Musical Chairs

---

THE NUMBER: $1.25 billion — Anthropic’s rent line to xAI for compute. Per month. The detail emerged today inside SpaceX’s S-1 filing with the SEC, ahead of the company’s June 12 IPO. Anthropic will pay $1.25B every month through May 2029 for the entire 300-megawatt output of Colossus 1 in Memphis — roughly $40 billion in total commitments. Either side can terminate on 90 days’ notice. xAI overbuilt; Grok usage has dropped; Anthropic absorbs the slack. That’s the rent line for one (1) frontier lab. Not the salary line. Not the talent line. Not the chip purchase. Rent. It is, by some margin, the largest single recurring infrastructure payment in the history of software — and it is the moment the AI capex era stopped being abstract and became a line on a tenant’s balance sheet.

---

The Eight-Seat Club

Take a minute to count the seats. There are eight. Apple, which founded the club in 2018 and now sits at $3.5T. Microsoft, elevated 2019. Nvidia, the gold-rush picks-and-shovels play, at $5.5T this month after today’s blowout quarter — Jensen Huang told the call demand has gone “parabolic.” Alphabet. Amazon. Meta. Saudi Aramco, which crashed the gate at IPO in 2019 as a sovereign-controlled oil major with 80 years of operating history. Berkshire Hathaway, built over six decades by a man who never IPO’d anything new in his life and just made the existing thing more valuable every year for sixty-eight straight.

Every single member of the club is battle-tested. Every single one generates extraordinary free cash flow against its capital base. Not one of them is in a capex emergency relative to its operating earnings. The club is exclusive by construction. The public market doesn’t add seats because a company shows up with a story. It adds seats when a company shows up with a decade of receipts.

Three lobsters are circling the floor this month.

OpenAI, reportedly filing a confidential S-1 this week, targeting a $1 trillion valuation. Eleven years old. The largest user base in AI by an order of magnitude. Free for most of those users. The biggest enterprise account on the books is itself.

Anthropic, just raised a primary round at $900 billion, trading as high as $1 trillion in private secondaries. Four years old. Fastest-growing enterprise share in the industry — 34.4% of paid business adoption on Ramp’s May Index, versus OpenAI’s 32.3%, in a category that didn’t meaningfully exist eighteen months ago. Compute footprint that lives mostly on someone else’s metal.

SpaceX, S-1 filed today, IPO targeted June 12, widely expected at or above $1 trillion. Twenty-four years old. Real launch business at $18.67 billion in 2025 revenue. A $4.94 billion loss last year, almost entirely from AI spend. TAM claim in the document: $28.5 trillion. “The largest TAM in human history.”

Two chairs. Three lobsters. The music is playing.

---

The Rent Line

The $1.25 billion is the part nobody outside this newsletter will put in the right frame, so we’ll put it here. Krishna Rao, Anthropic’s CFO, gave the math on Invest Like the Best last week: “If you buy too much compute, you go out of business. If you buy too little, you can’t serve your customers. Same thing.” Rao spends 30-40% of his working life on compute decisions. He calls compute “the canvas on which everything else gets built.”

Anthropic chose to rent the canvas. That’s the decision the rent line captures. Buying 300 megawatts of frontier compute is a multi-year construction program at hyperscaler prices. Renting 300 megawatts from a competitor who overbuilt is a checkbook move. The math at today’s prices favors the checkbook. The math also has a 90-day reset button.

Both sides of the Colossus 1 contract can be terminated on 90 days’ written notice. For xAI, that’s a feature — Grok usage may rebound, Cursor’s compute demand inside the SpaceX family is scaling, capacity may eventually be needed back home. For Anthropic, the 90-day clause is the wrinkle on the otherwise-cleanest bull case in AI in 2026. If Musk decided to play hardball — and there is no reason today to think he will, because the deal is monetization for xAI, breathing room for Anthropic, and a revenue line in the SpaceX S-1 that helps justify the orbital story — Anthropic would have 90 days to find 300 megawatts of replacement frontier capacity on a planet that, per Michael Dell at Dell Technologies World today, is in a supply chain crisis projected to last through 2028 and possibly 2030. The replacement capacity does not exist. Anthropic knows this. So does Musk.

He probably won’t pull it. The current configuration works for all three companies at once. What the 90-day clause does guarantee is that Anthropic’s compute position can be re-priced — not yanked, re-priced — on a clock Musk controls. That is not a moat. That is a vendor relationship. Long Anthropic is, structurally, also long the Musk infrastructure complex, whether the buyer means it to be or not. The piece of Anthropic’s $900 billion valuation that depends on continuity of compute supply is the piece of the trade most investors will not notice they are buying until the day they suddenly notice it. Note it now.

---

Long SpaceX, Short OpenAI

The pair trade the smartest allocators in this market are already running, even if they’re not naming it out loud: long SpaceX (or long Anthropic), short OpenAI. The mechanism almost nobody is connecting on the same page.

