There Can Be Only One

THE NUMBER: 90 days — the mutual cancellation notice on the lease that runs Anthropic. Colossus 1, in Memphis. Three hundred megawatts. More than 220,000 Nvidia GPUs. One and a quarter billion dollars a month, which is fifteen billion a year, and it goes to SpaceX. That number isn’t a leak or an estimate. It’s in SpaceX’s own S-1, filed May 28: “the customer has agreed to pay us $1.25 billion per month through May 2029.” The customer is Anthropic. And buried in the same filing is the clause that should reorganize how you think about this entire cycle: either party can walk away on ninety days’ notice. Elon Musk went on X the day it published to call it a 180-day lease. His own registration statement said three years. Nobody has cleared it up since. So when the CEO of Anthropic sat down on Saturday night and explained to eight and a half million people that artificial intelligence structurally concentrates power toward whoever holds the most compute and chips, he was describing his landlord.

Six days, five men, no journalists

Nothing about this happened in a press release, which is why almost nobody has it.

It starts Tuesday, August 12. Dwarkesh Patel, who runs the best interview show in the industry, posts a critique of Anthropic’s published Claude Constitution. His framing is a lawyer: “My lawyer is obligated to [defend me] in all but the most extreme circumstances; he will even defend me if he knows I’m guilty. In contrast, the Claude Constitution places the AI’s highest priority as Anthropic’s definition of the good of humanity.” His worry is that no frontier model ends up being your advocate, at a moment when who to vote for and how to invest and what news to trust all route through one. Six hundred and forty thousand views.

Wednesday, August 13. Gavin Baker, who runs Atreides and is one of the more listened-to technology investors alive, quotes it and raises: “A monopolar or bipolar world of frontier models is a very dangerous world for humans. This is why open-source is so important.” He adds the line that turns a technical dispute into a philosophical one. “Objective truth is just that: objective. What is ‘best for humanity’ is subjective.” Same day, separately, he notices the FT floating a three trillion dollar figure for Anthropic and dryly observes he feels like he’s heard that somewhere before.

Friday, August 14. Baker goes on All-In and says the quiet part: “I have been told by people I trust internally that Dario believes there could be a world where there are only governments and one private company, which would be Anthropic.” And then, pointedly: “I would certainly discourage Dario from saying that ever again to anyone.” Three hundred thousand views on one clip alone.

Late that night, an Anthropic researcher does something companies almost never let employees do. Sholto Douglas answers in public, and he doesn’t hedge. “Completely false. I like Gavin’s takes, but whoever he heard this from is lying so that it fits the narrative some people so desperately want you to believe.” Then the substance: “one of the things we are most worried about is economic concentration of power. There is no world where the government should let any company have that much influence. We need competition and capitalism.” Six hundred and eighty thousand views.

Saturday morning. Baker accepts the correction and reframes it into something harder to answer. “Sholto, thank you for setting the record straight. Larger issue is that multiple very serious people in Silicon Valley have heard some variation of this and believe it to be true. And the reason it is believable to so many is that it is consistent with Dario’s public messaging.” Nine hundred thousand views.

Saturday, 6:44 PM Eastern. Dario Amodei, who by his own admission doesn’t spend much time here, publishes roughly seven and a half thousand characters in two parts. Two point nine million views on the first. Five point six million on the second.

Worth noting the week he chose. The Wall Street Journal had run a long, personal profile of his wife two days earlier, and Baker’s clip landed the same day. Whatever you make of that, a man who had gone quiet for months broke his silence the following evening. Lulu Cheng Meservey, who does communications strategy for a living, wrote the autopsy within hours: “Dario’s tweets benefit from scarcity. The fact that he doesn’t ‘spend much time’ on X makes a tweet carry more weight and is itself a statement of seriousness.” She also flagged that Sholto did the fact-checking first, which let Dario skip who-said-what and go straight to principles. That’s not an accident. That’s a play.

