen for the high end of that table to show up.

• Agents need their own computers at scale... Muse passed 5 million US downloads in 22 days, ChatGPT took 56, Similarweb put it at about 2.2 million daily users on Oct 5, and it is still only in the US and Canada.

• They have to run longer... Today’s sizing assumes an agent is active about two hours a day. METR tracks how long a task a frontier model can finish alone, and the latest reading is about 17 hours, doubling roughly every four months. But, METR also warns its readings above 16 hours are unreliable, so treat that as direction.

• They have to run side by side... Most users still stop at one or two. Three agents for eight hours, instead of one for two, is twelve times the tokens.

• Access has to widen... Andrej Karpathy’s napkin math this week showed ~6 billion people have barely used these tools, 1-2 billion use free tiers casually, and about 20 million use frontier models professionally. Today’s demand curve is being drawn by a sliver of the population.

• Someone has to pay... Jefferies’ case of a billion Muse users produces about $10.8 billion of annualized revenue. The infrastructure to serve them runs to hundreds of billions. Ads, commissions or paid tiers have to close that gap, or usage gets rationed.

If these happen, the scale is not subtle. Analyst Freda Duan estimates always-on agents need 1 to 2 gigawatts per 100 million daily users, and 3 to 4 if agents make more model calls. Half a billion daily users across Meta, Google, OpenAI and xAI would be 5 to 20 gigawatts. TrendForce’s figures imply roughly 30 gigawatts of AI capacity added in all of 2027.

Big question is... when

I do not know when this inflects. Anyone who names the quarter is guessing. The reasons timing is hard are real:

• Rollouts are gated by supply... Meta’s AI chief said early Muse users consumed about 10x what internal test groups did. Muse has usage caps, and available compute is setting the pace, not demand.

• Retention is unproven... Sensor Tower has 73% of new Muse users returning the next day, almost twice ChatGPT’s rate. But, it then loses users quickly, and ChatGPT pulls ahead by day eight or nine. Close to half of Muse’s early impressions came from Meta’s own ads.

• Enterprises cannot yet measure the return... In Morgan Stanley’s latest CIO survey, 42% said they cannot assess the return on their token spend. About 13% see returns above their AI spend, but 28% see returns below.

• The product still breaks... Some Muse tasks fall back to human assistants, and Amazon has blocked Muse from shopping on its site. Roadblocks like this need to be resolved, and likely will.

• Hardware slips too... One unconfirmed report this week had expectations for Nvidia’s next-generation Rubin shipments cut roughly in half on yield and capacity limits.

Adoption curves look flat until they do not. February to August showed this can move faster than a sober model says. It can also stall for two quarters.

That is why I think every AI observer should be maniacally focused on two things... the timing and the scale of agentic adoption. It is the swing variable in every demand model, bull or bear.

The hiccups that will come

A multi-quarter view has to survive bad headlines. Expect them, especially on lab revenue and margins. The last two weeks already had two.

• OpenAI... On Oct 8 the FT reported that OpenAI told investors its annualized revenue is approaching $50 billion, not the $70 billion that had circulated. The gap was a definition as Anthropic counts sales made through cloud partners and OpenAI does not, but it still read as a miss.

• Anthropic... Its leaked draft prospectus showed a $42 billion net loss for 2025 on $4.6 billion of revenue, and $518 billion of planned spending on cloud services and data centers. The same document showed $11.5 billion of revenue in Q2 2026 alone and a second straight quarter of operating profit.

• Pricing... Frontier-class models keep getting cheaper. Meta’s new model is priced at $4 per million input tokens, against $10 for OpenAI’s flagship. Cheaper tokens squeeze lab margins and expand units at the same time.

A lab can miss on margin while it serves more tokens than ever. Margin decides who captures the value, and units decide how much hardware gets used.

I expect lab margin misses which will move stocks, but this is not the full picture and likely underplays the volume piece.

Why that still gives me comfort on infrastructure

I am not claiming precision on any one quarter. I am claiming that over several quarters the direction of unit demand is clear enough that a lab stumble does not break the infrastructure case over time.

• The supply chain is still strained... TSMC’s Q3 revenue was up ~50% YoY and above the top of its guidance.

• Buyers are raising plans... JPMorgan now models capex at the four largest US clouds up 85% in 2026 and 70% in 2027. The 2027 figure was 50% before.

• Existing capacity is full... Vahdat said Google’s seven and eight year old chips still run at 100% utilization, and that serving capacity has to double about every six months.

• Agents change what is inside each gigawatt... Estimates of CPU cores per gigawatt under agent workloads run from a doubling (Dolphin Research) to four times (Arm). Morgan Stanley puts agent-driven DRAM demand at 15 to 45 exabytes by 2030, equal to 26% to 77% of 2027 supply.

• The load lands on the rest of the internet too... Dolphin notes one agent price comparison task can trigger up to 146 page loads. Each one hits someone else’s servers.

One thing to keep straight. The agent’s own sandbox is small. Duan’s estimate for 100m daily users is about 65,000 CPUs and a tenth of a gigawatt. Most of the power is still inference. The hardware story is the shifting mix inside each gigawatt, not the sandboxes.

The risk on this side is financing, not demand. JPMorgan has cloud capex at 111% of operating cash flow in 2027. The buildout now leans on debt markets, and the cost of that debt is rising.

Pushback

• “These are junk tokens”... More than 85% of agent tokens on OpenRouter are cached and billed at about a tenth of a fresh one. But cached context has to sit in memory for as long as the agent runs, and a bargain model still occupies a chip. Lower grade models running more agents is not a hole in the thesis, it is the thesis.

• “This is bad for the labs”... On margins, yes. See the hiccups above. But whoever serves the tokens buys the compute still. The infrastructure case does not need the labs to capture the value.

• “The ROI is not there...” Jellyfish found the first million tokens per developer buy about 11 merged pull requests and later millions buy almost none. Uber used its 2026 AI coding budget in four months. I read a blown budget as demand outrunning the plan. A world of hard caps looks like my low case, not zero.

• “Supply and efficiency will catch up”... This is the strongest one. Reported capacity growth of about 1.6x a year, times cost per token falling 60% to 70%, can absorb 4x to 5x more tokens a year. My range is 2.2x to 3.5x. So at my low case accelerators could loosen. Against that, Google’s actual growth was 6.7x.

• “People will not show up”... They might be too slow, yes. That is the timing risk above. But the table needs little: 10% of casual users changing behavior. Goldman’s entire 24x rests on 12% of knowledge workers using agents by 2030.

What I am watching

If timing and scale of agent adoption are the variables, these are the gauges:

• Muse... whether usage caps loosen or tighten, when it opens International through WhatsApp, and daily users once the ad push fades.

• Dots... what OpenAI charges after the free first month, and when it sells a second agent per plan.

• Agent share of tokens... OpenRouter’s agent versus human split.

• Google’s next token disclosure... Alphabet reports on Oct 28.

• METR’s next task-length reading.

• Lab gross margins... starting with Anthropic.

• Server CPU and memory. AMD and Intel on 2027 allocation and pricing, and memory contract prices.

• Meta’s capex guide in January.

The charts are illustrative and built with Claude’s help, as they were in February. I am a generalist connecting dots, not an engineer. Better inputs are welcome!

NFA. DYODD. Nothing here is investment advice.

Sources

• Google I/O 2026: tokens processed

• The Decoder: agent vs human tokens on OpenRouter

• Tom’s Hardware: cached share of agent tokens

• ARK on OpenRouter weekly tokens

• Jellyfish, State of AI in Software Engineering H1 2026

• Anthropic run-rate reporting

• Goldman Sachs Research: AI agents and token consumption

• Amin Vahdat on Training Data, Oct 6 (Teahose summary)

• Muse downloads

• Muse retention and daily users, Yahoo Finance

• Muse capacity limits

• Freda Duan: Muse infrastructure estimate

• Dolphin Research via TradingKey

• TrendForce: data center power demand

• Morgan Stanley on agentic CPU and DRAM

• Reuters: Wall Street on Muse

• METR time horizon status

• Karpathy’s funnel

• TechCrunch: OpenAI revenue correction

• Fortune: Anthropic draft prospectus

• TSMC Q3 revenue

• StreetSignal, Oct 13: JPMorgan capex, Morgan Stanley CIO survey, model pricing, Rubin report

• Server CPU demand and the CPU to GPU ratio, via BigGo

• Morgan Stanley Sep 4 report, as summarized by TechFlow

• Uber and Microsoft on AI coding costs

• Financing costs and AI capital, Edge and Odds

DISCLOSURE

Position: At the time of publication, the author holds a positions in GOOGL.

Trading Policy: The author will not materially alter this position within 48 hours of publication. After this period, the author may buy, sell, or otherwise adjust the position without further notice. Changes to the author’s view or position will be reflected in subsequent publications when material.

Conflicts: The author has received no compensation from the issuer or any party with a financial interest in this security.

Forward-Looking Statements: This report contains the author’s opinions, estimates, and projections, including price targets derived from financial models. These are forward-looking statements subject to substantial uncertainty. If these assumptions prove incorrect, the actual value may differ materially, including scenarios of significant loss or total impairment. The price target represents the author’s estimate of fair value under the stated assumptions, not a prediction of where the stock will trade.

This report is provided for informational purposes only and does not constitute investment advice. See the full Terms & Disclosures for additional important information.

Bottom line…

• Pricing is still rising... 5yr market rate is now ~$19bn per GW-year (and is our GOOGL base case), short-term capacity clears at ~$40bn with two witnesses, and NBIS lifts list prices 17-21% from Oct 1st

• Market discriminating... big lease and capacity announcements that arrive with new debt or equity now fade or fall, pricing news without a raise still rallies, 9/17 the market graded both on the same morning

• Money got more expensive for the levered players… CRWV’s new convertible costs about 10.5-11.7% once the equity option is valued, up ~170bp from its April deal

• Supply got tighter again... Oracle’s New Mexico Stargate campus now carries a force majeure notice and a reported one-year slip, and Texas has frozen data-center permits pending a grid audit

• Thesis intact - owns integration and NBIS…

  Subscribe now

Compute keeps getting more expensive

Revenue per gigawatt-year by deal... bold green rows are new since our Google note…

• The market five-year rate now sits on our base case... SemiAnalysis puts five-year GB300 contracts at $4.50 to 4.60 per GPU-hour, about $19bn per GW-year, against our $18bn base case for uncontracted capacity

• Short-term pricing has two new marks… CoreWeave’s 3-6 month contracts at ~$40bn per GW-year (Sep 17) match Nebius’s short-term tier of $40-50bn

• Top of the market is SpaceX... Anthropic pays $1.25bn a month for Colossus 1, roughly $45 to 50bn per GW-year on 300MW+… that is immediacy pricing on sunk capex, with a 90-day termination right after December

• Nebius has set its 2027 baseline… 4 mid-term deals (Reflection, Cohere, a US lab, a quant firm) average over $1bn each at $20-25bn, with 50 to 60% of capex prepaid and payback of 1 year 10 months... management says it could sell all of its 2027 capacity on those terms… IREN matches with >$20bn signed, ~$25bn in talks, 45 to 55% prepaid

• Nvidia’s floor is half the market... AICP backstop at Firmus works out to ~$9.8bn, about where 2024-25 contracts were struck... that is how far pricing would have to fall before Nvidia’s guarantee pays

• Shell rent barely moved… $1.9-2.4bn on 2026 leases, versus $1.2-2.3bn before… All of the price action is at the operator layer

• List prices too... Nebius raises on-demand prices 17-21% from 1 October (B300 $7.85 to $9.50 per hour), its second increase in about three months

Market stopped paying for megawatts

Why the change… a 20yr lease at $2.4M per MW, funded at 80 to 90% loan-to-cost, is now read as a financing event, not a growth event. The equity is a thin slice on top of project debt, and that debt got dearer.

• Cost of paper is rising… see CoreWeave’s April convertible details above, IREN’s new GPU debt is 9.0% fixed.

• Rates moved against them... Fed hike and a 10-year 5.25% hit the most levered names hardest

• Nebius is not immune... fell 10% on its own $5bn convertible on Aug 19, Nebius’s news is more often pricing, and its prepaid deal structure means less of its growth has to be funded with paper

• Hyperscalers split the same way...

  • Alphabet was up 1.3% in September, moved by product competition not capex, TPUs are now bankable collateral with a reported $22bn bank loan secured by TPUs and customer contracts for Crux AI

    • Better yet, TPU costs reportedly coming down 43% via switch to MediaTek (per JPM)

      

    • Oracle, the debt-funded, customer-concentrated hyperscaler, fell 8% month to date against the S&P’s +0.6%

• Nscale’s IPO filing shows $103bn of contract value, about 85% of it from Microsoft and Anthropic

NM, TX and grid got tighter again

Oracle and OpenAI’s Project Jupiter in Doña Ana County was meant to be ~2.2 GW of 2028 capacity. The power plan is now ~2.45 GW of Bloom fuel cells running on gas (turbines were dropped in April), so one 17.8-mile pipeline is the critical path, and a one-mile crossing of state land is the choke point.

Resolution… The realistic unlock is political, not legal. The commissioner who denied the crossing twice leaves office… her successor is chosen on Nov 3 and is likely seated in January. If the new commissioner approves, and the air permit and FERC certificate land on time, gas could flow from February 2027 and first power follows later in 2027 at the earliest. The base case, consistent with Oracle’s force majeure notice and the reported one-year slip, is meaningful capacity in 2029. No party has given a first-energization date.

• Texas froze too… Gov. Abbott paused ERCOT interconnection for data centers on Aug 3 pending an audit, then on Sep 21 ordered TCEQ to halt all data-center permits until it finishes (expected around December). ERCOT’s large-load study for 66 GW of base load now reads “TBD.” Whether sites already under construction are affected is unclear

• PJM cleared its 2028/29 auction 6.8 GW short, at the cap. An emergency backstop auction runs Sep 30 to Oct 21

• Equipment… Siemens Energy’s gas-turbine backlog is ~69 GW, with lead times of three years or more

Each of these pushes energized supply to the right. That widens the shortage window in the main note’s Fig 5 and raises the value of megawatts that are already powered and permitted.

Putting it together…

• Structure unchanged, window widened… Jupiter, Texas and PJM all extend scarcity into 2028 and plausibly 2029

• Market picking through winners more closely… reward pricing power, prepayment and owned silicon, avoid volume funded with converts, ATMs and 9% project debt

• Powered Texas megawatts got scarcer

• Oracle is now a credit watch, not only a capacity story… BBB-, a construction loan at ~90 cents, and a force majeure notice, slip is capacity the rest of the market no longer has to compete with in 2028

Dates to watch

• 5 Oct: FERC comments close on the Green Chile pipeline

• 19 Oct: TCEQ status update on the Texas permit freeze

• 21 Oct: PJM backstop auction bids close; results early December

• Late Oct: hyperscaler Q3 prints and first 2027 capex guides

• 3 Nov: New Mexico Land Commissioner election

• Early Nov: Nebius, CoreWeave, IREN and the landlords report Q3

• 23 Nov / 3 Dec: NMED air-permit deadline / FERC certificate deadline

• 1 Jan 2027: Colossus termination windows open (Anthropic, Google)

• 1 Feb 2027: earliest gas to Project Jupiter

DISCLOSURE

Position: At the time of publication, the author holds a positions in GOOGL and NBIS.

Trading Policy: The author will not materially alter this position within 48 hours of publication. After this period, the author may buy, sell, or otherwise adjust the position without further notice. Changes to the author’s view or position will be reflected in subsequent publications when material.

Conflicts: The author has received no compensation from the issuer or any party with a financial interest in this security.

Forward-Looking Statements: This report contains the author’s opinions, estimates, and projections, including price targets derived from financial models. These are forward-looking statements subject to substantial uncertainty. If these assumptions prove incorrect, the actual value may differ materially, including scenarios of significant loss or total impairment. The price target represents the author’s estimate of fair value under the stated assumptions, not a prediction of where the stock will trade.

This report is provided for informational purposes only and does not constitute investment advice. See the full Terms & Disclosures for additional important information.

Personal note, experience informing opportunity for STLN

A few years ago, I spent a year as the CFO of a large direct primary care (DPC) business. Our headcount crossed the line where an employer must legally offers health insurance or pays a penalty. So we went to market with a broker, and the plan we could afford came through a TPA we already knew. Imho, there are a lot of lessons to draw from understanding both the DPC model and growing ecosystem (e.g. Aligned Marketplace) and the small but meaningful innovations that can grow in the insurance models when someone simply tries to align incentives.

If anyone has domain expertise in this space, please reach out! Would love to discuss further as I’m getting deeper into the oncology ecosystem which has massively entrenched players, and for good reason… they make so much money in the space.

On the TPA front, the partner we chose was particularly adept at finding ways to reduce costs for an employer in order to be more price competitive on plan rates and win more business. Seems obvious, but that’s actually the opposite of how most operate. This TPA had built its own medical equipment business. It had direct contracts with specialty pharmacies at better prices. It went after the most expensive slices of a health plan one at a time, so it could quote a price the big insurers could not, and grow on that price.

• TPA, third-party administrator: runs a self-funded plan’s claims, network and member service for a monthly fee, carries no risk, the employer pays the claims with stop-loss insurance behind it

• Who has to offer coverage: employers with 50+ full-time staff, part-timers counted as equivalents, or $3,340 per worker in 2026, 67% of covered US workers sit in self-funded plans, 80% at large firms

• Pain-Point analogy: the idea this note centers on is whoever takes cost out of the most expensive slice of a health plan wins that volume, and volume compounds

Cancer is the most painful spot in most plans, #1 cost driver for 70% of large employers heading into 2027. Enter Starling with doctors, a risk contract and a pharmacy. This is what strikes me most about STLN’s version of value-based care and why the opportunity is so large for pharmacy which is largest cost bucket.