Run the chain.

SpaceX’s S-1 makes one claim that, if it clears, eats half the AI industry’s premium pricing. The document targets $185 per kilogram to orbit on Starship. Today’s Falcon 9 internal cost is approximately $1,500/kg. Starship is supposed to do one-eighth that. The S-1 plans 100 gigawatts of compute deployed to orbit per year starting 2028. None of this works at current Starship reliability. The vehicle has completed nine test flights. Falcon 9 has executed 450+ successful landings. The orbital compute thesis is not pricing in launch risk because there is not yet a reliability surface to price.

But if Starship works — and SpaceX is betting $4.9 billion of last year’s earnings against the proposition that it will — and the cost curve bends as advertised, then a different company gets the secondary benefit before SpaceX-the-stock does. Cursor, whose parent Anysphere signed a $10 billion call option with SpaceX in April for subsidized compute, becomes the lowest-cost coding agent on the planet by some margin. Cursor is already winning the prosumer dev market that Claude Code, Lovable, and Replit are also fighting for. Cheaper compute lets Cursor undercut all of them.

The collateral damage is on OpenAI’s income statement. OpenAI’s most credible enterprise revenue line in 2026 is coding — Codex, the API revenue tied to programming workloads, the ChatGPT Plus power users who are mostly developers. If Cursor on subsidized SpaceX compute commoditizes the coding-agent layer, the line on OpenAI’s S-1 that justifies a trillion-dollar valuation gets harder to draw. Long SpaceX is structurally short OpenAI through Cursor. The harder SpaceX wins, the worse OpenAI’s IPO multiple gets.

That is the trade in eight words: if Starship clears, OpenAI’s pricing power doesn’t. Wall Street rarely names trades this clearly because Wall Street is not yet underwriting them. By July, it will be.

---

Anthropic Wins on Every Front (Except One)

Set the rent wrinkle aside for a section and the honest read of Anthropic‘s position is this: they are winning on every dimension that matters.

Models. Claude Opus 4.7 holds the top of Harvey’s BigLaw Bench, the top of Vals AI’s finance agent benchmark, and the top of the only enterprise benchmark that actually pays — the Ramp May AI Index at 34.4% versus OpenAI at 32.3% and Google in low single digits. Anthropic’s enterprise share grew 4x in twelve months. OpenAI’s grew 0.3 percentage points.

Talent. Karpathy joined Tuesday. The Mercury Seven is complete.

Product velocity. Three commits today. KPMG embeds Claude in Digital Gateway across 276,000 employees and names KPMG Anthropic’s preferred consultant for private equity. Bristol Myers Squibb deploys Claude across drug discovery, clinical, regulatory, manufacturing, and commercial workflows. Revenue is reportedly set to more than double to $10.9 billion in Q2. The drumbeat is daily now.

Moral posture. Dario Amodei’s 2017 thesis paper is one of the few founding documents in this industry that has held up against the company’s own subsequent behavior. While Greg Brockman wrote checks to MAGA Inc., while Altman litigated his own charter in federal court, while Musk built rocket-powered server farms, Amodei stayed on the page. That matters less to a quarterly trader and more to the enterprise buyer making a seven-year commit at the scale KPMG just made.

The wrinkle is the one we already noted. Anthropic does not own the compute. They are paying premium rent for it, on a 90-day terminable contract, to a competitor whose CEO has 85.1% voting control of the parent. That is the question every Anthropic IPO underwriter will get asked first by every institutional buyer who reads the prospectus. The answer is structurally sound — Rao’s calculus is correct, the math favors the rent today — but the answer requires the trust of an investor who has not bought any of this story before.

The private market already took that leap. The public market gets to decide whether to follow.

---

The Drag Nobody Is Pricing

While the labs raced today, the ground floor of the AI economy filed its quarterly status report. Acrisure — the Grand Rapids insurance technology firm — announced 2,250 layoffs today, citing AI directly. That is 11% of the workforce, on top of 400 cut in October. CEO Greg Williams: “It’s about changing how work gets done.” Same week, the same company opened a $184 million amphitheater downtown with naming rights and a $30 million gift.

On the same day, KPMG embedded Claude in 276,000 seats globally. Different clocks. Acrisure can fire 2,250 in eighteen months. KPMG embedding Claude across global tax and PE workflows is a multi-year change-management project, even with executive air cover. Real enterprise AI revenue, at scale, is a 2027-2029 story. The capex required to deliver it is being booked in 2025-2026. The valuations being asked of the public market this summer are priced as if those clocks match. They do not.