The sentence that prices the company

Most people quoted the graceful part, and it is graceful. Amodei lost his father to Hepatitis C a few years before the antivirals that cure ninety-five percent of patients. He wrote about it plainly, and it’s the reason the post doesn’t read like it came out of a communications shop.

But here’s the load-bearing sentence:

Overall my view is that AI is structurally a technology that tends to concentrate power, for reasons that have nothing to do with regulation (more to do with the extreme implications of the scaling laws). Open-weights do help some with this but are nowhere near a sufficient solution because they simply shift the concentration somewhat to those with the most compute and chips (which are roughly the frontier labs plus maybe hardware providers).”

Read that as a philosopher and it’s a warning about a dangerous world. Read it as an allocator and it’s a map with an X on it.

He is telling you the value pools with compute and chips. He is telling you this eight weeks before the largest initial public offering in history, at a target of two trillion dollars, and here is the part almost nobody has written down: that number does not come from this year’s revenue. Reuters reported it first. Bankers are applying an enterprise-value-to-revenue multiple to Anthropic’s projected 2028 revenue of $190 to $200 billion. Public markets normally price the trailing twelve months, or at a stretch next year. Reaching three years out is not a valuation. It is a bet that a growth curve keeps its shape for a very long time.

Which is Highlander. The whole movie is one mechanic: Immortals hunt each other across centuries, and when one takes another’s head, the loser’s power flows into the winner. They call it the Quickening. Two trillion dollars applied to a 2028 number is not a bet that Anthropic grows. It’s a bet that Anthropic is the last one standing.

Except MacLeod owned his glen

Anthropic does not own compute or chips. It rents them, at scale that is hard to hold in your head.

Over $130 billion committed to Amazon (NASDAQ: AMZN) Web Services and Microsoft Azure, of which more than $100 billion goes to AWS over ten years for up to five gigawatts, and $30 billion to Azure. An undisclosed multi-gigawatt arrangement with Google and Broadcom. And the fifteen billion a year to SpaceX for Colossus 1, on ninety days’ notice.

Now follow the ownership. SpaceX (SPCX) acquired xAI in February in a deal valued at $1.25 trillion. xAI owns Grok. And on Thursday, August 14, Bloomberg reported that SpaceX completed its $60 billion acquisition of Cursor, the AI coding company, announced back in June days after SpaceX’s own blockbuster IPO.

Sit with that for a second. Claude Code is one of Anthropic’s fastest-growing products. Cursor is its most direct competitor. Grok is its model competitor. Anthropic’s single largest operating check goes to the parent company of both, and that deal closed forty-eight hours before Dario started typing.

The landlord can also evict. Ninety days.

This is the part of the story that has no analogue in the last three technology cycles. Amazon owned its warehouses. Google built its own data centers. Microsoft owns Azure outright. Anthropic is asking public markets to price it as the eventual sole survivor of an industry while it leases the single physical input that its own CEO says the industry concentrates toward. It is a Highlander who does not own the Highlands.

Holy ground

There’s a rule in the movie that matters here. Immortals cannot fight on holy ground. It is the one place the game stops.

Watch where Amodei wants the fight to stop.

In his June 10 essay, Policy on the AI Exponential, he proposed the thresholds himself: mandatory testing for models trained above 10²&sup5; FLOPs, developed by companies with more than $500 million in AI revenue or a billion in AI R&D. Binding tests before or during deployment. Independent evaluators with access to unredacted risk reports and the developer’s most capable models. Government authority to block or deter dangerous deployments. Civil penalties scaled to global revenue.

In Saturday’s post he named the institution he has in mind, unprompted: “I am also supportive of Demis Hassabis’ ideas around a FINRA-like entity.”

Go read FINRA’s own website. “A private not-for-profit membership organization.” “A self-regulatory organization for member broker-dealers.” “Not part of the government.” “Our operations are funded by member fees, not taxpayer dollars.” The board runs to twenty-two seats, ten of them designated for industry.