Two pipes, and cancer runs through both

• Specialty drugs: too costly, fragile or complex for a retail counter, new oral cancer drugs run ~$10K a month

• Almost nobody manages both pipes for the same cancer patient, except Starling

How the industry got here

• 1990s: home-delivery pharmacies for hemophilia, HIV and transplant drugs, Accredo formed in Memphis in 1996

• 2001: Gleevec approved, the first blockbuster pill that targets a cancer, the start of oncology moving from IV bags to bottles

• 2005 to 2018: the PBMs bought the specialty pharmacies, then the insurers bought the PBMs, CVS-Aetna and Cigna-Express Scripts both closed in 2018

• 2015 onward: oncologists pushed back with in-office dispensing, NCODA formed to set the standards

• 2024 to 2026: regulators caught up, FTC reports on PBM markups, the IRA Part D redesign, the Feb 2026 PBM reform law

Specialty pharmacy business is huge for PBMs…

Who earns what on a cancer pill

Whoever controls the network decides who fills

• Doctor’s offices fill 98% of own oral cancer scripts if no PBM pharmacy mandated

• One PBM pharmacy in the mix drops that to 79%, three or more to 58%, commercial scripts fall to 15%

• The script belongs to whoever sits at the gate, the lesson that makes Nexus matter in section 5

Where Starling sits

• Evolent: takes plan risk and sells approvals, owns no doctors and no pharmacy, 95.3% medical expense ratio in 2Q26

• OneOncology, US Oncology, AON: own thousands of oncologists, cannot take a plan’s full oncology risk

• Onco360, PBM pharmacies: own the pharmacy, no prescriber relationship

• Thyme Care ($2BN+ valuation): guides patients through care, approves and dispenses nothing

• Starling: the only row with all four

• Starling keeps 18-22% of each pharmacy dollar, the giants 4% to 8%

• The plan still pays less in total, $5,379 lower cost per cancer patient over six months when the practice dispenses

• $1,032 less wasted drug each time a dose changes, and treatment in the office instead of a hospital that charges 86% more per unit

• The pitch to a payer is a lower total cancer bill

People don’t stop getting cancer

• Pills keep taking share, 8 of 11 novel cancer approvals in 2019 were oral

• Oncology spend per member rose 11.6% in Prime Therapeutics’ commercial book in 2024, Evolent planned for 12% in 2025

Price is the one number moving the wrong way. Medicare’s negotiated prices cut revenue per fill even as fills grow.

• Imbruvica 38% below its list price in 2026

• Xtandi, Ibrance, Pomalyst and Calquence 40% to 60% below list in 2027

• Starling’s revenue per fill fell 12% yoy in 1Q26 while fills rose 103%

Starling’s pharmacy today

• Gross margin 15.9% in ‘24, 18.1% in ‘25, 21.6% in 2Q’26, management guides to “high teens” as a procurement benefit fades, our model holds 20%

• Scripts per 1,000 clinic visits 107 to 264 over eight quarters (Sep ‘26 investor deck)

• Paid once: a fixed monthly fee per member to run the plan’s cancer care, mature loss ratio 75% to 85%

• Paid twice: the margin on every oral script those patients fill at Starling

• The second payment only lands when the prescription reaches Starling’s pharmacy, and today that means Starling’s own doctors

“When we launch these deals, there is always, 100% of the time, pharmacy attachment associated with that.” Rob Carter, CFO, Wells Fargo Healthcare Conference, 9/8/26

“Our current Part D fills are all from our employed physician base. There are no Part D fills flowing through to our MSO providers.” Carter, Q1’26 call

Nexus

Why the approval step is the valuable one

• Prescription lane: any doctor’s record system can already send a script to any pharmacy over Surescripts, 2.3m connected users, no moat there

• Payment lane: the PBM decides which pharmacies are in network

• Approval lane: before a costly cancer drug is covered, someone approves it, and whoever approves sees the order first

• When a health plan hands Starling its oncology risk, Starling holds the approval lane for every doctor treating that plan’s members, including the 300+ network doctors it does not employ

What Nexus is

• Web portal launched Aug ‘26, built for oncology

• Network doctors onboarded in Sep ‘26, Starling’s employed doctors later this year

• A doctor picks a treatment, Nexus checks it against NCCN-based pathways, on-pathway orders are approved automatically, off-pathway orders get a medical director within 24-72 hours

• E-prescribing with Part D formulary visibility follows in ~Sep to Oct 2026, the first time a network doctor can send a script to Starling’s pharmacy from the same screen as the approval

• Trial matching, patient guidance and data for partners are planned - no date given

• Today it is a separate login, no link to outside record systems has been disclosed, the hub version sits inside the doctor’s own record system and the plans’ new prior-authorization connections that CMS requires from Jan ‘27

“Once fully launched, all of our providers, both PC employed and network providers will use the portal to submit treatment orders and obtain pre-authorizations from us... the portal will be a hub to drive ancillary services such as our Part D pharmacy, decentralized clinical trials and other value-add offerings to our network providers.” Dan Virnich, CEO, 2Q’26 call

“It is not something that we have guided to specifically, but it is certainly upside.” Asked whether that was all upside: “100%. That’s right.” Carter, Wells Fargo, 9/8/26

Land and expand is a large opportunity

Read more

This last piece is what I want to work through next, but on the enterprise side. Gemini’s progress has been disappointing, but material personnel changes and founder input returning seem to be turning things if we are to judge by recent rumors on the upcoming release of Gemini 4 as a hit in the making.

What some in the market miss today on AI in general is that the leading edge models are good, and while not AGI, if productized can drive huge changes in daily life. That’s literally what we’re seeing with Muse today. Google will be forced to the productization path next or risk having key advantages and legacy profit pools perish while Cloud continues its positive earnings upgrade cycle.

Why do I say this… well Google already has all the parts internally and just needs to piece them together to form an incredibly compelling enterprise offer for larger corporates all the way down to the long tail of SMEs…

No doubt it will take some time to realize a vision like this, and in the meantime they have to contend with worries on Search, CapEx, and short term cash burn. Search revenue still grew 17% in 2Q26, and AI Mode passed 1bn monthly users. The question that sets Alphabet’s value over the next decade is who sells AI to the tens of millions of businesses that have barely started buying it. Our answer is Google, and it takes only a little imagination to see why.

Data + distribution

In a recent appearance, Gavin Baker of Atreides Management argued that only three model companies will hold enterprise value in five years. “The three companies that have unique data are Meta, Google, and xAI.”

I would take it a step further… unique data keeps Google in the race. What Google owns beyond the model decides how much of the race it wins. This requires them to have a model on or near the frontier. On the July 22 call, Sundar Pichai said Google has started “our most ambitious pretraining run yet” and is “very committed and very confident of being at the frontier.” Once the model is close enough, the model is no longer the whole product. The pieces around it start to matter more, and Google holds more of them than anyone.

Here’s a quick rundown of some of those pieces / traits…

• Cash and cash flow to subsidize offerings - already effectively happening today vs. upsell charge for CoPilot

• Trusted brand with inbox and the calendar… first of 500 AI companies in Cybernews’ Aug 2026 trust ranking, with OpenAI 22nd and Anthropic 70th

• No baggage, no polarizing founder, no decade of social-media scandal to live down, while OpenAI ranked 68th in the 2026 Axios Harris Poll 100 and Meta and X trailed the rest of Big Tech

• Suite with Gemini built in, Gmail, Docs, Sheets and Meet, carrying mail for 3.6mm business domains

• Search and Maps plus a Gemini app at 950mm monthly users

• Rails via Google Pay to collect and Google Cloud to run the agents, with Cloud up 82% in 2Q26 on a $514bn backlog

Each rival can match one or two of these. None can assemble the rest quickly.

The desktop moat is shrinking

For 30 years Microsoft’s edge was the desktop app - near monopoly one might say. Office set the standard because it was a necessity of white collar / knowledge work. When an agent builds the model and drafts the memo, the app underneath matters less than the model and the data it can reach. A spreadsheet an agent built in Sheets is as good as one it built in Excel.

That moves the buying decision toward the people doing the work. IT leaders have long preferred Microsoft for control and compliance. Users tend to prefer Google, and ease of use is the strength G2 reviewers cite most for Workspace.

Price points the same way. Gemini comes inside a $14 Workspace seat. Microsoft charges $21 on top of its own $14 seat for Copilot.

A little imagination

Picture a 12-person accounting firm or a six-truck plumbing company in 2029. Today they barely use AI. The Census Bureau found AI use at firms with 250+ staff reached ~37% by May 2026, while firms under 20 staff were flat and the smallest stayed below 20%.

When those firms go as deep as the average tech company is today, they will not hire an AI team. They will need handholding, and a product built into tools they already open every morning. For most of them that is Google. They run mail on Gmail and find customers through Google Search and Maps.

The vendor already sitting in those places can sell the agent that answers the phone and books the job, then charge for it. Philipp Schindler said on the July 22 call that Google’s new agentic tools for small businesses had reached hundreds of thousands of new customers this year. Why would there not be a far larger software suite that sits on top of that over time? Some of it will be built for specific trades and sold into small and mid-sized firms through channels Google already owns. As the cost of building declines, so does their ability to meet more verticalized needs over time.

Google is also where many of these firms begin. Of the Workspace companies matched to a headcount, 70.5% have 10 or fewer staff.

Small checks, nearly half the workforce

Each contract would be small, but they are by far the best positioned to implement AI into the long tail of 36m American small businesses employing 62m people, 46% of the private workforce and the largest single block of American labor, per the SBA. Every $10 a month of AI spend per small-business worker is ~$7.5bn a year of revenue. At $30 it is ~$22bn, close to a full quarter of Google Cloud revenue today. And why wouldn’t this also be the case across Europe?

Even better, nearly 90% of the Fortune 100 use Gemini Enterprise, per the 2Q26 call. Simply put, they have the ability to sell one AI stack to a one-person shop and to the Fortune 100. Microsoft is strongest higher up the axis. Meta holds the storefront. WhatsApp counts 200mm+ small-business users.

Google is the only company with a piece at every point on the axis. The market still sees a search company with a capex bill.

What has to go right

• Gemini 4 must be a strong launch with a more sustained pace of releases… the latter should be increasingly possible once back on the frontier as both Anthropic and OpenAI are using their latest models to train the next

• Workspace keeping the firms it signs as they grow, Microsoft still wins ~2 mail-provider switches for every 1 it gives back (TechnologyChecker, Aug 2026), and ideally they start to stem the tide in the NTM

• It would be great to see a Google’s small-business agent tools vs. Meta’s Business Agent at 1mm+ businesses

Putting it all together

It doesn’t take much imagination to understand what the existing component pieces of the business can produce in say 3 to 5 years, but the market clearly is not doing so today. There is no player better positioned to meet the needs of SMEs to enterprises, and I dare say even push into other adjacent spaces further like payments and operations software systems. Reading the tea leaves, it feels like there are A LOT of changes happening internally all at once. Results will take time but their position is irreplaceable. But just giving the market a taste of a vision like that above should ultimately support valuation and provide comfort that the large CapEx bill will have positive ROI.

DISCLOSURE

Position: At the time of publication, the author does hold a position in GOOGL.

Trading Policy: The author will not materially alter this position within 48 hours of publication. After this period, the author may buy, sell, or otherwise adjust the position without further notice. Changes to the author’s view or position will be reflected in subsequent publications when material.

Conflicts: The author has received no compensation from the issuer or any party with a financial interest in this security.

Forward-Looking Statements: This report contains the author’s opinions, estimates, and projections, including price targets derived from financial models. These are forward-looking statements subject to substantial uncertainty. If these assumptions prove incorrect, the actual value may differ materially, including scenarios of significant loss or total impairment. The price target represents the author’s estimate of fair value under the stated assumptions, not a prediction of where the stock will trade.

This report is provided for informational purposes only and does not constitute investment advice. See the full Terms & Disclosures for additional important information.

I'm looking to make room and may add soon.

The deal

RLC closed Monday at $250, Sandals rumors hit Tuesday, deal signed Wednesday, stock -10% since. Oil and rates also not helpful today. ~$61bn market cap paying $3bn for 50% of Sandals and Beaches Resorts… but they lost ~$6bn in market cap… safe to say some big shareholders were caught off guard….

Why buy a Caribbean beach resort

Share of wallet! RCL’s deal deck statesRCL captures 20–25% of its guests' five-year vacation spend. The ships are already full, and they’ve shown a consistent ability to fill new berths (2 new ships in ‘26, 1 in ‘27, 3 in ‘28, and entering river cruises). So the next dollar has to come from the same household’s other trips.

• Structure… 50/50 board, equity method so JV debt stays off RCL’s balance sheet, Morgan Stanley committed financing, close early 2027, accretive to 2027 earnings per RCL

• Wallet is the constraint.. 110% load factor in 2Q26, 57% of 2026 capacity in the Caribbean

• Missing trip type is beach all-inclusive… ~30% overlap between RCL guests and Caribbean resort travelers over 2yrs, booked as an extra trip rather than a swap

• Faster growth segment… Caribbean all-inclusive spend $9bn in 2019 to $16bn in 2025 for 11% CAGR, against RCL capacity of 4-7% guided for 2027-29

• Best asset base in the basin… 18 resorts and 6,155 keys, 7 of them in Jamaica, 3 more Beaches coming

• Founder’s son Adam Stewart stays executive chairman

RCL paid ~10x forward EBITDA, which is ~$600mm EBITDA for the whole company and ~$975K per key. Hyatt (H) paid 8.5-9.5x stabilized ‘27 EBITDA in 2025 for Playa’s management contracts, asset-light, with no owned real estate. RCL itself trades at 9.6x CY27E EBITDA pro forma… so what’s the big deal? This is likely accretive to their mid-teens ROIC over a few years. And what happens if they get to know the asset and buy in the whole thing as they did before (Silversea).

Where the multiple sits

Assume the world doesn’t end, oil and rates don’t stay here forever, boomers keep spending their interest income on cruises, and mass AI job losses don’t occur… backsolve from target 12x ‘27 EBITDA and the market is paying for $7.1bn of 2027 EBITDA. That is 8% below this year's $7.7bn guide and 15% below ‘28 consensus. The sell-side is actually quite good at modeling cruise lines and typically takes forward curves for fuel and have a good view into booking curves and yields out several quarters. Let’s assume the world doesn’t end and that consensus is a good starting point, which I firmly believe it is… the market is now giving them no credit for potential upside from the deal… let’s walk through that below. Better yet, RCL says 2027 bookings are “pacing ahead of historical levels” at higher prices and 53% of 2027 fuel is hedged at $405/t. A 2027 net yield guide below +1% in February changes it.

Subscribe now

Read more

fka The Oncology Institute (TOI)

Last: $6.75 Mkt cap: $805mm fully diluted

1.0x ‘27 CONS Sales / 27x EBITDA (b/e margins in ‘26)

The business in short…

• Largest value-based oncology platform in the U.S.

• Hybrid care delivery model (employed physician group and network of independent providers) plus large specialty pharmacy (60% of revenue)

• Operate across five states covering 2.1m lives in CA, FL, NV, AZ, OR

• Fixed monthly payment per covered life from health plans and medical groups, then owns that population’s Part B cancer bill

• 116 employed oncologists and clinicians across 80 sites, 300+ independent doctors under contract, 17 markets

• Buys 95% of the drugs it administers via owned specialty pharmacy

• FY2026 guide to revenue $650–670mm, first year of positive adjusted EBITDA as a public company

• 2027 guide to capitated revenue doubling to $300mm on contracts mostly already signed

Note: new CEO in place for 3yrs, meaningful turnaround corrected meaningful mis-steps that will not be covered here but are widely covered elsewhere

Personal note: after 10 years on the buy-side, I spent a year helping restructure a large direct primary care business (DPC), also value-based care, with ~15k members and an in-house pharmacy…. I have seen the explosive growth firsthand when VBC models manage to engage and align incentives for large pools of patients

Why oncology needs value-based care

Fee-for-service pays a doctor for each thing done. Value-based care (VBC) pays for an outcome, or in this case a population, and lets the provider keep what it saves. In oncology, this means capitation (charging per head) on the Part B benefit:

• Plan pays a fixed dollar amount per member per month (PMPM), whether or not that member has cancer

• Provider takes responsibility for the doctor visits and the infused drugs of everyone who does

• Provider keeps the difference between the check and the cost, and eats the difference if cost is higher

• Pills taken at home (Part D), hospital stays (Part A), CAR-T sit outside the risk

The reason this exists at all is that oncology is a massively inflating portion of risk pools with long tail outcomes that are difficult to manage… 10% of some pools with $300bn of annual spend… fee-for-service has failed here.

Why oncology is such a mess

Drug prices are key part of problem. US oncology drug spend ran ~$145bn in 2024, projected $180bn by 2028. Biologics were 79% of Part B drug spending by 2021 and 89% of its growth. CEO Virnich at Wells Fargo said “oncology is the most expensive specialty that payers have to deal with, all driven by the costs of the drugs.”

Everyone in the chain is paid cost-plus. A community oncologist buys the drug, administers it, bills at average sales price plus 6%, sequestered to ~4.3%. Between 80 and 85% of a fee-for-service oncologist’s income is that spread. The cheaper drug pays him less. Nobody in the delivery chain is paid to spend less. Incentives simply are not aligned.

The same infusion costs 86% more in a hospital. EBRI, 2016 commercial claims, 18,195 patients on the top 37 infused cancer drugs saw that hospital outpatients paid 86% more per unit than physician offices, $21,881 a patient against $13,128.

Hospital share has gone up, not down. Physician offices fell from 63% of Part B drug spending in 2008 to 53% in 2021. Hospital outpatient rose 23% to 41%. The mechanism was 340B where hospitals buy at statutory discounts, bill at ordinary rates, and in 2022 earned a margin on cancer drugs equal to 45% of reimbursement… $570,161 per medical oncologist against $64,236 for other specialists. That number is why hospitals bought the practice down the street. Community Oncology Alliance count since 2008: 435 clinics closed, 722 acquired by hospitals, 203 merged. To be fair, though, you cannot blame them. Reimbursements are being squeezed and they’re looking for anywhere they can make margin.

Payers being squeezed from both sides. Medicare Advantage (MA) cut rates the last two years while acuity and utilization rose. Oncology runs 8 to 11% of MA premium. The legacy fix was a benefit manager reviewing prescriptions, and the loss ratios quoted for that model sit in the mid-nineties.

Medicare itself tried three times and lost money each time. The Oncology Care Model, 2016–2022, had 202 practices… real gross savings, fully offset by incentive payment and net costs to Medicare. Oncology Care First never launched. Enhancing Oncology Model 2023 onward saw uptake collapse from 122 to 44 practices, so terms had to be sweetened, first evaluation still lost $13.2MM in six months.