And the bottom of the inference stack is doing the labs no favors. Tomasz Tunguz published a chart today showing Google’s API prices have tripled over the last year, OpenAI’s are up 40%, and Anthropic’s are flat with cuts at the top of the line. His own pricing read: “Cuts when cash is plentiful and share matters. Increases when cash is tight and margins matter.” The latter is the case for all three frontier labs now. Meanwhile, CNBC reported today that running an identical ten-evaluation enterprise workload costs $4,811 on Claude versus $1,071 on DeepSeek — a 9x premium. Chinese-model usage on OpenRouter has gone from 1% to 60% in 24 months. The premium pricing the trillion-dollar valuations require is being eroded fastest in the exact enterprise segments the IPO prospectuses need to dominate.

The Bessemer math requires demand to scale at the speed of the model curve. Demand does not scale at that speed. It scales at the speed of CHROs writing severance plans, CIOs renegotiating vendor contracts, and middle managers refusing to fire people they have worked alongside for ten years. The technology is real. The organizational absorption is slow. The capex came in early. The revenue arrives late.

That is why the music will stop. The capex needs the revenue. The revenue needs the change. The change takes longer than the capex.

---

After the Bell

There is a feature of the IPO transition that the recent boom in private secondaries has obscured: private capex is funded by patient money. Venture LPs, sovereign wealth funds, and corporate strategic investors write multi-year checks against lumpy theses and accept long holding periods. Public capex is funded by operating cash flow, by debt against hard assets, or by selling stock at a price the market is willing to pay. Each option compresses the moment growth doesn’t show up on the next earnings call.

The mechanics are unforgiving. A freshly-public lab that misses on enterprise revenue conversion in its first or second quarter watches the stock fall. A falling stock cannot be used to fund operations through issuance — or if it is, the dilution accelerates the fall. Anthropic’s roughly $15 billion in annual rent through 2029 is locked. SpaceX’s orbital roadmap depends on $185-per-kilogram Starship launches that have not yet been demonstrated. OpenAI’s compute build was structured against revenue ramps the Ramp data is not validating. The capex commitments do not pause when the stock does.

The historical analogue is the SaaSpocalypse of 2022-2023. Public software companies that had funded growth through stock issuance suddenly couldn’t — because their stock prices had been pre-emptively marked down by an institutional market that saw the macro change before any of them missed a quarter. The companies survived. They spent two years in the wilderness explaining why their capex math still worked at half the valuation, and they missed the AI inflection because they were busy defending their cap tables. The public market is forward-looking when it wants to be and finicky when it doesn’t.

There is an implicit Microsoft put under OpenAI. Microsoft owns roughly 49% of the for-profit subsidiary’s economic interest, has board influence, and every commercial reason to absorb the company rather than watch a competitor acquire it at distress. That put exists at a price. The price is some discount to whatever OpenAI’s market cap is on the day Microsoft chooses to exercise it. Microsoft’s fiduciary duty runs to Microsoft’s shareholders — not to OpenAI’s IPO buyers. The investor who pays $1 trillion at the bell does not get the put at $1 trillion. They get it at whatever lower price the market clears at, on whatever schedule the Microsoft board judges accretive. That is the gun on the table beneath the OpenAI IPO. The IPO buyer is not the protected party.

---

The Mirror

The eight seats in the trillion-dollar club did not become eight by accident. They became eight because the public market, after 250 years of practice, is quite good at deciding which businesses can earn the kind of money the public can lend against. The bar is steep. The bar moved up, not down, in 2025-2026 as the macro tightened.

The labs are arriving at the bar at the moment the bar is highest. They are not arriving with decades of receipts. They are arriving with growth and a TAM claim and a capex profile that, in the existing club, is reserved for monopolies. The private markets already ran the easy multiple. If you held Anthropic at the $50 billion mark or OpenAI at the $90 billion mark, you are not the buyer of the IPO — you are the seller. The IPO is the exit for the people who got in early. The retail buyer and the public-market index fund are being asked to take supply at peak optimism, against a macro that has been unfriendly to tech IPOs for two years running, with institutional buyers who have read Tunguz too.

You can vote with your portfolio or your operating budget. Both answers are the same. Long Anthropic on private secondaries while they still trade, eyes open on the rent. Long SpaceX with a clear-eyed read on Starship reliability and the Cursor flywheel underneath. Treat the OpenAI IPO with the skepticism the Ramp data and the cheap-inference math both already justify. Do not assume all three chairs are real.

---

THE PAYOFF. Three lobsters. Two chairs. The music has been playing for a long time and most people in the room are dancing. Tonight Anthropic disclosed $1.25 billion a month in rent. SpaceX disclosed a $28.5 trillion TAM. OpenAI is filing inside the week against a Ramp index that says its enterprise share is flat. The Bessemer math requires demand the change-management clock will not deliver fast enough. The cheap-inference math requires premium pricing the marginal customer is already routing around. The capex doesn’t slow. The IPO window doesn’t widen. The chairs are not multiplying. When the music stops — and it will stop — you want to know which lobster you bet on. Long the operator. Long the optionality. Short the brand. Find your chair before the song ends.

The music is real. The seats are not.