That is Wall Street policing Wall Street under the SEC’s supervision. It is the closest thing in American institutional life to a system where the regulated write the rules and pay for the referee. Nobody accused Amodei of wanting it. He volunteered it.

And then there’s what Anthropic published on July 27, in its formal position on open weights: “All sufficiently capable models, open and closed, should go through mandatory safety testing,” while “exempting less capable models, such as those from startups and academia, entirely.” Nowhere in that document is there a technical definition of “sufficiently capable.” The company says whether open models are risky “should emerge from testing, rather than be decided in advance.”

The gate is real. Its location is to be determined. And the people qualified to determine it are, by construction, the people already inside it.

Three men, three private governments

Here’s where I have to be fair, because the cheap version of this argument is wrong and it will age badly.

Amodei is not hiding a moat play. His own thresholds, 10²&sup5; FLOPs and $500 million in revenue, capture Anthropic squarely. If you were drawing a line to protect yourself you would draw it under your feet, not through them. He is genuinely asking to be regulated. And there’s a real problem underneath his position: somebody has to write the test, and a regulator staffed entirely by people with no frontier experience writes a test that measures nothing. That is not a dodge. That is the actual difficulty.

But look at what happens when you put the three most articulate men in this industry side by side.

Amodei wants mandatory pre-deployment testing above a capability threshold, with government power to block, administered through a FINRA-like body.

Demis Hassabis, on July 14, published a framework asking for exactly that FINRA, with the power to screen the world’s most advanced models and to coordinate an industry-wide slowdown if dangers mount. Labs would submit models up to thirty days before release. Voluntary at first, mandatory later. To his credit, he asks for a majority-independent board, stacked with Turing Award winners and other credentialed experts.

Alex Karp, on March 13, in the middle of the Pentagon fight, said this: “The Valley should have a consortium: This is what we’re going to do, and this is what we’re not going to do.”

Three men. Three proposals. Three private governance bodies. Not one ballot between them.

They are not arguing about whether a small technical class should sit at the gate. They agree on that completely. They’re arguing about which small technical class, and underneath that, about whose commercial position the gate happens to protect. A majority-independent board of Turing Award winners is many things, but democratic is not one of them.

The hallway Karp already owns

Which brings us to the collision nobody has written, and it’s stranger than the obvious version.

Alex Karp has been naming Amodei since March. At the a16z American Dynamism Summit he warned, in language I’m not going to reproduce, that a Valley which automates every white-collar job while refusing to work with the military is going to get itself nationalized. On CNBC on July 1 he said the models “have been completely, irresponsibly, oversold,” and told Forbes the same day: “Are we really going to outsource the battlefield of this country to the consensus view in Silicon Valley? That is effing insane.” Palantir (NASDAQ: PLTR) signed Jensen Huang’s open-weights letter in July. Anthropic never did. Neither did xAI. OpenAI and Google were also absent at launch and signed later, which makes Anthropic one of two permanent holdouts out of more than a hundred and fifty signatories.

And yet.

In April 2025, Anthropic joined Palantir’s FedStart program rather than pursue its own federal certification, which is how Claude ended up running at Department of Defense Impact Level 5 on Palantir’s compliance rails. In March 2026, after the Pentagon designated Anthropic a supply chain risk for refusing to lift its restrictions on autonomous weapons and mass domestic surveillance, Karp said this out loud at Palantir’s own customer conference: “The Department of War is planning to phase out Anthropic; currently, it’s not phased out. Our products are integrated with Anthropic.”

The man who signed the letter Anthropic wouldn’t sign, who says the models are oversold, is the reason the blacklisted safety lab is still running inside the Pentagon. He doesn’t need to run into Dario Amodei. He owns the hallway.

He also defended him. Asked to pile on, Karp declined to bash anyone and took Anthropic’s side of the exact argument that got them blacklisted: “there was never a sense that these products would be used domestically,” and he’s “very sympathetic with arguments against using these products inside the U.S.” Then he proposed the consortium.