Subscribe now

Politics and well-meaning frustrations from all parties aside… there is no systemic fix to this in sight, but multiple parties are at breaking point

Enter VBC

• Payers cannot make oncology risk work in-house and have repeatedly failed trying, so they hand the whole benefit to somebody who can

• The site-of-care spread is the largest of any specialty

• Policy finally pushed the other way… CMS site-neutral payment for drug administration in off-campus hospital depts from 1/1/26, ~60% cut in that setting, $8bn estimated savings over ten years

• Every other specialty with this spread saw capital arbitrages it, then the sector consolidated, and in two cases the provider ended up holding the risk

Dialysis is the precedent that completed the move. The ESRD bundle in 2011, then Kidney Care Choices with 100 risk-bearing entities and roughly half of eligible fee-for-service beneficiaries in a full-risk contract today. Behavioral health got to 71MM at-risk lives by 1997, then the plans bought the vendors back. Oncology has the widest price gap of any of them and the weakest record of closing it.

How Starling makes money

1. Health plan pays someone else to own its cancer bill

2. Two kinds of contracts

• Narrow network, California since 2007… counterparty is a medical group under a Knox-Keene licence, STLN is its only oncologist, group keeps the network and pricing power, mid single-digit dollars a life a month, loss ratio 65–70%, group can swap providers so the rate is capped

• Delegated, the growth product… counterparty is the health plan who “delegates” process to STLN, STLN takes prior auth, claims payment and the network, mid double-digit dollars because STLN bears the risk, up to 10x California, loss ratio matures 80–85%

• Why 10x: Florida’s benchmark is hospital-delivered care because this accounts for 40–50% of oncology patients who are served at 150–175% of Starling’s cost, Starling prices against that

• 2025: old product $85mm, new one $11mm; delegated was under 5% of lives and a third of run-rate revenue at year end

• Delegated can launch a market on a contracted network before Starling opens a clinic, so it can go into states where Starling has no footprint

• Owned clinics follow the lives, not the other way round

3. Payers hand Starling blocks of lives

• 2m lives YE’25 to >2.3m by October as contracts ramp

• The trap on this visual… California exclusivity 230,000 lives at ~$2.20 a month, about $6MM a year — lives are not the unit of value, revenue per life is

• Clinic cost is fixed; California ~75% full, Florida ~40%; a patient in a lit room is close to pure margin, above the gross-profit line

TAM is LARGE

The company says it is under-penetrated in those five states and that the delegated product applies to basically the other 47 states. Here is the denominator, CMS May 2026 vintage…

• STLN’s 2.1m is 7.7% share in their existing 5 states and just 1.9% of the L48 core

• ~100 million lives is the potential opportunity for them

• ~48% of Medicaid lives are children so excluded from oncology capitation;

• Evolent, closest comp, carries 6.7m Performance Suite lives across three specialties, or just 6.5% of the pool

• No published total exists for US lives under capitated oncology but Evolent plus STLN is under 9m total addressed today - this penetration is now turning and accelerating

• Alaska has 2.8% MA penetration and no Medicaid managed care; Hawaii is 0.4MM Medicare — the 48-state framing is right

4. Paid on everyone but spent on the few

• Active-treatment share never disclosed but we carry ~4%

5. Massively underappreciated pharmacy levers

• Check arrives first, prepaid, so growth consumes no working capital

• Cohort launches near 90% MLR, staff hired pre-launch, but with 90 day established continuity of care the MLR settles near 84% over a year

• 4 leaks getting plugged to improve cash flow: out-of-network scripts steered to STLN pharmacy, hospital infusions redirected via network redesign; off-pathway treatments changed to NCCN standard; list-price drugs at independents brought in-house

• Company’s own evidence across every delegated contract as of Aug 2026… cost per episode down 27%+, per IV visit down 41%+, 98%+ on formulary, 95%+ decided same day

• Risk corridors on every plan contract, and one 2m-life pool, so one bad book does not set the year

6. Every contract seasons on its own clock

• Blended loss ratio now disclosed seven quarters… 61% in 1Q25, 85% by 2Q26 as delegated lives came on

• Dollars run opposite to the percentage, so delegated keeps 15–20% of a premium ten times larger

• Most important number coming is the Q4’26 MLR… will be worst in company history as contracts ramp and then fall back down toward 82% by 4Q27

7. Pharmacy business underappreciated quality and upside

• Buys 95% of what it administers and one of the largest oncology drug buyers in the country that is neither a distributor nor a health system

• Pharmacy is 61% of revenue and 78% of gross profit in 2Q26, quick collections

• Nexus launch is a massive underappreciated lever, we owe more work on this - hugely strategic asset that could change the landscape in the markets they enter

8. Pharmacy grows as pool grows

• Three structural changes, not a sales push…leakage recaptured (done), pharmacy licenced in California Dec 2023 and Florida late 2025 (so opening Medi-Cal patients previously unservable, done), the network door (not yet)

• Asked if opening the network would be all upside to guidance… 100%

9. The piece that is shrinking is the worst piece

• Model it falling 6% a year from 2027, capitated contracts eating their own FFS patients, a transfer from a 6% line to one carrying 15–30%

10. Underappreciated SG&A leverage

Read Management’s own words… they’re saying their own guidance is conservative… on nearly every metric…

Quotes from Q2’26 call Aug 6th and Wells Fargo Sep 8th… none of it is in our base case unless the plate says so

On 2027… “The combination of those is really what is driving our 2027 guide of 100% capitated revenue growth. Going from $150 million to $300 million of capitated revenue. The basis of that is those deals, which obviously represents a tremendous upside next year as there is additional deals in the pipeline as well.”

On how much is signed… “The majority of that has been announced to date, the vast majority of it... the bulk, again, of the $300 million is from the contracts already announced.”

On what was not forecast… “The exclusivity obtained this quarter in California was not part of the initial forecast nor was the 2 contracts, Oregon and Nevada.”

On the pipeline…. “Our pipeline is incredibly robust. The demand for our services has never been stronger.”

On pharmacy attachment… “When we launch these deals, there is always, 100% of the time, pharmacy attachment associated with that.”

On the network door… “Through the delegated networks and the networks that we are building, there is the potential for us to also add our network providers as well. Our new product, Starling Nexus, will aid in that. It is not something that we have guided to specifically, but it is certainly upside. Moderator: That would be all upside as you think about your financial model today. Carter: “100%. That’s right.”

On guidance excluding it… “Analyst: Still nothing assumed in the guidance in terms of any Part D lift related to the portal rollout, correct? Virnich: No, nothing.”

On the pharmacy beating plan… “Our Specialty Pharmacy segment this year continues to surpass our own expectations.” “Our Part D business is continuing to set monthly fill records.”

On the loss ratio... “On the delegated product, we’re looking at MLR between 75% to 85% once fully ramped.”

On SG&A… “What does it take to keep the lights on at Starling? Quite frankly, it’s quite minimal.”

On the market… “We exist in an industry that spends over $300 billion a year. We’re currently in five states, and frankly, under-penetrated in those five states. The ability for this business to grow is really unlimited.”

On what comes next… “Ancillary services that we can extend to our now network of providers that is 300+ in head count, as well as the ability to capture the massive amount of drug procurement we’re doing and lateralize that into other specialties over time.”

On scale... “This is a business that will be a multi-billion dollar platform soon, I’m sure of that.”

On how they win... “It’s really been more about winning business from competitors than consolidation per se... service and price seem to be the way we’re winning in other markets, too.”

Below the paywall you’ll find a breakdown of KPIs to financials to price targets, as well as 8 discrete sources of upside to guidance/consensus quantified. Without giving away too much, the market is just starting to catch on to the income statement turnaround that is set to take place in the next 12-24 months. However, the cash flow and balance sheet transformation the market will be forced to price once it understands where margins can go will allow true platform valuation… there is massive strategic value to be unlocked with a bit more scale and once they are flush with cash.

Read more

THIS IS A 30-MINUTE PLUS READ… SO IF YOU DON’T HAVE THE TIME, SKIM THE TITLES AND BOLD TEXT AND CHARTS. WELCOME FEEDBACK AND DEBATE.

My argument is that the current market structure is a manufactured disequilibrium that will decay overtime. The path there is important and will become more clear likely in the coming months as we understand whether Astra-class models and agents are the next OpenClaw-style inflection in demand we saw in early 2026. Regardless, the path is toward consolidation for a long list of reasons and neoclouds would be well-served to monetize sooner rather than later as we roll into a very favorable pricing environment in 2027.

The endgame is concentration. Compute should be owned mostly by the parties closest to demand, with the lowest cost of capital, and without the fragmented merchant tier between them. Owners of scarce, hard-to-replicate physical assets (permitted, powered, interconnected land backed by long-dated investment-grade offtake) are acquisition targets that clear at a strategic premium, not fire-sale prices, but only if the strike deals in a strong market environment. Pure GPU-rental operators carrying 10-15 year lease liabilities against a 2-5 year pricing window are the equity-wipe candidates if they do not execute. Nebius is a category error to lump - it owns capacity and runs a software stack and has marquee partnerships, which is why the market already pays it a different multiple. The highest-confidence three-year expressions are to own the integrated demand-owners who capture the rent, weighted toward those that own competitive silicon rather than merely cheap funding, and to rent the upstream chokepoints that get paid regardless of which neocloud wins. The merchant tier itself is event-driven, not compounding, and should be underwritten to an exit rather than to a terminal multiple.

Structure… few buyers and many sellers

Frontier compute demand is set by roughly seven entities vs. 30+ neoclouds and converted-miner landlords, all leveraged, all buying from the same chip vendor, all chasing the same handful of contracts. When concentrated buyers face a fragmented, capital-intensive supplier base, price-setting power sits with the buyers the moment scarcity relaxes. That is an oligopsony.

Demand

Forur captive builders plus three merchant labs matter. Everyone else, Mistral, Cohere, SSI, the inference resellers, the enterprise long tail, is small, lower-credit, and price-taking. Annual capex frames the scale as big four hyperscalers guided to roughly $720 to 745B for calendar 2026, up from about $410B in 2025. Add Oracle at roughly $70B for FY27 and the top of the demand stack is near $835B a year. On the merchant-lab side the flows are stocks not annual runs, but they are enormous: Anthropic committed north of $275B of cloud compute across five deals in 2026 alone; the Oracle to OpenAI arrangement is a similar order of magnitude.

Those seven are already integrating

The seven buyers are not passive tenants… each builds its own DCs... 6/7 have custom silicon in production or in tape-out... 3 of them run enterprise software and commercial security businesses that already sell the trust layer a lab needs.

Thomas Kurian, who runs Google Cloud, described the end state without meaning to: some labs use Google’s TPUs and Gemini, others use the TPUs and then buy Google’s cybersecurity protection for their models. That is the full stack sold in pieces, by the party that owns all the pieces.

The neocloud owns none of these columns and simply will not be able to compete here over time.

Next demand names uncertain

If the buyer set expands, it does not expand into a long tail of startups. It expands into a small number of entities with balance sheets and proprietary workloads. House ranking, in order of expected signable external demand:

None of these is a durable neocloud customer. Apple would build and own. Sovereigns are buying their own national infrastructure. Quant finance is the one genuinely additive merchant customer, and it is small relative to the labs. The buyer set widens slightly and gets richer, but it does not get more fragmented, which is what the merchant tier would need.

Weighing both sides

Two defects of the neocloud model

First, the merchant tier an income-statement problem via a permanent cost-to-serve disadvantage. Second, a balance-sheet problem with a liability that outlives the revenue that services it.

Unit economics currently masked

Standalone merchant compute carries a structural cost-to-serve disadvantage against the integrated owners on four counts, none cyclical:

• Cost of capital: hyperscaler funds near its cost of cash and a neocloud at high-yield or GPU-collateralized rates… this delta compounds rapidly over the years

• Dual-use flexing: integrated owner moves capacity between internal and external work vs. merchant binary idle risk

• Captive silicon: TPU and Trainium cut inference cost per token below a merchant Nvidia fleet

• Subsidy capacity: search, ads, cloud and enterprise software fund compute at deliberately thin margins for lock-in

None of this matters today, or until demand meaningfully breaks, because the merchant tier delivers the single scarcest thing in the industry… energized capacity now. Buyers whose own power queues run years out pay a premium to intermediaries who already hold sockets. That premium is a rent on a scarce factor, not an arbitrage. An arbitrage is riskless and closes cleanly - this is not that at all.

Liability outlives the rent

The second defect is what turns poor returns into zeroes. Premium pricing that justifies the build lasts one GPU generation. The lease and the debt that financed it last three to five. Stack this on the wedge above and the sequence is clear: the rent decays, the cost disadvantage surfaces, and the fixed obligation is still there for another decade.

Who owns what

Operator quality against asset quality dispersion

Treating this cohort as one asset class an error. Asset quality is the land, the power, the interconnect position, the permits. Operator quality is whether the megawatts actually produce sellable tokens: goodput, uptime, fabric design, cluster orchestration, time to energize. These are not correlated. A converted miner can hold superb land and be a poor operator. A pure tenant can be an excellent operator on borrowed shells.

Architecture trap… operational sophistication destroys acquisition value

A key risk is that hyperscalers will not buy merchant neoclouds, because they run proprietary reference designs and will not inherit someone else’s bespoke ops and infra. What a hyperscaler mostly likely wants from an acquisition is raw, clean, conformable megawatts: permitted sites, long-dated interconnects, pipeline capacity it can shape into its own internal design. Or they want a great price!

If what the buyer wants is conformable capacity rather than a company, it does not need the equity at all. It can buy the land, buy the development pipeline, or simply sign a lease. That weakens the takeout case for operators specifically and concentrates it on landbank holders with uncommitted capacity.

One correction to the objection, which matters for who bids as we’ve seen two distinct acquirer types. Hyperscalers buy capacity and conform it. AI-native buyers buy stacks, and they have been active, with CoreWeave acquiring Weights & Biases, OpenPipe and Monolith, and Nvidia absorbing Groq into rack-scale LPX in about eight months. So sophistication is not unsaleable but requires a different set of circumstances.

Roughly a dozen listed miners have signed decade-plus leases this year, and CoinShares put the sector at up to 70% of revenue from AI by December 2026, from about 30% earlier in the year, on more than $70B of announced contracts. That is a genuine transfer of scarce, energized land into AI service. It is also the clearest evidence for the thesis: when a $10M market-cap shell can sign a headline billion-dollar contract, the scarce thing is not the operator. It is the interconnect queue position, and that is exactly the asset that stops being scarce once the queue clears.

The one exception, and why it is not a contradiction

Among the pure-plays the answer is one name. Nebius is in a league of its own, and it should not be grouped with the generic GPU landlords or the leveraged sublease vehicles.

The separation rests on a verifiable asset fact, not positioning:

• Owns >75% of its capacity, against a cohort defined by leasing shells

• Built the layers between silicon and enterprise workload (orchestration, managed services, confidential compute), converting commodity hardware into something with retention rather than spot churn

• ~$46B backlog including the Microsoft agreement; the market already marks the difference, Nebius +130% ytd vs CoreWeave +7%

This is where the architecture trap stops being a contradiction. Sophistication is only self-defeating for an operator that needs an exit, and Nebius is building toward not needing one.

When does scarcity ease?

Scarcity eases unevenly by region and by procurement route. The binding constraint is energization. Interconnection queues in the busiest US markets run four to seven years. Queue-to-commercial-operation has stretched about 60% since 2017 and now averages over 2,100 days. Large power transformers run roughly 128 weeks and generator step-up units 144, with substation transformers now past 160 weeks, against a forecast 30% supply deficit through 2027; the same unit shipped in four to six weeks in 2020. Large-frame turbines from GE Vernova, Siemens Energy and Mitsubishi are booked through 2028. ERCOT’s large-load queue went from 63 GW to 226 GW in a single year with roughly nine of ten requests from data centers. PJM’s capacity auction came in 6.6 GW below its reliability requirement for 2027-28 at a record clearing price, with data centers driving about 94% of projected load growth through 2030. Dominion has said it cannot take additional large-load interconnection in Northern Virginia through 2030. Of 16 GW targeted for delivery in 2026, only about 5 GW entered active construction, and 30 to 50% of the remainder slips.

A second, independent constraint sits alongside it, and it needs to be scoped precisely rather than waved at. DRAM is a 2027 constraint on incremental capacity above current plans. Builds already in the plan of record have memory allocated against them. What memory gates is the upside case: the additional gigawatts a builder would add if it could, and the higher-specification configurations it would prefer. That is visible in Rubin Ultra de-speccing from planned 16-hi HBM4E toward 8-hi HBM4, which the house read treats as supply rationing rather than demand destruction. The practical effect is that memory does not stop the current build; it caps the response to any 2027 demand surprise, which is precisely the mechanism that keeps pricing firm.

The energization gate is firmly in place through 2027, and plausibly persists into 2028 depending on permitting, transformer and turbine progress. Both power and DRAM likely ease into 2028 rather than clearing. The merchant tier has sometime to execute, but any softening in the market (demand or bottlenecks) could prove disastrous for their equity.

Margin path

What matters is where each type of player’s margins are heading over the comin years. And again, this does not bode well for neoclouds who are simply disadvantaged.

NVDA: the counter that matters most

Nvidia sits on five sides at once… chip supplier, equity holder, lease-holder, rent-back guarantor, and now owner of the model-distribution layer via Hugging Face. The hyperscalers backstop offtake to keep capex off their own balance sheets. This is engineered solvency, and it can persist far longer than a naive short thesis assumes. It cuts both ways, both confirming these are financial vehicles rather than franchises, and it explaining why they stay bid.

Backstop is wildly important

The backstop is simultaneously what keeps the merchant tier alive and the clearest evidence it is not a business. In August 2026 Nvidia announced financing platforms with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to mobilize over $500B of third-party capital, and is backstopping up to 25% of a site’s residual value. Jensen Huang’s framing is that AI factories are an investable asset class… they produce revenue, serve a broad market, improve with CUDA over time, and can be redeployed.

Ben Thompson’s read is the one to carry: the guarantee is a price cut in disguise. Nvidia is putting its own profit at risk in order to lower the cost of capital for people building with its chips. Follow that logic and three things fall out.

Nvidia’s pricing power is already peaking. A monopolist with unchallenged pricing power raises price, it does not subsidize its customers’ financing. Thompson locates the cause precisely in the guarantee downstream of Google’s and Amazon’s aggressiveness. If capital is the binding constraint on a new data center, a cheaper chip up front beats a more efficient one, which is exactly the opening TPU and Trainium exploit. Nvidia’s own new reporting split, separating hyperscaler sales where it is fighting commoditization from everyone else where it runs the whole stack, is the company conceding the two-tier structure in its disclosure.

The fragmented tier is being financed, not validated. Nvidia is not making these companies profitable. It is making them fundable. This is a concentration hedge expressed in the capital structure. Remove the backstop and a large part of the tail cannot clear its return hurdle at any contract price the buyers will pay.

Risk is being moved somewhere worse. The big four raised $194B by early July 2026 versus $108B in all of 2025; spreads are widening, 86% of this year’s bonds trade above their issue yield, and cover has fallen below 2x from 5x in February. The marginal dollar is now moving past IG debt into insurance floats and pension money via the asset managers. Thompson’s analogy (and Nadella’s, on the MSFT call) is Jay Cooke financing Northern Pacific into the Panic of 1873… adjusted for economy size, that boom’s railway bonds were ~$600B a year, the same order as today’s tech capex.

When fragmentation unwinds

The fragmentation is a manufactured disequilibrium, held up by three props that are all decaying… a chip monopolist managing its own customer concentration, a physical shortage minting temporary scarcity rents, and the speed premium of an unbundled supply chain during a land grab. When those props go, the structure reverts toward vertical integration on the balance sheets closest to demand and with the lowest cost of capital.

Nvidia’s worst outcome is a world where four buyers take most of its output. Those four (Amazon, Google, Microsoft, Meta) all design competing silicon, all have buyer power, and all want Nvidia’s margin. So Nvidia seeded the alternative.

Training and inference are forking

It is now clear that the highest-end GPU is a training requirement, and not an inference one. Inference, which is where the volume and the recurring revenue go, runs well on heterogeneous and custom silicon… TPU, Trainium, Inferentia, and depreciated prior-generation GPUs. Amazon and Google pay thinner margins to Broadcom, Marvell and Alchip than Nvidia charges, and still resell compute at a lower cost per token. As inference becomes the majority of the workload, value migrates from Nvidia’s newest silicon toward custom silicon that mostly sits on integrated balance sheets.

This leads straight into the neocloud problem through the GPU’s second life. A demand-owner cascades a chip… bleeding edge for training, then the same chip redeployed for internal inference for years. A standalone neocloud cannot cascade. It has to re-lease the depreciated asset into a falling spot market. Same silicon, very different lifetime value, and it is worth more inside the integrated owner.

Bundle, unbundle, re-bundle

This is a classic technology cycle, not a one-off. Value integrates, then disaggregates for speed, then re-integrates once specs stabilize and scale justifies owning more of the stack. Today the AI silicon chain is fully unbundled: TSMC fabs, Nvidia designs, an OEM assembles, a neocloud operates, a lab consumes. Every hop leaks margin. Unbundling won the land-grab phase because specialization let everyone scale fast. Re-bundling wins the maturity phase, because at 800B+ of annual spend the margin leakage is intolerable and the integrated owner captures it. Google is the Apple of this cycle… its own chip (TPU), its own data centers, its own model (Gemini), its own distribution. That is why the fully integrated player is best positioned, and why the middle of the chain gets squeezed from both ends.

Rent extraction will be temporary

The value a buyer extracts per token is rising faster than the cost of producing a token is falling. While compute is short, the data-center layer can push its price up into that widening gap, or at least not let it stretch so much, and neoclouds that are recontracting or repricing right now capture it. That is real, and it is the entire bull case for owning them today.

But the DC layer’s pricing power is a shortage rent, not a value-add. When the shortage ends, cost deflation resumes and dominates, the compute price rolls back toward cost, and the widening surplus accrues to the model and application layers instead. The window is defined by the shortage, and its width depends entirely on the compute demand curve.

WACC and proximity to demand decide the owner

In steady state, compute should be owned by whoever can fund it most cheaply and flex it against internal demand. That is not a neocloud. Two structural edges compound (both priced in Figs 12a to 12e):

• Cost of capital: on a 15-year depreciating asset, the WACC gap alone lets the mega-cap outbid everyone for the same asset and still clear its hurdle

• Dual-use optionality: those with internal demand burn compute when that return is highest and sells externally when it is not… a pure neocloud is structurally short an option only the integrated owner can realize

The same asset is therefore worth more inside Google or Amazon than as standalone neocloud equity… the textbook condition for consolidation.

Cost of capital comps

Alphabet’s size-weighted coupon across $62.5bn of issuance is 5.14%. Amazon is 5.04% on $77bn, Meta 5.41% on $55bn, Oracle 5.61% on $43bn; Microsoft has issued essentially nothing and funds from free cash flow. The neocloud cohort spans 5.90% to 9.88%. So the gap is 99bp to 474bp, not the 700bp Fig 12 carries, and it is a range rather than a number.

Converts distort comps

The instrument that makes this cohort look cheaply funded is the one that most distorts the comparison. Across $24.8bn of convertible paper from CoreWeave, Nebius, IREN, TeraWulf, Cipher, Core Scientific and Bitdeer, the size-weighted headline coupon is 1.43%, and four of the fifteen tranches carry a zero coupon.

A convertible is a bond plus a call option written on the issuer’s own equity, and on a stock running 65-80% volatility that option is expensive. Valuing each one at issue and annualizing it over the life of the note gives a true economic cost of 8.2% to 9.5%. The headline understates by roughly 680 to 810bps, which means reported interest expense understates the true cost of capital across the entire cohort.

Carrying it through to returns

GCP sells complete TPU systems externally at $35bn per GW gross at a low-30s EBIT margin, so Google’s own cost to build a gigawatt of TPU is about $23.8bn. TPU v7 TCO per chip runs roughly 44% below a GB200 server, which grosses the equivalent Nvidia deployment to about $42.5bn. Add $11bn of shell and power to each and the stacks are $34.8bn against $53.5bn per GW. Revenue comes off the rental curve: at 3.5kW per GPU including PUE a gigawatt carries about 286,000 GPUs, so at 90% utilization the cost-based floor of $4.92 per GPU-hour yields $11.1bn per GW per year, today’s shortage pricing around $7.50 yields $16.9bn, and the value-based ceiling of $9.63 yields $21.7bn. Depreciate over six years, carry $1.5bn of cash operating cost, and the bridge closes.

The finding that reorders this note

Decomposing the gap says cost of capital is the smaller half of the story. Between Google on TPUs and a neocloud on Nvidia GPUs at shortage pricing there is a 20 point difference in spread. Hold capital constant and take away the TPU and the spread falls from +22 to +7.0: that 15 points is the silicon advantage. Hold silicon constant and take away the balance sheet and it falls from +7 to +2: that 4 points is the capital advantage.

Nvidia allocation stops being a moat

The original neocloud edge was privileged access to scarce Nvidia silicon. That edge is decaying on two fronts. Chips are getting easier than shells. TSMC has doubled CoWoS advanced packaging repeatedly; the acute allocation crunch of the early buildout has eased. A buyer with a signed power agreement and a data hall can now get accelerators on a shorter timeline than it can get a grid connection. The scarce thing moved from the chip to the socket to plug it into. An intermediary whose franchise was chip allocation is holding the asset that stopped being scarce.

Inference silicon is proliferating. As the mix shifts toward inference, purpose-built parts undercut the general-purpose GPU, and almost all of them are captive to a demand-owner. Note what OpenAI’s Jalapeño demonstrates beyond its specs: concept to tape-out in about nine months, with OpenAI’s own models assisting the design. That is the cost curve bending under AI’s own hand, and it is a template others will copy. Each new captive part removes a buyer from the merchant GPU market permanently.

Fewer labs perhaps, but more integration, and more captured rent

The final loop closes on the demand side. Inference, not training, is where model-company margin now accrues, and inference is a fundamentally different business that is more stable, repeatable, and continuously tunable. A lab running a large inference book can grind its cost per token down every quarter through kernels, caching, quantization, batching, scheduling and silicon choice. Those gains compound and they favor scale. Blended inference margins have already moved from under 40% to over 70% inside a year while the sticker price per token fell.

In commodity markets marginal cost determines not just profitability but viability. The frontier labs have the superior cost structure, and they are widening it by converting compute from a marginal cost into a capital cost, which is exactly what Anthropic did by buying TPUs outright rather than renting them. So the lab field narrows. As it narrows, economic profit concentrates in fewer hands, and those hands have both the cash and the motive to integrate further down the stack, capturing the rent that currently leaks to intermediaries.

The loop is bad for the merchant tier twice over. A consolidated lab field means fewer buyers, which worsens the oligopsony in Fig 1. And better-capitalized labs means more self-build, which shrinks the merchant market those fewer buyers address.

The open-weight objection, and why it fails

The most common rescue offered for the merchant tier is that open-weight models (Llama, DeepSeek, Mistral and successors) break the oligopsony by decentralizing the buyer universe from seven names into thousands of enterprises.

Model IP does decentralize, but serving capital does not. Enterprise adoption of open weights does not produce a Fortune 500 that each build data centers, rather it produces a Fortune 500 that queries hosted open models through managed endpoints. The question then becomes who serves those tokens at the lowest cost, and in a commodity market the low-cost producer wins. That is the integrated owner.

A hyperscaler can serve open models at or near zero margin, because serving is a customer-acquisition channel for database, security and application software it also sells. A merchant neocloud has to extract a living margin from the compute spread itself, because it has nothing to cross-sell.

Hyperscalers have a marginal funding is cheap and effectively unconstrained, while a neocloud funds at high-yield or GPU-collateralized rates unless it can pledge an investment-grade offtake contract.

Security culls the tail

One more force pushes the same direction. Model weights are the crown jewels, worth billions and increasingly the core IP of the labs. Handing them to a leveraged third party running multi-tenant hardware is a large attack surface and an IP risk. Security therefore raises the bar and pulls compute either fully in-house or toward a small number of certified, scaled, trusted providers. It is a tax the long tail cannot pay. Ben Thompson’s read on agentic security sharpens the point in a way that helps the integrated players specifically: incentives favor offense, and the only viable defense is for defenders to have access to the best models too. That makes the security layer inseparable from the model layer, and it is a product only a handful of companies can sell.

Which is exactly what is already happening. Thomas Kurian describes labs that use Google’s TPUs and then buy Google’s cybersecurity protection for their models. That is the integrated stack monetized twice off one customer, and it is revenue no neocloud can book. Of the seven buyers, only Google, Microsoft and Amazon own commercial security franchises, which is another reason the integration matrix concentrates rather than spreads.

Conclusion

Fragmentation is held up by Nvidia's channel strategy, a physical shortage, and the speed premium of unbundling. All three will decay. What remains is a bundling cycle turning back toward integration, a widening value-to-cost spread that pays the data-center layer only while supply is short, and a WACC-and-proximity gravity that makes the assets worth more inside the demand-owners. Fewer players will capture more margin on lower-cost-of-capital balance sheets. The neoclouds' job is to build, contract, and be absorbed on the way there.

The structure is a manufactured disequilibrium reverting toward integration, and that reversion is likely over a three year horizon or less. Own it through the consolidators, through the upstream, and selectively through Nebius.

DISCLOSURE

Position: At the time of publication, the author holds a positions in GOOGL and NBIS.

Trading Policy: The author will not materially alter this position within 48 hours of publication. After this period, the author may buy, sell, or otherwise adjust the position without further notice. Changes to the author’s view or position will be reflected in subsequent publications when material.

Conflicts: The author has received no compensation from the issuer or any party with a financial interest in this security.

Forward-Looking Statements: This report contains the author’s opinions, estimates, and projections, including price targets derived from financial models. These are forward-looking statements subject to substantial uncertainty. If these assumptions prove incorrect, the actual value may differ materially, including scenarios of significant loss or total impairment. The price target represents the author’s estimate of fair value under the stated assumptions, not a prediction of where the stock will trade.

This report is provided for informational purposes only and does not constitute investment advice. See the full Terms & Disclosures for additional important information.

The numbers tell a different story, especially in the context of SE’s tactics history and product development to-date. Monee operating margin fell from 29% in Q1’25 to 20% in Q2’26. Provisions cost 4pts of that. Sales and marketing cost 9pts. The S&M line is twice the size of the “problem” most are looking at.

This has a huge impact on perception of the segment not only the next few quarters but in modeling long term margins. Provisioning problems are a loan book problem, but a sales and marketing problem could well be building distribution, or buying it.

Quick take

Sea is buying a second payments network one voucher at a time, and the cost is showing up in S&M… to the tune of nearly $300mm in Q2’26, up a staggering 140% yoy (just like in Q1) and 27% qoq. These are wild numbers, but that’s typical of SE when they’re spending to win a market. Monee S&M has gone from 12% of segment revenue in Q1’25 to 21% in Q2’26. They then achieved off-Shopee surpassing 20% of the SPayLater portfolio and reached 35% in some markets. Two years ago it was under 10%. Active credit users are over 40MM, up 34% yoy.

We’ll call this the acceptance bill, meaning the cash a credit network pays to get accepted where a platform integration cannot place it. Monee has been paying it in cashback on QRIS scans, in 50%-off SPayLater vouchers at Alfamart and Indomaret, and in promoting a standalone wallet app in four countries.

Two channels cost nothing per merchant. Shopify checkout already carries ShopeePay and SPayLater across six markets, and Stripe begins enabling both across Southeast Asia in Q4’26. These integrations can take the place of ongoing voucher spend in an effort to expand distribution. If they work, the acceptance bill should peak in H2’26 and Monee’s return on book recovers. If they do not, Sea is buying a network retail, forever, and our FY30E Monee number may be too high.

Monee in 60 seconds

Monee is Sea’s financial services arm, rebranded from SeaMoney in May 2025. It runs the ShopeePay wallet, the SPayLater buy-now-pay-later product, cash loans under several local brands, insurance, and digital banks in Singapore, the Philippines and Indonesia. It operates in 8 markets across Southeast Asia plus Taiwan and Brazil.

Monee revenue and operating income are primarily from consumer and SME credit. The book stood at $11.1BN of principal outstanding at Q2’26, up 63% yoy, split $10.0BN on-book and $1.1BN channeled through partner institutions. Revenue yield on the average book runs ~53% annualized. NPL90 was 1.0%. It has held there for five quarters.

Monee delivered $1BN of adjusted EBITDA in FY’25 and is ~24% of our FY’30E group number. That makes it the second-largest profit pool at Sea. It is also the one with the least written about it.

Subscribe now

Read more

Moving onto GOOGL… I’ll admit upfront I spent a great deal of time attempting a bottoms-up build. It failed. However, some of the insights do provide value in mapping the path forward for one of the largest CapEx spenders, globally, in history. I found half to 2/3 of the GW they’ve planned depending on the year.

GOOGL does not disclose MW by site, and expansion inside an already-powered campus generates no groundbreaking or press release. The company discloses $85.2BN of leases signed but not yet commenced with no counterparty named. Twelve countries of international buildout publish almost no capacity figures at all.

The verdict. Google Cloud is becoming the majority of Alphabet’s earnings power. This capacity it is being built and deployed into the tightest compute market, and roughly 3/4 of that capacity has no price attached to it yet (best we can tell). You cannot model the outcome with precision. So I created a range of outcomes and mapped them to FCF in 2030. Ultimately, this is what the market cares most about. After the years of massive spend, will investors get FCF back on the balance sheet to do something with.

The TLDR… pay attention to compute contract pricing going forward. If the current path continues, Alphabet could see 25%+ upside to 2030 FCF estimates and trade at the high end of its historical P/FCF range at 30x.

Subscribe now

1. The monster

Start with what has already happened, because the change in this business is easy to miss if you follow the consolidated line.

Revenue growth accelerated significantly and operating margin doubled. Backlog went up 4.8x in twelve months, much faster than revenue growth. Q2’s incremental margin was 54%.

Cloud contributed 5% of operating income in 2024 and 11% in 2025. Our base case models 18% this year, 30% next, and just under half by 2030.

This is the part that matters for the stock. Alphabet spent two decades as an asset light advertising behemoth with a rounding error attached and that will now begin inverting… the compute business is growing at 3-5x the rate of the advertising business.

FCF was -$5.9BN in the second quarter, the first negative quarter Alphabet has ever printed. Capex guidance has been raised three times this year to $195BN to $205BN. Buybacks went to zero for the first time in a decade and the company raised $49.6BN of equity in June.

2. Data center unit economics

A gigawatt of compute capacity behaves like a well with a fixed production profile. You spend the shell cost roughly a year ahead of the servers. You spend the server cost in the year the capacity lights. Then it produces for about six years until the chips are retired and the building is re-racked.

Two years of outflow, six years of inflow. On our base inputs, one gigawatt costs about $39bn for GOOGL and returns about 1.5x that after tax over one server cycle, with a payback in year five.

Now stack them. Build one gigawatt this year, more next year, more the year after.

While the build rate is rising, cash burn accelerates while the model waits for capacity to be fully utilized. The moment the build rate stops rising, the accumulated production overwhelms the current spending.

This is why Alphabet’s cash flow statement looks the way it does, and it is the single most important thing to understand about the name. The burn is not evidence of a problem, at least as far as anyone can reasonably tell today given the demand outlook. The burn is arithmetically what a company looks like when it is adding capacity faster each year than the year before. The question is not whether the flip happens, but how big the number is on the other side, and when.

3. What we could not find

Here is the honest accounting of the mapping exercise.

Take Alphabet’s capex, strip out the ~8% that is offices, split the rest 60/40 between servers and buildings the way management has described it consistently since late 2025, divide by a cost per gigawatt, and phase it into service. That is the top-down count, and it says Alphabet turns on ~1.9 GW this year, 3.8 next year and 5.4 in 2028 after applying a six-month delivery delay.

The register covers 43% of 2027 and 79% of 2028, and roughly half of the five-year total. The single cleanest illustration of why is Anthropic. Broadcom’s filings put roughly 3.5GW of next-generation TPU capacity for Anthropic beginning in 2027, and on the Q2 call Broadcom said Anthropic is on track to become its largest custom accelerator customer next year, ahead of Google itself. Press reporting derived from the Financial Times says Google backstops 10 TPU data center projects totaling 2.4 GW. We have addresses for three of them.

Five things remained unmapped. Expansion inside existing campuses, which could be the largest figure and which generates no public record at all beyond the occasional utility filing. The $85.2BN of leases not yet commenced, which is a disclosed number attached to no disclosed tenant, implying roughly 3.2GW of leased shell. International capacity across twelve countries where megawatts are almost never published. Densification, because an Ironwood rack draws far more per square foot than the rack it replaces. And the unnamed Anthropic sites.