Two philosopher-CEOs, mutually dependent, publicly at odds, wanting the same kind of referee. The fight is real. So is the rent.

(The court fight, for the record: Anthropic sued on March 9, won a preliminary injunction on March 26 with the judge citing First Amendment retaliation, then lost an appellate bid on April 8. At a July 30 summary judgment hearing, Judge Rita Lin said “the record has gotten worse for the government” and signalled she would likely make the injunction permanent. As of this writing she has not ruled.)

So what do you actually do

Three things, and they’re the same three whether you run a company or run money.

Time your model swap. Don’t plan it, time it. Point your primary workflow at a different model this morning and start a stopwatch. If the answer runs past a day, your harness belongs to your vendor. Databricks’ Smart Router cut average task cost by more than thirty percent simply by routing work between models, and that only works if the work can move. Portability used to be a cost optimization. After this weekend it’s a governance hedge.

Keep one open-weight model you actually control. Not for price. For custody. When Hugging Face got breached, it tried to defend itself with a leading American frontier model and the model refused, unable to distinguish an incident responder from an attacker. They used GLM-5.2, a Chinese open-weight model, instead. Frontier access is a turnstile now: Commerce gated Fable 5 and Mythos 5 in June and lifted it weeks later, and twenty-two members of Congress want Anthropic’s incident logs by August 24. Custody is the only thing nobody can revoke on ninety days’ notice.

If you’re allocating, price the landlord before the tenant. Anthropic’s compute obligations are contracts. The two trillion is a forecast. But don’t get cute about the toll roads either: Amazon booked $53.4 billion of last quarter’s $62.6 billion in net income as non-operating gains “primarily from our investments in Anthropic,” which makes it a levered stake rather than a pick-and-shovel. Nvidia’s (NASDAQ: NVDA) own $21 billion SpaceX position is worth about $17 billion now. SpaceX trades roughly thirty-eight percent off its high, Cerebras forty-three percent off its. The landlords got repriced once already this year, and Goldman counts $1.5 trillion of hyperscaler lease commitments with about a trillion of it “uncommenced” and not yet on anyone’s financial statements.

Everyone in this trade is levered to the same forecast. The difference is that the landlord gets paid on a schedule and the tenant gets paid on a thesis.

And Anthropic may well be the last one standing. The revenue is real: Q2 came in above $11.5 billion against $787 million a year earlier, with a first positive adjusted operating profit. Nothing here says the company is a fake. It says the company is renting the one input its own CEO named as the thing that matters, from the one man least likely to be sentimental about it, on a clause you can read in a public filing.

But MacLeod owned his own glen, and Dario leases his. The Prize goes to the last one standing. The rent gets collected either way.

---

Sources:

• Dario Amodei, part 1/2 and part 2/2 — @DarioAmodei, Aug 15, 2026, 6:44 PM ET (2.9M and 5.6M views; the concentration claim, the FINRA line, the Hepatitis C passage)
• Gavin Baker’s reply to Sholto Douglas — @GavinSBaker, Aug 15, 2026 (“consistent with Dario’s public messaging”)
• Gavin Baker on a monopolar world — @GavinSBaker, Aug 13, 2026 (“What is ‘best for humanity’ is subjective”)
• Sholto Douglas, “Completely false” — @_sholtodouglas, Aug 14, 2026 (Anthropic researcher; “we need competition and capitalism”)
• Dwarkesh Patel on the Claude Constitution — @dwarkesh_sp, Aug 12, 2026 (the lawyer analogy; where the chain begins)
• Gavin Baker on All-In — The All-In Podcast, Aug 14, 2026 (the “only company left” claim)
• Lulu Cheng Meservey’s meta-analysis — @lulumeservey, Aug 15, 2026 (scarcity, steelman selection, why the personal passage works)
• How long is Anthropic’s lease with SpaceX? Opinions vary — TechCrunch, May 28, 2026 (S-1 language: “$1.25 billion per month through May 2029”; 90-day mutual cancellation; Musk’s contradicting 180-day claim)
• Anthropic is paying SpaceX $15 billion per year — Axios, May 20, 2026
• Higher usage limits for Claude and a compute deal with SpaceX — Anthropic, May 6, 2026 (“more than 300 megawatts of new capacity (over 220,000 NVIDIA GPUs)”)
• SpaceX completes its $60 billion Cursor acquisition — Bloomberg, Aug 14, 2026
• SpaceX to acquire Cursor for $60B in stock, days after blockbuster IPO — TechCrunch, Jun 16, 2026
• About FINRA and Board of Governors — FINRA (“not part of the government”; “funded by member fees”; 22 seats, 10 industry)
• Our position on open-weights models — Anthropic, Jul 27, 2026 (“all sufficiently capable models”; startups and academia exempt; no threshold defined)
• Policy on the AI Exponential — Dario Amodei, Jun 10, 2026 (10²&sup5; FLOPs; $500M revenue / $1B R&D thresholds; authority to block deployments)
• DeepMind CEO calls for an independent standards body to regulate frontier AI — TechCrunch, Jul 14, 2026, and Axios (the FINRA model; 30-day pre-release submission; majority-independent board)
• Designing a FINRA for Frontier AI — Mark Thomas, Lawfare, Jul 30, 2026 (industry capture as the primary critique of the SRO model)
• Palantir CEO Alex Karp on Anthropic and the Pentagon — Fortune, Mar 13, 2026 (“It’s our stack that runs the LLMs”; the consortium proposal)
• Alex Karp calls the AI industry “effing insane” — Forbes, Jul 1, 2026, and CNBC Squawk Box, same day (“completely, irresponsibly, oversold”)
• Palantir, Anthropic and Google bring Claude to government via FedStart — FedScoop, Apr 17, 2025 (FedRAMP High, DoD Impact Level 5)
• Where things stand with the Department of War — Anthropic, March 2026 (the two refused restrictions: fully autonomous weapons, mass domestic surveillance)
• Judge blocks Pentagon’s effort to “punish” Anthropic — CNN, Mar 26, 2026, and Judge likely to rid Anthropic of Pentagon’s “supply chain risk” label — Courthouse News, Jul 30, 2026 (Judge Rita Lin; no final ruling as of publication)
• Huang’s open-weights letter — NYU RITS, July 2026 (25 signatories at launch Jul 24, 150+ by Jul 28; OpenAI and Google absent at launch and signed later; Anthropic and xAI never signed)
• Anthropic investors target $2 trillion IPO valuation in October — Financial Times, Aug 13, 2026
• Anthropic’s $200 billion revenue forecast sets the stakes for its IPO — Reuters, August 2026 (2028 revenue of $190–200B as the multiple’s base)
• Anthropic revenue surges to more than $11.5 billion in the second quarter — Bloomberg, August 2026 (vs. $787M in Q2 2025; first positive adjusted operating income)
• August AI Index — Ramp, Aug 13, 2026 (Fable 5 at 6% of tokens, 11.4% of model-attributed spend)
• AI’s infrastructure boom is getting more leveraged — CNBC, Aug 14, 2026, citing the Goldman Sachs note of Aug 6, 2026 ($1.5T lease commitments, ~$1T uncommenced)
• Amazon Q2 earnings release, 2026 ($53.4B of $62.6B net income as non-operating pre-tax income “primarily from our investments in Anthropic”)
• The logs Anthropic won’t hand over — genai.works, Aug 12, 2026 (22 signatories; August 24 deadline)
• Highlander (1986, dir. Russell Mulcahy) — Christopher Lambert as Connor MacLeod, Sean Connery as Juan Sánchez-Villalobos Ramírez. The Quickening; the rule against fighting on holy ground; “There can be only one.”