One honest counterweight. Some of the gap is our top-down being generous rather than the bottom-up being sparse. Alphabet’s capex includes turbine deposits, transformer prepayments, land and Broadcom supply commitments that will not light anything for years. Inventory went from $2.4BN to $10.0BN in six months, and TPUs built for sale to third parties are not Google’s own capacity.

And the timing is uncertain even for the buildings we can name.

That is a 3x spread on cumulative named capacity by 2029, using the same set of projects. Texas ordered a full audit of every data center in the ERCOT interconnection queue early August and suspended large load classification, with completion targeted for December. Indianapolis has a moratorium through the end of 2027. St. Charles passed a permanent ban. Pennsylvania removed data centers from fast-track permitting. Gas turbine lead times run 5-7years, and GE Vernova’s book is sold out into 2030. Sightline Climate found that 26% of expected 2025 capacity slipped and projects 30-50% of 2026 capacity will.

We handle this by carrying capex in full in every scenario and pushing half of each year’s newly lit capacity into the following year. If Alphabet delivers on schedule our capacity is too low and our cash flows are too low. If it slips, we have already paid for it.

4. Pricing

I spent much of my effort on the physical build and it was the wrong place to spend it.

Run the model changing one lever at a time and measure the effect on five years of cumulative free cash flow. Delivery timing, the thing the entire site-by-site exercise was built to pin down, is worth +/- $32bn. The price of a gigawatt-year of compute is worth $137bn either way.

Pricing matters 4x more than the construction schedule. So we went and built the pricing comps properly, across every deal with disclosed dollars and disclosed capacity.

The first thing to understand is that there are three different things being sold and they differ by a factor of thirty.

A powered shell lease is a building with electricity and cooling in it and nothing else. The tenant brings the chips. Core Scientific leases to CoreWeave at roughly $1.44mm per megawatt-year. Applied Digital is at $1.67mm to $1.83mm, Cipher at $1.22mm to $1.79mm, TeraWulf at $1.86mm to $2.26mm. Call it $1.2bn to $2.3bn per gigawatt-year for real estate with power.

Full stack compute is the same building with the operator’s chips in it. CoreWeave’s second quarter revenue against 1.5GW of active capacity annualizes to roughly $6.9bn per gigawatt-year blended across its fleet. Microsoft’s contract with IREN is $9.7bn over five years for 200 MW, or $9.7bn per GW year. Microsoft with Nebius is $11.6bn. Nebius disclosed its own baseline entering 2026 at about $12bn.

The gap between those two layers, roughly seven times, is almost entirely the depreciation of operator-supplied silicon.

The second thing is that the price moved in 2026, and it moved a lot.

Nebius disclosed the clearest ladder anyone has published. Its baseline entering 2026 was about $12mm per MW year. Long term contracts signed in the second quarter were struck at $20-25mm. Short-term capacity in the third quarter went at $40-50mm per MW year, and sometimes above. That is $40-50BN per GW year for capacity of six months or less. Its maiden capacity auction cleared 15% above any price it had previously achieved on Blackwell, and payback on new capacity compressed from two to three years down to one year and ten months.

CoreWeave put through a roughly 25% price increase across every SKU in July and told investors near-term capacity is effectively sold out, with contribution margins on new contracts running 5-10pts above those signed in prior quarters. SemiAnalysis’s one year H100 contract index rose about 40% between October 2025 and March 2026, on a chip two generations old. Nvidia’s chief financial officer said H100 rentals were up about 20% year to date. Bernstein put IREN’s newest contracts 20-25%above its earlier ones.

The third thing is where Google sits on that chart, and it is not where you would expect.

The reported Anthropic arrangement, $200bn over five years across roughly five gigawatts, works out to about $8-9bn per GW year at full ramp. That is the largest private compute contract ever signed and it is priced near the bottom of the observable range, because it was negotiated in 2025 at 2025 rates by the largest single buyer in the market.

Published sellside models for Google Cloud imply $11-14BN of incremental revenue per GW in 2027. Morgan Stanley’s own framework models Google at $15 per watt in 2027 rising to $18 in 2028, and notes explicitly that this is below what it calls an industry negotiation level above $20bn per Gw.

So consensus is modelling Google’s compute at roughly half of what the merchant market is currently charging.

And then there is Google’s own price list, which nobody seems to be reading.

Google publishes TPU pricing. Ironwood, the current generation, lists at $12.00 per chip-hour on demand, $8.40 on a one-year commitment and $5.40 on three years. Convert that at roughly 714,000 accelerators per gigawatt, which assumes 1.4 kilowatts all-in per accelerator including networking and cooling overhead.

Google’s own three-year committed list price implies $34bn per GW at full utilization, or $24bn at 70%. Our base case assumption of $18bn sits below every rate Google publishes, including the cheapest one after a 30 percent utilization haircut. SemiAnalysis’s estimate of what Anthropic actually pays, about $1.60 per chip-hour, implies $10bn, which tells you how large the discount is to the single largest customer.

We are not claiming Google will realize list price, nobody realizes list price. We are pointing out that the model assumes Google captures roughly half of its own published three year committed rate on capacity it has not sold yet, and that this is a conservative assumption rather than an aggressive one.

5. The window

Put the two halves together and the setup becomes clearer than either half alone.

8% of Google Cloud’s capacity is unpriced this year. By 2028 it is as much as 55% and by 2030 it is 73%. The legacy book and the Anthropic and Meta contracts are locked at rates struck in 2025 and earlier. Everything lit after, we believe gets priced at whatever the market clears at on the day it is energized. And it is being energized into the tightest compute market anyone has seen. SemiAnalysis counts roughly 20 GW energized industry-wide in 2026 and expects about 30 in 2027, against demand that Morgan Stanley models as a 38 to 49 GW shortfall through 2028. Texas has paused its interconnection queue pending an audit. Half a dozen counties have moratoria.

Every one of those constraints is a reason our delivery timing might be too optimistic. Every one of them is also a reason the price on delivered capacity goes up. If power scarcity delays Alphabet’s buildout, the same scarcity raises the rate it can charge on whatever it does deliver. The two things that make us wrong on timing make us conservative on price. They do not compound in the same direction.

One thing does compound, and it is the actual risk. If demand cracks, then delivery is easy, pricing collapses and utilization falls, all at once. That is the scenario where this does not work, and no amount of construction mapping would have told you about it.

The bear case on pricing deserves a fair hearing because it is real and it is loud. Spot rental rates for H100s have fallen roughly 57% from their early-2024 peak. AWS cut its P5 on-demand list price 44% in June 2025. Inference cost per token is collapsing at something like an order of magnitude a year. TD Cowen’s channel checks found Microsoft cancelling or deferring data center leases, which the firm reads as a potential oversupply signal.

Every one of those is about a different thing than what Google sells on a multi-year contract. Falling spot prices in 2026 coincided with contract prices rising 40% because serious buyers migrated out of spot and into committed capacity as on-demand sold out. Per-token deflation is a model efficiency gain, and Google itself has disclosed cutting Gemini serving costs 78% in a single year, which raises the output a GW can produce rather than lowering what the gigawatt rents for.

6. Why pay Google more

There is a piece of this that does not show up in any comps table.

Alphabet is a company with $155bn of cash, effectively no net debt, an Aa2 rating and a stable outlook, selling a multi-year compute contract. The merchant alternatives are, in several cases, former bitcoin miners financed through SPVs with private credit, residual value guarantees and, in a number of the largest deals, a credit backstop from Google itself. Alphabet discloses $43.8bn of maximum potential payments under credit derivatives backstopping data center payment obligations, plus $7.6bn of guarantees for suppliers buying long-lead power equipment and $24.1bn more agreed but not finalized.

If you are a chief information officer signing a five year commitment for capacity your business now depends on, the identity of the counterparty is part of what you are buying. So is the security posture, where Alphabet has bought Wiz and Mandiant and folded them into the stack. So is distribution, where Google reaches both the consumer surface and the enterprise, and where Thomas Kurian has said nine of the ten largest AI labs are already customers, that customers using AI products consume 1.5 times as many Google Cloud products as those who do not, and that customers who sign commitments routinely over-attain them.

7. The inversion

So I stopped forecasting bottoms-up and asked the only question that can be answered.

Alphabet’s market capitalization at $335 is $4.1tn. Take out $57bn of net cash, the entire securities portfolio at reported value ($101bn of SpaceX stock and $131bn of non-marketable holdings dominated by the Anthropic stake). What remains is a core business priced at $3.8tn.

At 25x 2030 FCF (roughly the 10 year average), you need visibility to ~$150bn of FCF.

Now run the model at every price point in the observable range, changing nothing else.

At $8bn per GW year, which is the reported Anthropic rate and the cheapest price anyone has paid for compute at scale anywhere, Alphabet produces $165bn of FCF in 2030 and the stock trades at 23x.

At consensus pricing of $11bn, it is $191bn and 20 times. At our base case of $18bn, half of Google’s own three year committed list rate, it is $253bn and 15x. At the rate Nebius is signing long-term contracts today, it is $315bn and 12x.

The bear case…. where pricing is low and costs are high and delivery slips and chip life proves to be five years and the installed base was smaller than we thought, produces $50bn of FCF in 2030. That is 77x and it does not clear anything. The bull case produces $612bn, which is 6x.

So the width of the distribution on a five-year view of a company spending $1.56tn is enormous, and anyone quoting a point estimate on this is not being straight with you. What the inversion tells you is narrower and more useful. No single variable breaks this. The variables have to fail together. And the two most likely failure modes, power scarcity and slow permitting, push price up while they push volume down, which is a partial hedge rather than a compounding loss.

For what it is worth, our base case has the hole deeper and a year later than consensus. Minus $65bn in 2027 against consensus at minus $19bn, almost entirely because we carry 2027 capex at $330bn, or as much as 10% above consensus. Run the operating model on consensus and the 2027 gap narrows to about $11bn.

8. Funding

Net debt peaks at $76bn at the end of 2028, which is 0.2x EBITDA, and is gone during 2029. Peak incremental borrowing across the whole program is $37.5bn. Moody’s affirmed Aa2 stable in May and S&P AA+ stable, with S&P forecasting adjusted debt to EBITDA of 0.2-0.3x.

That bridge excludes $101bn of SpaceX stock and $131bn of other securities. Include the SpaceX position at reported value and Alphabet never crosses into net debt at all. The 551 million SpaceX shares carried at $94bn cost roughly $900mm in 2015.

The funding question here is genuinely uninteresting. The $40bn at-the-market program authorized in June is unused. The preferred converts to equity in May 2029. Alphabet could fund the entire remaining burn by selling a third of a position it did not know it needed.

9. What to watch

The value of the site-by-site work is not that it produced a number. It is that it produced a list of things that resolve on dates, which converts an unmodelable company into a monitorable one. Here is a long list of items to watch into Q3 earnings and beyond.

The three that matter most are the 2027 capex guide in February, because it settles whether the spending path is real, any disclosure of external TPU pricing, because it settles the repricing question, and the direction of merchant contract rates, because that is the highest-frequency read on whether the window is open or closing.

10. Conclusion

It is unlikely that anyone outside of perhaps SemiAnalysis can model this business with any precision. And even then, there are human factors that are unquantifiable here. However, it is indisputable that Google Cloud is becoming the majority of Alphabet’s earnings power on a timeline that is short, and the market has not had to price that because it is happening underneath the worst-looking cash flow profile in the company’s history.

The capacity being built is landing in the tightest compute market on record, and roughly three quarters of it has no price on it yet. Contracts struck in 2026 may clear at as high as 2-4x the rate embedded in published models, and Google’s own three year committed list price implies double what the models use. Every observable failure mode on the physical build makes the pricing better.

Furthermore, the full deltas diagram below shows the sensitivity in the model to all of the variables, from bear case to bull case. The market should be more worried about advertising risk downside, however that downside reflects a dramatically different macro environment to be fair.

Alphabet is one of a very small number of counterparties that can credibly sign a 10 year compute commitment, and across a large share of the merchant market it is already the credit standing behind somebody else’s lease.

And when you invert the question rather than forecasting it, the price at $335 requires roughly $153bn of FCF in 2030 against a model that produces $165BN at the cheapest price anyone has ever paid for compute and $253bn at half of Google’s own list rate.

While we should all monitor the construction schedule, we really care most about the moment the market decides the cash flows are coming, because that is when a business growing operating income at this rate, with half of it not yet priced, stops being valued on the burn and starts being valued on the ramp in FCF possible beyond.

Appendix

Four assumptions that matter

Everything in this note reduces to four levers. The rest is noise at five years' distance. Each is shown with the observable evidence rather than a point estimate, so a reader can substitute their own.

Named capacity register

Thirty-three projects. Structure A is an owned shell with Google chips, B is a leased shell with Google chips, C is a third-party site where Google sells TPU systems rather than renting capacity. Megawatt figures for owned campuses are estimates, because Alphabet discloses none. Both tables are downloadable images.

Pricing comps

All figures are total contract value divided by term and by capacity, which is a bookings average and not a steady-state realized rate. Layer matters more than any other adjustment.

What we couldn’t resolve

1. The Anthropic and Meta contract terms are press reports and neither company has confirmed them. We do not know whether either carries an annual escalator, whether they are leases with a rent schedule or minimum spend commitments, or how capacity is measured for billing. We model Anthropic flat at $8.5bn per gigawatt-year and flag that the contract form itself is unknown.

2. The counterparty, megawatts and term of the $5.8Bbn short-term lease commencing in the third quarter of 2026 are undisclosed. We model roughly 0.3 GW of full-stack compute over twelve months.

3. Installed lit capacity at the end of 2025 is not disclosed anywhere. Our 8.5 GW starting point is triangulated from cumulative capex.

4. Segment capex and segment depreciation are not disclosed and are allocated by lit-gigawatt share. This is the largest single judgment in the model.

5. The $707bn purchase commitment figure comes from secondary sources quoting Note 10 of the second-quarter 10-Q rather than from the note text directly.

6. Owned-campus megawatt figures throughout Appendix B are estimates. Only the Structure C sites carry disclosed megawatts.

7. The 2Q26 net income of $112bn includes a $99bn unrealized gain on Anthropic and SpaceX marks. Nothing in this note is built on it.

8. The Koyfin consensus free cash flow line does not equal the same screen’s cash from operations less capex, with a gap of about $39BN in 2029. Both are shown where relevant.

9. The backlog series has a definitional break. The first quarter of 2026 added contracts with an original term of one year or less, worth about $7.3bn. The move from $240bn to $462bn is 97% bookings.

10. Conversions from chip-hour prices to revenue per gigawatt-year assume 714,000 accelerators per gigawatt at 1.4 kilowatts all-in. That assumption drives the result and is stated wherever it is used.

DISCLOSURE

Position: At the time of publication, the author holds a position in GOOGL.

Trading Policy: The author will not materially alter this position within 48 hours of publication. After this period, the author may buy, sell, or otherwise adjust the position without further notice. Changes to the author’s view or position will be reflected in subsequent publications when material.

Conflicts: The author has received no compensation from the issuer or any party with a financial interest in this security.

Forward-Looking Statements: This report contains the author’s opinions, estimates, and projections, including price targets derived from financial models. These are forward-looking statements subject to substantial uncertainty. If these assumptions prove incorrect, the actual value may differ materially, including scenarios of significant loss or total impairment. The price target represents the author’s estimate of fair value under the stated assumptions, not a prediction of where the stock will trade.

This report is provided for informational purposes only and does not constitute investment advice. See the full Terms & Disclosures for additional important information.

Sea Limited (SE) has entered “investment windows” 4 times in 9 years with a predictable impact on the stock. We are now emerging from this cycle with shares ~40% from the September 2025 high, and management has done something it never did in the first three. It has put a date and a dollar figure on the exit.

Pitch in 1 paragraph…

Shopee’s EBITDA margin on GMV troughed at 0.55% in 4Q25 and has risen two quarters running to 0.67%. Management guided FY’26 Shopee adjusted EBITDA to $1BN on the Q2 call, which needs $522MM in 2H26 against $478MM in 1H26, and it reaffirmed a 2% to 3% medium-term margin target with a three-part bridge. Our own bridge says advertising will add 128 of the 194bps between here and FY30, and the ad line is compounding at +90bps of take rate a year. Our base case sees consolidated adjusted EBITDA growing from $3.9BN in FY26 to $11.85BN in FY30. That is 3x in 3 years. The stock trades at 13.6x CY27 EBITDA. And perhaps most importantly, this mix shift away from less certain Free Fire EBITDA should provide long-term support to valuation.

Put more simply… SE spends into a competitive or structural opening, the equity market prices the spend as permanent, and the business exits the window with more share and a higher profit base.

Why the opportunity exists

The market is pricing the current spend as a tax rather than an investment, and it has good reason to be skeptical. The 2% to 3% target has been reaffirmed on every call for two years while the margin fell from 0.92% to 0.55%. Hayden Capital, which runs SE as a top-two position, frames the whole debate as reinvestment versus defense and puts the bear’s terminal margin near 1% of GMV against management’s 2% to 3% [C]. At $119 the stock discounts roughly that bear, and we show the back-solve below.

Two things have changed since the March guide that the price has not caught up with. The margin turned in 1Q26 and confirmed in 2Q26. And the composition of the ad business inverted, from a fixed base of sellers spending more to a broadening base of sellers, which is the only version of the ad story that reaches 5% of GMV without hitting seller economics.

SE in 60 seconds

3 businesses run from of Singapore with local market teams.

• Shopee is the largest marketplace in SEA and Taiwan with $127BN of GMV in FY25 and a profitable and fast-growing Brazil business

  • Revenue is commission on transactions, ads sold to sellers, and logistics fees, FY25 Shopee revenue was $17.0BN and adjusted EBITDA $881MM

• Monee is the consumer credit and payments arm

  • Lends to Shopee buyers and, increasingly, off-platform through SPayLater at physical merchants

  • Loan book $11.1BN at 2Q26, up 62.5% yoy, with 90-day NPL at 1.0% and FY25 adjusted EBITDA of $1.0BN

• Garena publishes Free Fire, a nine-year-old mobile shooter with more than 100MM daily active users

  • Bookings were $2.9BN in FY25 and adjusted EBITDA $1.7BN, it funded the first two windows, and it still carries 46% of segment EBITDA

Evidence

From Dec 2021 to Aug 2023 the stock fell 85%. Over the same 20 months Shopee’s share of the tracked marketplace pair rose in Indonesia (49% to 67%), Malaysia (60% to 69%), Vietnam (31% to 61%), Thailand (24% to 41%), Singapore (42% to 51%) and Taiwan (67% to 71%). The Philippines fell 25pts between Oct 2022 and Mar 2023 as TikTok Shop scaled, and had recovered the entire loss by July 2023.

The 2022 collapse was rising rates and Covid pull forward coupled with strategy reversal. It was never a competitive event (had the data that proved so at prior fund), though the Tik Tok fear at the time was real and growing.

Lazada was supposed to buy share (funded by BABA) as they Lazada raised discount penetration from 23% of orders to 78% between 2019 and 2023 and lost 7pts of regional share. Shopee gained 23pts over the same window while its free-delivery penetration fell from 54% to 51%. Announced competitor spend buys no measurable share.

We provide the above prices as a range of outcomes for the next few years. They are not meant as price targets, rather to help a reader frame the opportunity set under different assumptions. Those assumptions are spelled out in detail below the paywall.

Subscribe now

Read more

TLDR… storied history of building businesses over decades with key mentors who helped shape his now empire that expands well beyond MAIR to his own private equity firm. Now a buy-and-hold style investor, it’s no wonder he agreed to inject hundreds of millions to help close MAIR’s recent transformative transaction of ebm-papst. His quiet $9bn stake in MAIR has gone largely undiscovered, it seems, though it stacks up against some of the greatest tech founders of our time…

Ownership

As of 30 June 2026, Larry Gies controlled ~337MM shares of MAIR. They sit across three vehicles. Madison Industries Holdings LLC, Madison Air Co-Investors LLC, and a private placement made at the IPO price. That’s 67% of the economic interest and 95% of the voting rights, worth ~$9BN. For scale, Marc Benioff’s stake in CRM runs ~$6.5BN, Tobi Lutke’s in SHOP $12BN, Brian Chesky’s in ABNB $12BN.

The gap between the two percentages is built on 10 votes per Class B share. The Schedule 13G filed August 10th states the chain plainly, “Larry Gies is the sole manager of Holdings. Holdings is the sole manager of Co-Investors.”

Small town Illinois

Gies grew up in Mendota, Illinois. Population ~7K, 90 miles west of Chicago. His father worked at Conco, a local manufacturer.

Conco was owned by Michael Heisley, a Chicago industrialist who founded Heico Companies, a decentralized holding company assembled from broken manufacturers. Heisley later bought the Vancouver Grizzlies and moved them to Memphis. He’s the single most important figure in this story, and he’s almost entirely absent from coverage of MAIR.

Gies’s first job was at Touche Ross. It ended at 24, after he wrote a memo criticizing how the firm handled its largest client. The story is usually told as a firing over a memo sent to a client’s CEO. His own account is narrower. His words are that “my days in my first job were numbered,” and the memo was internal.

He then went to work for Heisley. By his own description he ran one of Heisley’s manufacturing companies “quite poorly.” The instruction he received, and repeats, was “Go do it, kid.”

That apprenticeship is where the operating model comes from. A decentralized holding company that keeps what it buys. He founded what became Madison Industries in 1994, at 27, on 10 maxed-out credit cards and a $50,000 loan from a high school friend. His wife Beth, a former teacher, worked three jobs.

The eulogy, and when it actually happened

The story Gies tells most often is the eulogy exercise, from his University of Illinois commencement address of 11 May 2019.

“So that day I got our entire management team together and asked them to do something a bit unorthodox. I had them write their eulogy. Not their obituary, but their eulogy... My teammates spoke of family, making a positive impact on the world and doing good. But when they spoke specifically about how they wanted to be remembered for their professional lives, nobody mentioned making a lot of money for our investors.”

“That day we formed Madison Industries. Which meant that our highly successful business of buying and selling companies was headed for the trash heap.”

In a 2018 speech he put a date on it. “I was forty-five before I found my why.” That would place the turn around 2011, not 1994.

Madison Industries was founded in 1994 as Madison Capital Partners, a fundless sponsor. And a 2006 profile in The Deal describes a firm advertising a 90% net internal rate of return across 12 exits and five partial realizations.

So the hold-forever philosophy dates to 2011, not to the founding. Todd Bluedorn, former Vice Chairman of Madison Industries, described the earlier version in May 2026. “Man, the guy can sniff markets. He buys on the cheap, he sells at the peak, and when he buys, it’s from corporations who are running and he sells them to corporations who have more money than they know what to do with.”

Professionalizing the business

Frank Ptak spent ~30 years at Illinois Tool Works (ITW), rising to Vice Chairman, and Madison’s own materials credit him with helping pioneer ITW’s 80/20 model. He then ran Marmon Holdings for 13 years. More than 100 autonomous businesses and $8BN of revenue, sold to Berkshire Hathaway. He reported to Warren Buffett. He also ran Marmon’s acquisition of Duracell in 2014, then joined Madison Industries as Senior Managing Director in January 2019.

He wasn’t the only one. Gies has assembled an unusually senior bench at the private parent.

George Nolen arrived first, in October 2010, as Vice Chair. 27 years at Siemens, the last five as President and CEO of Siemens Corporation. He also chairs Filtration Group, another Madison company, and now sits on MAIR’s board.

Todd Bluedorn joined in April 2022, also as Vice Chair. 15 years at Lennox International, CEO from 2007 and Chairman from 2012, over which market capitalization went from $2.5BN to more than $13BN. Before Lennox he was President of Otis Elevator Americas and President of Carrier North American HVAC at United Technologies, and McKinsey before that, from 1992. West Point, and a US Army Combat Engineer Officer from 1985 to 1990. He’s the man quoted above.

Gies on hiring him. “Todd attracts world-class talent, builds high-performing teams and brings a successful track record of building best in class global businesses.” Bluedorn on why he came, and note what he names first. “The exceptional Madison team and its mission of making the world safer, healthier and more productive attracted me to this opportunity.”

There’s a recruiting channel visible in the dates but stated nowhere. Quan Nguyen, now MAIR’s Group President for Air Quality & Home, was VP and GM of Lennox Industries Residential from 2017 to 2024, under Bluedorn for most of it.

The lineage we can now see… Heisley supplies the shape in the early 1990s. Gies runs it for 15 years as a buy-and-sell operation. The identity changes around 2011 to holding. Ptak arrives in 2019 to professionalize the machine, immediately before the air platform is scaled through the $3.6BN Nortek acquisition.

The acquisition record

Madison Industries has done 80+ acquisitions since 2004, across 20-plus countries. MAIR specifically claims ~$8BN invested and ~$28BN of TAM unlocked since 2021. The air platform itself was assembled in nine years.

Madison publishes margin expansion figures for three deals. That’s the closest thing to a track record on the operating model. +975bps at Therma-Stor in Nov 2017. +1,700bps at Specified Air Solutions in Oct 2018. And +500bps at Broan-NuTone after the Nortek acquisition in Apr 2021. The company states the majority of those gains landed inside the first three years, and more than half above the gross margin line.

The anchor deal was Nortek Air, announced 19 April 2021 and bought from Melrose Industries plc at a reported $3.6BN. It brought ~6,000 employees and the Broan, NuTone, Reznor, StatePoint, Zephyr and Huntair brands. Gies at the time. “The global pandemic has put a spotlight on the importance of indoor air quality.”

Two things about how they buy are worth more attention than they get… they don’t call it integration. Wyant, on the 1Q26 call. “we don’t really call them integrations. We call them transitions... we buy businesses to allow them to continue to do what they do.” For a serial acquirer that’s a real statement of intent. It’s also consistent with a 45-person center that couldn’t run integrations even if it wanted to.

And speed is a stated value. Madison Industries publicly claims it has never changed an acquisition price after agreeing it, and that acquisitions close within 30 days, with one counterparty testimonial citing 15 business days. This is very attractive to sellers.

The green line on that timeline matters too. In October 2024, mid-assembly, Madison sold Nortek Global HVAC to Rheem for $1.15BN and said very little publicly about why. A company that holds forever is still a seller when the asset doesn’t fit. That’s defensible. But it isn’t the story on the slide, and it’s the fact that gives Bluedorn’s quote its edge.

The parent you are not buying

MAIR is one of seven verticals inside Madison Industries. Filtration, medical, safety, energy, industrial solutions, air, and Seakeeper. The parent reported ~$5BN of revenue in 2019 with 10,000 employees and 178 facilities. It has since ~doubled on both counts. Madison Capital Partners, the capital vehicle, is separately described as a registered investment adviser managing ~$4.8BN across 13 funds.

Two things follow for a public-market investor. The governance of the listed company is set by a private organization with considerably more at stake than the listed entity. And a Transition Services Agreement with Madison Industries International keeps supplying services post-IPO. Pre-listing, allocated corporate costs ran $10.0MM and $5.8MM in the periods shown in the 2Q reconciliation.

Decentralization

The prospectus contains the sentence that makes the model concrete.

“We remain lean at the center with approximately 0.5% of total headcount.”

Against ~8,800 employees, that’s ~45 people running a $3.9BN revenue company. Forty-five. A typical industrial of this scale carries several hundred at corporate.

The CFO described the mechanism at the William Blair Growth Stock Conference on 4 June 2026, and it’s the clearest account anyone at the company has given.

“I think about the Nortek Air Solutions business, where it was kind of run as one blob when we first bought the business. Now we’ve broken that out into the air handling and the direct expansion in the clean room, clean room breaks out into life sciences, into semicon. Now we’ve got, instead of a centralized commercial team, we have dedicated teams coming to work every single day thinking about how they serve those customers in that end market.”

That’s 80/20 applied to org design rather than to product lines. Gies described the same instinct in the 2006 profile, long before any of the current messaging existed.

“I think that is why we have a hard time keeping people in the office. They all want to be on the plant floor.”

Where the decisions actually sit, from the filings and calls.

One caution on the branding. The decks name the operating system MAXX, Madison Air Execution Excellence. It appears once in the public record once… a strategy-pillar bullet reading “Madison Air Execution Excellence (MAXX) powered by 80/20.” That’s the entirety of the disclosure.

The nearest thing to observable machinery comes from Wyant’s 1Q remarks. More than 60 leaders put through 80/20 and sales training courses in a few weeks. Immersive AI-enablement sessions. And an enterprise-wide innovation summit she hosts at least twice a year.

The published values are short enough to quote in full, and two of them are testable. Trust, “doing what you say and saying what you do.” Bias for Action, “an organizational reflex to own problems and resolve them quickly.” Entrepreneurial, “operating with an owner’s mindset, emphasizing accountability, humility, and continuous improvement.”

Compensation

The compensation design is what makes a 45-person team possible. Executives are paid in long-dated Equity Appreciation Rights with single-cliff vesting. Jill Wyant held 3.7MM shares at IPO, of which 2.6MM were EAR units. She’s now at 3.5MM after two tax-withholding events. JJ Foley holds 1.7MM including 0.6MM EAR units.

The instrument is defined in Wyant’s Form 3 as “the right to receive one share of the Issuer’s Class A common stock upon vesting.” The cliffs are long. The clearest published example is the new chief accounting officer, hired June 2026. Base salary $355,000, target bonus 50%, and ~$700,000 of EAR units vesting in a single cliff on 15 July 2031. Five years, all or nothing.

It’s also broad. At the IPO, 10,339,435 Class A shares went to company executives and employees, and 146,556 EAR units went to former employees and consultants. They carried alumni. At the private parent, Madison Industries describes ~1,000 “owner managers” and maintains a named corporate function called Vice President, Owner Relations.

This is where an accounting item from the first piece stops being a technicality. The $125.9MM equity-appreciation charge Madison adds back to adjusted EBITDA is the price of running a company with 45 people at the center. It’s a real compensation cost. Excluding it is what takes the reported 2025 margin from an honest 23.0% to the 26.6% that anchors the premium-platform framing.

Board of Directors

One structural fact follows directly from all of the above, and we haven’t seen it written anywhere.

The entire board is four people. Four. Hudson La Force and George Nolen joined on 15 April 2026, the day before the IPO. Gies has chaired since October 2017.

Both directors labeled independent have long-standing Madison relationships. Nolen has been Vice Chair of Madison Industries since October 2010 and chairs Filtration Group, another Madison company. La Force sat on Filtration Group’s board from 2022 to 2026. Independence here is a NYSE definition.

And Gies personally chairs the Compensation and Nominating Committee. The founder holding 95% of the votes also chairs the committee that sets executive pay and nominates directors. La Force chairs Audit, which is the one committee where the rules leave no discretion. Directors are paid in the same instrument as management. Form 4s dated 29 June 2026 show La Force granted 5,711 EAR units and Nolen 5,077.

For a certain kind of investor this is the attraction, not the objection. Owner-operators with real capital at risk and no quarterly-earnings reflex have compounded capital better than committees for a century. Gies has ~$9BN of his own money in the same security, plus $620MM more bought in last week’s placement at the market price. But be clear about what you own. A private holding company that has sold a third of its economics and none of its control.

Ownership figures are from the Schedule 13G filed 10 August 2026 and Form 3 and 4 filings. Quotations are as delivered and dated in the text. The third piece will cover the products, the pricing and the ebm-papst acquisition. Not investment advice.

DISCLOSURE

Position: At the time of publication, the author holds a position in MAIR.

Trading Policy: The author will not materially alter this position within 48 hours of publication. After this period, the author may buy, sell, or otherwise adjust the position without further notice. Changes to the author’s view or position will be reflected in subsequent publications when material.

Conflicts: The author has received no compensation from the issuer or any party with a financial interest in this security.

Forward-Looking Statements: This report contains the author’s opinions, estimates, and projections, including price targets derived from financial models. These are forward-looking statements subject to substantial uncertainty. If these assumptions prove incorrect, the actual value may differ materially, including scenarios of significant loss or total impairment. The price target represents the author’s estimate of fair value under the stated assumptions, not a prediction of where the stock will trade.

This report is provided for informational purposes only and does not constitute investment advice. See the full Terms & Disclosures for additional important information.

Arxis, Inc. (ARXS) is an explicit “Buy-Down Machine.” As CEO Kevin Perhamus puts it, “[we] grow the EBITDA of the acquired business at a higher rate than the base of Arxis in the next three years”, and can it “buy down the multiple to less than 10 times within 36 months”. Thirty-five acquisitions in ~6 years. The two most recent large deals were disclosed together at ~12x forward EBITDA.

And what comes after is even more important… ARXS printed 21% organic growth and a 42% adjusted EBITDA margin in 2Q26, its second quarter as a public company, and the shares have ~doubled from a $28 April IPO. Given it’s market cap of ~$22bn, this business is entering what we believe is a sweet spot of scale in its lifecycle… large enough to do $1-5bn transactions but small enough so that these actually matter. We’ve seen too many large acquirers/compounders suffer from scale, and this is an opportune time to look at one with nice runway ahead.

Overview

ARXS designs and manufactures proprietary electronic and mechanical components for harsh environments. Connectors, RF and microwave devices, sensors, capacitors, self-lubricating bearings, seals, ducting, radar-absorbing materials. To cut to the chase…

• ~90% of revenue is proprietary, sold to >600 platforms for more than 5k customers

• Top 10 customers 36% of revenue, none >7%

• Defense and space 46%, commercial aerospace 22%, industrial technology 31% (includes med devices, surgical robots, semiconductor equipment and factory automation

• Parts typically cost less than 0.01% of the platform… too cheap to fight about, too critical to remove… qualified years before production, usually as the only approved source, on platforms that produce for 20 years and fly for 40

• Platforms are diverse… F-35, DDG-51, Virginia-class, SPY-6 radar, missile-tracking satellites, 737 and Airbus A320, the LEAP, CFM56, PW1000G and PT6 engine families, Gulfstreams G650, G700 and G800, and in medical the contacts are built to survive 15 years inside the human body

So the budget line that matters is R&D, not procurement. A designed-in supplier wins its next positions during platform development, and the Pentagon’s RDT&E request stepped from ~$143BN in FY25 to $179BN in FY26.

In the company’s own words, positions are won because “our engineers meet with our customers’ engineers. They solve a problem which results in a custom product that gets designed into the bill of materials, typically as the only qualified part, and then stays there for decades.” ARXS owns 67 foundational technologies across 48 business units. A good example of this is the KAron bearing liner, which turns a standard bearing into a custom product with “five to ten times the value.” KAron went onto the Gulfstream G650 in 2008, then spread to five more models. It produced “approximately a 965x increase in our revenue from this product line and customer.”

Two structural facts about ARXS that set this business apart from average roll-ups…

First, they have an actual operating system that everything together called Arxis EDGE which also enables a scaled and decentralized model. EDGE is a purchase-order-level system that digests 6k orders / month. ARXS runs 48 decentralized business units across ~70 focus factories under 16 block leaders. That’s why 2026 was guided in a $20MM revenue range and then raised by $100MM a quarter later. By late July, 95% of 2026 revenue sat in firm orders, with “another $150 million or so left to go to book for the year.” More than 450 of the company’s ~5,750 employees are paid on new-business wins, and the EDGE cross-sell pipeline grew from ~50 opportunities in 2021 to more than 2,000 by mid-2025… so they have a large proportion of employees who are highly incentivized to grow volumes (on which they’ll surely take ongoing price) and track every order. This is important as it means management can quickly triage with problems before they become large.

Second, they actually decompose growth… their algorithm is MSD new business, plus MSD price, plus whatever the market gives. They put up 17% organic in 1Q26 and 21% in 2Q26, and the entire acceleration was the market bucket… “new business remains at mid-single digit. The price that we’re gaining remains at mid-single digits. All of the delta that you’re seeing is in the base business.” So the long-term guide is high-single-digit organic growth, converting at ~50% incrementals into low-double-digit EBITDA growth, before any acquisition.

Adjusted EBITDA margins should land near 40% for 2026, up 3pts YoY. High teens FCF margins in 2026 will help reduce net leverage toward 1.3x from 4.2x at IPO. CapEx runs ~3% of sales. This is one of the nicest compounding models I’ve come across at what seems to be a sweetspot… wish had finished the work sooner! Share are almost 15% off their highs.

We provide the following targets as a range of scenarios overtime. These are not meant to be explicit price targets, more so they provide a framework for where the stock could across different sets of outcomes. The details for all of this, and more, are below the paywall. Despite the run-up post IPO and large headline multiple, there remains a nice opportunity ahead, and the 13F scrape suggests this is genuinely underowned the funds that would typically be involved in a name like this.

Read more

The pattern is easy to admire five years after it becomes apparent and when the record is public and the multiple agrees. It is only ever available at a reasonable valuation when something breaks the price without breaking the machine.

Madison Air Solutions (NYSE: MAIR) came public in April at $27 and just broke deal price. (See our prior work on this dynamic here… shouldn’t be entirely surprising is the takeaway link.) Shares peaked in the first few months at 27x NTM EBITDA - expensive by any measure. Add a softening macro on numerous fronts and higher rates and a BIG acquisition with equity to be raised alongside 4.0x net leverage and it’s no wonder why the stock is now ~45% off its highs. Then, add a restatement the market is anxious about was disclosed in the IPO prospectus back in March (balance sheet reclassification of redeemable NCI, restating 2023 and 2024, with no impact on income, revenue or cash flow). We look at the acquisition and path forward and are immediately reminded of acquisition indigestion seen many times before for similar businesses. Taking a view beyond this short-term hiccup yields material upside and an attractive risk:reward.

Overview

MAIR was the largest US industrial IPO in quite some time, and quickly ran to an intraday high of $44.50 creating a multiple problem. Then the tape turned against duration industrials. The Q2 print raised the sales guidance and the stock fell 9% anyway. On August 17th, the company announced a transformative acquisition, at an EV larger than its own public float, with underwritten debt and an equity component whose size and price are still undisclosed.

Looking forward, we should be most focused on whether the compounding algorithm is impaired. The CFO’s own decomposition on the Q1 call speaks to this… of 13% pro forma net sales growth, “volume was about 9 points, price was about 4. Commercial growth was mostly volume driven. That’s typically a backlog-driven business.” Then in Q2 they achieved 11 points of volume and 3 price. The company has outgrown US GDP in 16 of the last 18 years through 2025 (when measured on an if-owned basis).

Over the same four months in which the stock fell 42%, the business raised full-year sales guidance, grew Commercial orders more than 70% (up roughly 50% even excluding data centers), grew Resi through a falling housing market while expanding that segment’s margin by 423bps, and ended June with a record backlog near $2.9bn, more than half of it converting in 2027 and beyond, at 2.8x net leverage.

Today, pro forma for the deal, the stock trades at just under 14x 2027 EBITDA… for a business whose stated algorithm is mid-single-digit revenue growth converting to high-single-digit EBITDA growth before any synergy credit. From 27x to under 14x while guidance went up. We believe they are entering a unique window where EBITDA will compound above stated algorithm and assume no further M&A, nor a buyback, and very modest Resi / housing recovery out to 2030.

This is our first round of diligence, with more to come in the future as we go deeper on the product range and dynamics and refine estimates. For now, below the paywall we tell the high level story, the case for compounding… publishing in a timely manner here as the share price continues to approach levels we believe long-term investors will start to pick away to establish a floor. Math included, as always - matters more than the story (and don’t let people tell you that’s not the case).

Subscribe now

Read more

Kri-Kri makes Greek yogurt in one plant in Serres and sells most of it as supermarket own-label across Europe. Revenue grew 28% last year, and we estimate the plant ran near 92% of the milk intake its environmental permit allowed. Management deferred taking on new customers, and reported wasting product by over-running lines. Soon they’ll be able to show investors how much demand really exists in Europe for “authentic” Greek label yogurt (labels matter in the EU) just as GLP1 adoption also ramps.

Across fourteen reported years Kri-Kri compounded revenue at 14% a year. It did that from one site, through an energy war, and a milk inflation cycle. Demand has been solid, and now they need more room to run.

The market has picked up the permit alongside a sell-side note and run with it. We still see nice compounding upside on every time frame. The stock is breaking out and multiple modestly expanding but not demanding given the durability and drivers of growth. We are still looking for a call with management to understand the nuances and shape of the growth to come, but feel it’s reasonable to assume this capacity gets eaten up by end of the decade. The other important driver we’re trying to pin down and refine the range of outcomes for is milk input prices. These can swing and move GMs, but increasingly imported feedstock should also keep a lid on absolute costs (more on that below).

Secular demand growth

Kri-Kri is best analyzed as a manufacturing business, not a consumer brand. While their brand means something in some regions, the growth from here is increasingly exported and private label. Regardless, evidence points to the fact that they’ll be able to sell what they produce for the foreseeable future. There are three stacked demand curves… modest yogurt category growth, strong Greek yogurt share within the category (healthy and protein trends along side GLP1 growth, and now opportunity for them to take incremental share…

• UK Greek yogurt consumption growing 40%, importing some Kri-Kri already

• Authentic Greek, meaning strained yogurt made in Greece, is taking share inside yogurt.

• Europe sits at 2% GLP1 adoption against 12% in the US

• A Cornell basket study finds yogurt is the largest positive category shift among GLP1 users

• Kri-Kri taking share in authentic export private-label manufacture… we estimate it is the largest exporter, with 33.6% of a EUR 559m market that grew 42% in 2025.

So the growing demand of exported Greek yogurt has a very long runway supporting continued expansions. This is the type of long-term runway you want to look for to find multi-baggers. Nothing is guaranteed but it certainly leaves plenty of room to compound for years in a category where healthier eating habits can be driven by broad populations organically changing habits or via regulation or via GLP1 use incentives.

And for Kri-Kri, this is clearly their sweetspot and where they have been focusing the most effort.

Supply

The Serres plant operated under a permit capping milk processing at 299 tonnes /day, roughly 109k tonnes / year. Our intake reconstruction puts 2025 at about 100k tonnes. Management’s own commentary corroborates it. The ministry approved 554 tonnes / day, about 202k tonnes, split 480 dairy and 74 ice cream. The site expands 62%, water rights rise 39%, and a second whey digester doubles the straining bottleneck’s relief. The state co-funds the expansion. About EUR 23m of tax relief becomes available after the project completes at end-2027, so it phases from 2028, and a separate certified relief of roughly EUR 6.5m is already in hand with about EUR 4m usable in 2026.

How fast can the new plant ramp, and what is it worth?

Our base case fills the new ceiling to 87% by 2030, and net income roughly doubles against 2025 by 2029. The build is mechanical with milk intake times revenue / tonne, then margins from the milk work below. Revenue per tonne drifts from EUR 3.3k toward 3.5k on mix. That gives revenue of EUR 396m this year, EUR 497m in 2028 and EUR 604m in 2030. Management frames the current expansion as capacity for about EUR 500m of sales, and was careful to add that it depends on demand. Note that management states the Dynamo doubling against 2024, not 2025. Our plant-defined mid-case at full utilization is EUR 600m. The three anchors agree.

The cadence is the model’s biggest unknown. A doubling that arrives in one step in late 2027 and one that phases across six quarters are different stocks for a year. The company has never disclosed the sequence. Until it does, our 2027 carries the widest error bars in the model.

One broker models EUR 715.8m of revenue in 2029. At the company’s 2025 dairy revenue per tonne of milk, that needs about 263k tonnes of intake, or 130% of the approved ceiling. Their number is only reachable through a large mix shift toward fortified recipes, an unannounced further permit, or both. Perhaps there is further upside in scale or mix, but we would not underwrite this without confirmation from management.

Milk prices and margins

The cost base is roughly 90% Greek liquid milk. Asked directly what share of cost of goods is Greek versus EU milk, management said imports of raw milk are very limited, that they mostly use Greek milk, and that what they import is dairy commodities such as butter and proteins. So the cheaper leg is commodities, not liquid milk, and the premium full-fat cups and the domestic book carry Greek milk throughout. Greek farmgate stands near EUR 56.60 per 100 kilos against an EU average of about 41.84. The spread is above a third, the widest on record, and it helps through the commodity line rather than the milk line. The gross margin expansion of 5.5 points in Q1 is better explained by the price adjustments management took from H2 2025 and by falling dairy commodity prices, which management cited directly when noting the energy surcharge was largely offset.

Greece produces roughly 40% of the cow milk it consumes. Feed is imported and runs 60-75% of production cost. Farm numbers fell 42% in a decade, and 358 farms now deliver nearly three quarters of the pool. A price ~30% above the EU average has produced supply growth under 2%. The 2026 EU price decline is driven by a shrinking herd while the heifer pipeline is the thinnest in decades. EU prices firm into 2027 and 2028, exactly when the new capacity lands. Meanwhile, a growing competitor plans to triple production with a EUR 150 to 160m plant into the same market. Thus, we have tried to prudently give back some margins after a great 2026 for this spread.

However, this remains a key risk to monitor… a double squeeze, EU inputs recovering fast while Greece stays high, produced 2025’s 27.0%. Our bear carries exactly that.

Free cash flow coming

Today’s modest net cash position supports the buildout. Prior years saw lumpy working capital and CapEx leading to minimal FCF. Normal cadence with more scale and similar CapEx as those years will yield a growing net cash balance to nicely support the dividend (current 1.5% yield, or EUR 0.45).

Cumulative free cash flow after dividends from 2026 to 2030 is about EUR 111m. Phase 4, the post-2030 study, is EUR 100 to 150m. The cash the plant generates and the next project management is contemplating are the same size.

Valuation

We value the business on EPS as the going FCF will likely be given back in a few years to support further expansions. Importantly, past EPS declines were largely driven by GM compression resulting in real pullbacks in the stock. However, the underlying demand and volume growth support a nicely growth EPS algorithm when margins are normalized and observed over any multi-year time frame. We use a range of multiples of just 14-18x which result in a very modest PEG ratio. There is a scenario where finding growth stocks, with durable trends, in the EU becomes hard in the coming years as a result of so many secular headwinds… this could see Kri-Kri trade at a multiple of 20x EPS or higher.

The multiple range is not arbitrary. It comes off what buyers actually pay.

The transaction comps below should be viewed as the high end of what is possible.

Risks

UK case law defines Greek yogurt as made in Greece, traditionally strained, without additives. The fastest growing tier is protein-focused recipes lawfully declaring EU dairy inputs. A purist litigant or a front-page investigation into what authentic means could compress the whole category’s premium. Two mitigants are real... the practice is industry-wide, since Greece’s entire 400k tonne yogurt output could not exist on traditional straining from a 640k tonne pool.

Second, the input squeeze. BellRing fell 80%+ when protein inflation and club own-label broke its margins. Kri-Kri buys the same protein complex for its growth tier. However, they are unlevered, have a strong brand position, and pass-through on most UK voluemes.

Third, a competing plant is tripling capacity with state funding and is backed by the country’s most successful food-export builder. It arrives after our ramp fills, which is why it caps the terminal multiple and growth rather than the base case in order to strike a conservative approach. However, they will also begin bidding for the same fixed milk pool, so Greek farmgate drifts up in our base case rather than flat.

Fourth, capital allocation as the cash accrued through 2030 is likely to be redeployed again for expansion, it seems.

Conclusion

Kri-Kri is the manufacturer behind Europe’s fastest-growing food category, and the binding variable was never demand. Expansion is coming which will support multiple years of growth at the same time GLP1 adoption grows. This type of durable demand profile is rare to find in packaged goods, and we believe the market will pay a healthy multiple for it. It has secular, thematic growth, is self-funded, and entering new markets that are rapidly absorbing product. The risk remains input costs which will certainly swing over time, but this should be manageable. The large ~70% family ownership makes an outright acquisition unlikely, but acquisitions of this sort of growth and margin profile have come at multiples well above our underwriting cases. Current valuation is fair and we aim to get a call with management to drill down on the most important drivers.

DISCLOSURE

Position: At the time of publication, the author holds a position in KRI.AT.

Trading Policy: The author will not materially alter this position within 48 hours of publication. After this period, the author may buy, sell, or otherwise adjust the position without further notice. Changes to the author’s view or position will be reflected in subsequent publications when material.

Conflicts: The author has received no compensation from the issuer or any party with a financial interest in this security.

Forward-Looking Statements: This report contains the author’s opinions, estimates, and projections, including price targets derived from financial models. These are forward-looking statements subject to substantial uncertainty. If these assumptions prove incorrect, the actual value may differ materially, including scenarios of significant loss or total impairment. The price target represents the author’s estimate of fair value under the stated assumptions, not a prediction of where the stock will trade.

This report is provided for informational purposes only and does not constitute investment advice. See the full Terms & Disclosures for additional important information.

ADI reported its first quarter as a public company last week. One number changed our targets and most of the rest confirmed the setup. The new CFO put the real GM run rate at 21.35%, about 100 points below the base the July targets were built on. Reported margins had been flattered by tariff pricing and a $20M one-time refund. The targets move from $8, $30 and $58 to $7, $26 and $39, struck the same way as before (1yr downside, 1yr upside, 3yr upside). The bull case, a residential recovery, yields adds $33 and $49 at the same multiples.

The tape has been wild. I do not have good trading color to justify these moves. The stock was added to the Russell 2000 and S&P 600, but REZI was already a part of these. Mechanically, this should mean the index received the spinoff shares so was not a large buyer in the market. We also do not have any insider filings yet. The one-year target now offers 23% and the three-year offers 81%, about 19% a year. Importantly, a long term investor can look out three years to a resi recovery and see a 25% IRR from here. That’s worth doing the work on.

The size of the change deserves a clear statement. Our July numbers had 2028 EBITDA at $429M, but the clean, updated figure is $370M in our base case. This change was driven almost entirely by GMs, falling 100bps (100bps on $5.4B of revenue is $55M of the $59M cut). The demand call was right and the margin base was wrong.

The same arithmetic hits management’s plan. Their bridge ran 22.3% to above 22.7%, a task of 40 to 50bps. From 21.35% the identical target needs 135. At roughly 35bps per point of EB mix, that is four points, 18% to 22%, which is exactly their stated plan. Their target is internally consistent. Their $500M for 2030 also compounds off a lower base now. From this year’s guide it requires 15.1% a year vs. our clean path compounds 12.1%.

https://open.substack.com/pub/banyanlanecapital/p/adi-global-distribution-adig-greenblatt?r=5cm8h&utm_campaign=post-expanded-share&utm_medium=post%20viewer

Results

Four things moved our numbers. First, GM ran 21.3% to 21.4% in both quarters this year, and management called that the run rate of the business. The July targets started from 22.3%. Second, the mix engine stalled, it seems. EB revenue fell almost 3%. Management disclosed that roughly three quarters of the $800M program sits in residential AV, the one market still declining. A clear negative vs. expectations, but also provides cyclical kicker should the housing market thaw. Third, the guide itself of $275M to $295M of standalone EBITDA, below last year’s $295M, and about $265M excluding the refund. In addition, there were 4 analysts on the call asking questions, but all shorter term - good questions and trying to figure out the near term noise in numbers which is not uncommon in spins.

The call also laid bare the cost of the savings program, we’ll call this the The Savings Treadmill… $80M gross cost program nets against merit, inflation and growth spending to roughly zero. EBITDA growth therefore has to come from gross profit. Management said both part aloud. Sixty million of savings are already actioned, and operating expenses next year should not grow “significantly, if at all.” Flat opex is the win. The program costs money though.

The call. Our base case is the clean path below, because the savings are contracted and commercial demand is intact. Security is back to mid-single-digit growth, and management says most of the share lost in the system conversion has returned. The next print confirms it if net operating expenses decline as guided and the newly filed exclusive-brands revenue split holds steady. A residential turn moves us to the bull case. EB mix and GM flat together through mid-2027 ends the thesis.

Subscribe now

Read more

The stock has been more challenged than the chart above since covid, despite the underlying moat and strength of the business being unchanged.

The disconnect is due to both pull forward of pet adoption rates during covid, and the fact that Americans have been taking pets to the vet slightly less often for three years. The fear has been this is a broken consumer story. Meanwhile the growth engine re-accelerated from 4.5% to above 10%. Visits now increasingly suggest testing, thus creating incremental growth not tied to visits.

The AI angle is not a chatbot, but a cytology analyzer that replaced the glass slide. It is a $15 cancer screen that finds lymphoma months before symptoms. It is software that writes the medical record while the vet talks. Nine thousand of those analyzers are in clinics eighteen months after launch. Ten thousand clinics order the cancer panel. Next week, at the 13 August investor day, a fifth undisclosed platform may get unveiled.

I did the full workup in late July with the stock at $559, hoping to get a shot 5-10% cheaper. It never quite got there, then earnings hit, and shares rose nicely, again. So the honest title of this note is the one at the top. The turn is visible in the numbers, and I partially missed it. Below the line is the complete work… the business, history, why the acceleration is structural, and what the AI stack is actually worth. Also included are my forecasts and scenarios, and a preview of what I expect at the investor day.

Subscribe now

Read more

Elevator pitch

• Probe cards are consumables… they wear out and are rebuilt for every chip design, so revenue tracks wafer starts and design turnover, not capex sentiment

• Every AI path pays the toll so they win with both GPUs or custom ASIC, HBM or DDR, pluggable or co-packaged optics… all of it runs wafers, and every wafer gets tested

• Q2 proved the model with revenue up 32% and GMs 1,500bps in four quarters and EPS tripling

• A second engine is now lighting… co-packaged optics test revenue passes $20M this year… management sizes this market at $400M by 2030

• The stock is up 4x from its April 2025 low on four straight record quarters, but still maintains powerful secular tailwinds that present good multi-year upside

Why this opportunity exists

• The market files FORM under semicap equipment, so it trades on capex sentiment… revenue actually follows unit volume, which capex fear does not dent

• Complexity is good… test intensity per wafer compounds with stacking, pin count, and now optics

• The July selloff was a memory pricing scare and momentum washout… falling memory prices come from rising memory volume, and volume is FORM’s paymaster

The business in 60 seconds

FORM makes probe cards or boards holding thousands of microscopic springs that touch a chip while it is still on the wafer and check it electrically. A defective die caught here costs cents to discard. The same die caught after packaging destroys a $10,000-plus AI package. Cards wear out and are redesigned for every new chip, which makes this a razor-blade business. FORM holds the number one share in advanced probe cards, roughly twice the next competitor [VLSIresearch]. It also sells the probing systems used to characterize chips, which is where the optics story lives.

Free section ends here.

Read more

The pitch

• On August 4, every Resideo (REZI) holder receives shares in North America’s largest security-products distributor.

• ADIG will debut trading with a $261M loss last year and 3.0x leverage, rated BB minus. Index funds and many active managers will punt the stock - a classic Greenblatt-style spinoff.

• Losses and the leverage were both inherited from REZI - they are the result of REZI’s 2018 Honeywell indemnity that it received upon spin. Operating profit grew every year underneath both.

• Management plan calls for EBITDA compounding from 2025 depressed $295M toward $500M as a result of finally getting hold of the strategy with initiatives that have largely been in the works for a year already.

• Even the bear grows EBITDA every year and sweeps debt below $100M by 2030, so the business survives but the multiple wobbles.

Why this setup exists

• Forced sellers: REZI holders own a products story. Index funds must sell what no index includes. Nobody will cover the stock for some time, potentially quite a while.

• Optics: $364M charge, junk ratings, and three different EBITDA figures likely to give many pause.

• Neglect: The business spent 30 years as a segment inside Honeywell, then Resideo, and it has never been managed for its own benefit.

The business in 60 seconds

ADIG is the middleman between roughly 1,000 electronics manufacturers and 100,000 security and AV installers. When a two-truck alarm contractor needs a camera, a panel, and a spool of cable today, on credit, they go to ADI. He also needs someone at the counter who knows which transformer fits. ADI runs 200 counters, 20 distribution centers, and 500,000 products to answer that need. Revenue is $4.8B. Gross margin is 22.3%. EBITDA margin is 6.2% standalone, against 10% to 13% for the best distributors. That gap is the whole investment. Security and fire are 65% of revenue, AV is 30%, data-comm is 5%.

Three straight years of growth through an ERP conversion and a housing trough, then three paths forward. Management’s own 2030 floor sits just below our base case.

The real pitch is the distressed print

To be clear about what we are actually interested in here, when mechanical selling starts on August 4 we believe the trough could bring a very attractive risk:reward. And few will have done the deep work to understand the opportunity they aim to execute against over the years to come. CD&R owns roughly 20% as-converted. Open-market common purchases from management would validate the equity.

Read more

The typical lockup is 180 days after IPO allowing private shareholders, executives the opportunity to monetize. Everyone knows the date, and it sits in the prospectus months ahead. The lead up to these events serve as a well-known opportunity for long and short investors to position ahead of the feared supply that will swamp demand on the days that follow lockup expiry.

So, I went to check whether that actually happens. My last piece, Nine in Ten IPOs Break, found that most of an IPO’s damage lands in the first few months. That is well before the lockup opens. This is the follow-up, and it puts the lockup itself under a microscope.

I pulled 19 of the most recognizable IPOs of the past fifteen years, and ran a study of the period around each one’s 180-day unlock. Every figure here is market-adjusted, meaning the stock’s return minus the S&P 500 over the same days. A falling market

never gets counted as a lockup effect. Day 0 is the first trading day at or after 180 calendar days from listing. Direct listings sit out, since going public without underwriters means there is no lockup at all. Coinbase, Palantir, and Roblox are therefore excluded.

Here is the short version. The dip everyone waits for has already happened by the time the lockup arrives. The unlock day itself is a volume event with a small price notch. The two largest insider floods in the sample did not crater their stocks. They rallied them.

Subscribe now

The fade shows up before the unlock

The dip is mostly over by the time insiders can sell. Stack all 19 names on the same clock and the shape is plain. The average IPO walks into its unlock already down about 15% against the market over the prior 30 trading days. The median is worse, at down 21%. By the time insiders can legally sell, the selling is nearly done.

Some of that slide is ordinary post-IPO drift. These are expensive stocks. The quiet period ends, when the company and its bankers stop promoting the stock, and the first earnings misses land. Some of the slide is the market front-running a date it can see coming. I will not claim the lockup causes the whole 15%. But the direction is clear, and it runs opposite to the folk model. The weakness comes before the event.

Volume is the only thing you can count on

The one dependable feature of an unlock is a jump in volume. On the day itself, volume explodes. The mean is 4.5× normal, and the median is 3×. Every marquee name in the sample lit up. Twitter traded 11×. Uber traded 15×. The float, meaning the shares actually available to trade, multiplies overnight, and the tape shows it at once.

The price move is a different animal. The unlock day is usually a small down-notch. 15 of the 19 names fell against the market. But the average was only −2.7%, and it does not stick. Volume tells you the event happened. It tells you almost nothing about direction.

The reaction is an average, not a rule

The base rate dissolves the moment you look at the names. The 20-day reaction runs from Beyond Meat at −28% to Facebook at +33%. Just over half the sample was negative. The rest rose, some of them sharply. This is not something you can trade blind on any single name. It is a faint tilt buried inside huge dispersion.

The worst-behaved cohort is the one you would guess. The 2019 profitless-unicorn class averaged −5% on the day, on roughly 7× volume. Uber, Lyft, Pinterest, and Peloton led it. The three names that barely flinched arrived at their IPO already profitable, namely Alibaba, Zoom, and Facebook. That is the same fault line my last piece found across the whole asset class.

The biggest single unlocks rallied

The largest single unlocks in the data went up, not down. The cleanest test sits inside two stocks that released their shares in stages.

Facebook spread its release across five dates. The reactions ran the full range. Down 7% on the 90-day tranche. Almost nothing on the 150-day. Then up 14% on the 180-day unlock, which freed 777 million shares. That was the single largest block of the whole schedule. The biggest flood of insider stock in the sample produced a rally.

The +19% pop a few days earlier was Facebook’s Q3 earnings, not a lockup. The two get conflated constantly, and they are not the same thing.

Snap tells it in miniature. Its first unlock freed 400 million shares to early investors. The stock had already bled 22% into the date, and it slipped about 1%. Then came the big one. 782 million employee shares, freed four days after a poor earnings report. That is the exact setup the fear model says should be catastrophic. The stock rose 5.5%.

The mechanism is not mysterious. By the time the flood everyone circled finally arrives, it is the most anticipated seller in the market. The overhang that pressed on the stock for weeks clears. The last buyer of downside protection is done. It is sell the rumor and buy the news, expressed in share count.

The size of the flood is irrelevant

The size of an unlock does not predict the reaction. If the fear model were right, the deals that free the most stock should fall hardest, and they do not. Initial float across the 19 names clustered in a thin band. The median was around 10% of shares outstanding, with none above 20%. So the lockup takes each stock from a sliver of tradeable float to nearly all of it. That is a float expansion of roughly 6× to 16× overnight.

The expansion has no reliable link to the reaction. The rank correlation is near zero over the 20 days after the unlock. On the day itself the faint tilt runs backwards, with bigger expansions posting slightly better days. The two largest single-session releases settle it without a regression. Facebook freed 36% of its shares in one day, and Snap freed 67%. Both rose. What sets the reaction is not how much stock unlocks. It is whether the market saw it coming, and with the date in the prospectus for months, it always does.

Robinhood shows where the cliff went

The clean 180-day cliff is disappearing, and Robinhood is the tell. Its lockup was a patchwork of early releases. A slug of shares became sellable under Rule 144 from the 91st day. Rule 144 is the SEC rule that lets holders sell restricted stock after a set holding period. Price-based triggers freed more, and the nominal 180-day date sat underneath it all.

By the time that date arrived in late January 2022, it was a non-event. Volume ran 1.1× normal. The float had already been let out in pieces, and the stock had already fallen by roughly two-thirds on its own merits. Robinhood’s worst day in the window was not a lockup at all. It was the −10% reaction to Q3 earnings.

More deals now write in these staggered, performance-based releases. The single date you could once circle is turning into a smear. If you still trade the lockup, use the date where the shares actually move. The volume spike will tell you, and often it is nowhere near the 180-day mark.

The folk trade runs backwards

The intuitive lockup trade is backwards. The obvious version is to wait for the lockup and short the flood, and it points the wrong way. The fade is a pre-event, not a post-event. By the day the shares unlock, the crowd has already shorted it, and the overhang is already in the price.

The evidence says the stock is closer to a local bottom than to a fresh leg down. Across the sample, the average name is at its worst the day after the unlock. Then it grinds back to roughly flat against the market within 60 trading days.

If there is an edge near the event, it is the mirror of the folk trade. It is the weakness into the print, not out of it. But the same wall from my last piece applies here, and harder. Borrow on a fresh IPO is scarce and expensive. It gets cheap only once the lockup floods the float, and by then the move is over.

Options are the cleaner vehicle, but implied volatility around these dates is brutal. The pattern is real, but the tradeable version is small. It lives in careful timing and instrument selection. That gap is probably why it survives.

Caveats

This is 19 hand-picked marquee names, not the full universe. They are the famous, heavily analyzed deals with the loudest lockup coverage. That is deliberate. It also bakes in survivorship and selection, and the small-cap tail behaves worse. Returns are market-adjusted with a beta of one, not a fitted factor model.

The 180-day mark is a convention. I use each stock’s first trading day at or after IPO plus 180 calendar days. For the two names with staggered or early-release structures, Snap and Robinhood, that date misses the real unlock. That is why their day-0 volume is flat. The “down 15% into the event” figure blends real pre-positioning with ordinary post-IPO drift, so it should not be read as pure lockup anticipation. LinkedIn belonged in the sample but was dropped, since clean split-adjusted history for it no longer exists after its acquisition.

What I take from this

The lockup is a supply event the market prices in advance. The damage lands in the weeks before, not the day of. If you wait for the unlock to tell you something, you are quite possibly reading yesterday’s news.

Volume is the signal, and price is noise. Every unlock spikes volume. The direction of the day is a small, mostly-down flicker that fades within weeks.

The size of the flood is not the risk. The two largest insider unlocks in the sample both rallied. Across all 19 names, the float-expansion multiple has no bearing on the move. A fully anticipated seller clearing the overhang is a bottoming mechanism, not a crash.

Check the date, not the calendar. Early-release and staggered structures have moved the real unlock off the 180-day mark. Follow the volume, not the six-month rule of thumb.

The lockup is not where IPOs break. It is where the market finishes pricing a break it saw coming months earlier. And often enough, it is where the selling finally ends.

19 marquee US IPOs from 2012 to 2021, with daily prices from Financial Modeling Prep. Event study around each name’s 180-day lockup expiration, defined as the first trading day at or after IPO plus 180 calendar days. Returns are market-adjusted against the S&P 500. Staggered-tranche dates for Facebook and Snap, and the Robinhood early-release timeline, come from contemporaneous reporting and SEC filings. Per-tranche volume uses a trailing 20-day median. Unlock magnitude compares each deal’s base offering size, excluding the greenshoe, to shares outstanding from the quarterly filing nearest each unlock. The greenshoe is the underwriters’ option to sell up to about 15% more shares. Direct listings are excluded. None of this is investment advice. It is a base-rate study. See here for Disclosures & Disclaimers.

So I had Fable build it. Roughly 3,000 US operating company IPOs from 2010 through mid 2026, assembled from Ritter’s IPO database, IPOScoop’s offer price records, and SEC prospectus filings. SPACs, closed end funds, and ETF listings are stripped out. After requiring a valid offer price of at least $4 and a clean daily price history from listing day forward, the core sample is about 1,700 IPOs. Of those, 1,638 also have financial statements from around the time of listing.

The typical IPO loses money

Index every IPO to its offer price, so offer equals 100, and stack all the price paths on top of each other. The median IPO closes day one around 106. It drifts up for a few weeks, tops out, and then bleeds. By the end of year one the median IPO trades below its offer price. By the end of year two it sits around 87 and is still falling.

The mean tells a completely different story. It stays above par and grinds higher the whole time. That gap between mean and median is the entire IPO asset class in one chart. The average gets rescued by a small right tail of enormous winners while the typical name in the group is a losing investment from the first print. Buy every IPO at the day one close and your results depend almost entirely on whether you caught the handful of Airbnbs (priced at $68, peaked at $220, then chopped between $80 and $160 for ~5 years).

A few base rates from the sample are worth having in your head. The median day one pop is 6.5%. The mean is 17%, dragged up by the tail. And about 29% of IPOs close their first day below the offer price. The pop is not a law of nature. It is a coin flip with a mild upward tilt.

The early peak comes fast. Looking at the first 60 trading days, the median IPO tops out around trading day 24, roughly a month after listing. About one in five IPOs peaks within the first three days, meaning the pop was the top.

Then comes the purge. Measured from that early peak, the median maximum drawdown over the following year is 50%. The 25th percentile is 70%. This is not a mega cap thing or a busted small cap thing. It is the base rate for the asset class, and it sits right in the range Truist found on their 30 famous names.

More than 90% trade below their offer price at some point, and the median first breach happens around trading day 31, about six weeks in, after the quiet period ends and well before the lockup opens.

Who comes back

Since breaching the offer price is nearly universal, the interesting question is how long stocks stay under. Most breaches are brief. A stock hovering near its offer dips below and pops back within days. But there is a thick, stubborn tail. I ran the recovery question through survival analysis, treating delistings both as censored observations and as permanent failures, and the two versions barely differ. About 15% of breachers are still below their offer price a full year after first going under. Roughly 10% are still under three years later. Once a name settles into that tail, delisting becomes a common exit. About 40% of the full sample no longer trades today.

Only about six in ten IPOs ever get back to their early peak, and the ones that do often take the better part of a year.

One accounting flag splits the whole market

At the time of listing, only 38% of these companies had positive net income and only 51% had positive EBITDA in their latest fiscal year before the offering. Split every path metric by that one flag and you get two different asset classes.

The EBITDA positive group pulls back a median of 38% from its early peak. Its median price path hovers around par, 103 at one year and 99 at two. Two thirds of these names eventually reclaim their early peak. These stocks mean revert.

The EBITDA negative group pulls back a median of 60%. The median path sits at 88 after one year and 68 after two, still declining. Nearly half are delisted today.

And the part that should bother you. The unprofitable names pop harder on day one, a median of 8.7% versus 5.5% for the profitable ones. The market pays its biggest opening premium to exactly the cohort with the worst subsequent outcomes. Story stocks pop on story. Then the quiet period ends, the lockup expires, the cash burn continues, and gravity does the rest. As far as I can tell, essentially the entire famous long run IPO underperformance result lives inside the unprofitable half of the market.

Two refinements on this.

First, pop size interacts with profitability until it stops mattering. Profitability protects you in every pop bucket except the biggest one. A modestly popping profitable IPO, up 0 to 10% on day one with positive EBITDA, behaves like a normal stock, with a median one year forward return from the day five close of about plus 3%. But IPOs that popped more than 75% were toxic no matter what the financials said. Median one year forward returns were minus 37% for the unprofitable ones and minus 47% for the profitable ones, though that last cell is small. A giant pop is not a quality signal. It is a supply and demand dislocation that gets corrected over the following year, usually violently.

Second, growth partially redeems cash burn. Among the EBITDA negative names, the top quartile of revenue growers, with growth above roughly 80% and a median near 170, showed much better forward returns, roughly minus 12% over the next year versus minus 30 for the slower growers, along with lower delisting rates. The true kill zone is the slow growing money loser. Burning cash while growing less than about 20% carried a median pullback of 64% and about half of those names are gone. Burning cash while hypergrowing is a gamble. Burning cash while growing slowly is a diagnosis. VC backing, for what it is worth, mostly proxies this same split, since 79% of VC backed IPOs came public without positive EBITDA.

The uncomfortable trade

Put the pieces together and a systematic pattern falls out that is almost embarrassing in its simplicity. Short the unprofitable IPOs after the early peak. Entering short at the day 20 close and covering six months later produced a median gain to the short of about 16% with a 63% hit rate across 800 EBITDA negative IPOs. Waiting until day 40 or 60 does slightly better still. Run the same trade on the profitable cohort and you get about 2%, which is no edge at all. The whole anomaly is concentrated in the cash burners.

Before anyone runs off to build this, those numbers are gross of borrow costs, and borrow is exactly where the market defends itself. There is no rule stopping you from shorting a new listing. You can short the moment your broker locates shares. The problem is that shares barely exist to borrow at first. IPO stock has not settled, lendable float is tiny, and fees on fresh deals routinely run to triple digits annualized. In practice, borrow becomes possible within the first week on larger deals, tolerable over the first month, and cheap only after the lockup expiration floods the float about six months in, by which point much of the fade has already happened. Options usually start listing about a week after the IPO and are often the cleaner early vehicle, though implied vol on fresh IPOs is brutal and eats a good share of the expected fade. The edge in the data is real. The implementable edge is smaller and lives in careful instrument selection. That gap, where everyone can see the pattern but few can cheaply trade it, is probably why the pattern survives.

Caveats on base rates

Prices are split adjusted, and I cross checked day one closes against contemporaneous records to catch adjustment mismatches. About a fifth of names needed correction, mostly reverse split casualties. Names with unrecoverable price history, which skew heavily toward fast failures, are missing entirely, so if anything these numbers flatter the asset class. Fundamentals use the fiscal year closest to listing, which for some names is the first post IPO year. Companies that changed tickers along the way, which usually means they were acquired or rebranded, drop out of the sample, and since acquisitions skew toward good outcomes this cuts mildly against the recovery numbers. Offer price records also thin out after 2020, so a handful of famous recent listings like Reddit, Arm, and Cava sit outside the offer price sample even though their price paths check out fine. Direct listings were never in scope since they have no offer price at all. Recent IPOs have not had time to complete their arc. And the 2020 and 2021 vintage, where 60% of IPOs were unprofitable and median pullbacks hit 66%, weighs heavily on the worst cells. The pattern shows up in every era but the magnitudes move with the cycle.

What I take from this

If you like an IPO, plan around a better than 90% chance it trades below its offer price at some point, most likely one to three months after listing. The best time to buy something you believe in is rarely week one. The supply calendar and the data both argue for patience measured in months, not days.

One accounting flag, positive EBITDA at listing, separates stocks that mean revert from stocks that melt. It is the cheapest quality screen in the market and almost nobody applies it to new issues.

And if a deal doubles on day one, that is a reason for caution rather than celebration. The biggest pops had the worst year one outcomes even among profitable companies.

The IPO market is not mysterious. It sells stories at peak enthusiasm into constrained supply, then delivers the shares, and the truth, on a six to twelve month delay.

Data covers roughly 3,000 US operating company IPOs from 2010 to 2026, built from Ritter’s IPO database, IPOScoop, and SEC filings, with daily prices from FMP and Yahoo. The core sample with financials is 1,638 names. None of this is investment advice. It is a base rate study, and base rates are where the bodies are buried.