$35B, Five Chairs, One Name

Anthropic agreed on Sunday to pay $35 billion for compute from Lambda, in a Texas data center Lambda doesn't own, running chips Lambda doesn't make. The name on the building's lease is Nvidia. The name on Lambda's cap table is Nvidia. The name that committed up to $10 billion to Anthropic's cap table last November is Nvidia. The headlines called it an Anthropic-Lambda deal, and Lambda is barely involved.

Count the chairs.

Here's the structure, pieced together from Bloomberg, the Journal, Reuters, and the FT, because every company involved declined to comment.

Hut 8, a Bitcoin miner that pivoted to data centers, is building a campus called Beacon Point in Nueces County, Texas. In May it signed a 15-year lease for 352 megawatts with a tenant it would only describe as investment grade: $9.8 billion of base rent, $25.1 billion if every renewal gets exercised. In July it signed a second lease, and the base value doubled to $19.6 billion. Then the FT named the tenant. Nvidia. A chip company is renting a shed, built to Nvidia's own reference architecture, and the shell got financed on Nvidia's signature (S&P moved Nvidia up to AA in June; Anthropic carries no investment-grade rating).

Now Lambda. Nvidia invested in its Series D and came back for the Series E. Nvidia is Lambda's only GPU supplier, which you'd expect, and its largest customer, which you might not: in 2025 Nvidia agreed to lease roughly 18,000 servers back from Lambda for about $1.5 billion. Three weeks ago Lambda borrowed $917 million, Morgan Stanley leading, to buy more Nvidia chips. Orders came in near $2 billion. Lenders love paper that terminates in Santa Clara.

Then on Sunday Anthropic, which already carries a commitment to run up to a gigawatt on Nvidia systems, signed for 350 megawatts of Beacon Point, through Lambda.

Tally it. Landlord's tenant. Operator's investor. Operator's supplier. Operator's biggest customer. End customer's investor. Five chairs, one name. Pull Nvidia out of the diagram and Nueces County is a field with a substation, because every dollar of credit in the deal was borrowed against Nvidia's willingness to sit in the next chair over.

Every module imports the same class.

Engineers have a name for this: the God object. One class that holds all the state and gets called from everywhere. Here's the thing product managers never believe when you ask for a quarter to refactor one: a God object runs fine. Beautifully, even, because one thing doing everything has zero integration bugs. The bill comes when you try to test anything else on its own. Every test needs the God object mocked, and mocking it means restating what it would have done, so every other component's behavior is just the God object's, seen from a different angle. Everything else is a wrapper.

Now the other deals. Nscale's $45 billion, signed the week before Lambda's: 460 megawatts of Nvidia Vera Rubin in West Virginia, from a company Nvidia has invested in twice. SpaceX's roughly $45 billion: more than 220,000 Nvidia GPUs at Colossus, whose builder, xAI, took up to $2 billion from Nvidia in a round built to buy Nvidia chips. Microsoft's $30 billion of Azure: a gigawatt of Nvidia Grace Blackwell and Vera Rubin, in the same press release as Nvidia's $10 billion. Fluidstack's $50 billion is for a neocloud whose price list reads GB200s and B200s.

The deals that skip Nvidia are Amazon's and Google's, and those run Trainium and TPUs, silicon the landlord makes itself. Every other line on the list ends at the same address.

Nvidia reported its quarter on August 26: $96.2 billion of revenue at a 75% gross margin, and a guide of $108 billion for the next one. The God object is the healthiest thing in the codebase. And down in the CFO commentary sits a table of future commitments that runs to $366 billion, sorted into Nvidia's own categories: supply, cloud services, data center leases, equity investments, capex. It discloses 15-year data center leases starting in fiscal 2028 or 2029, which lines up with Beacon Point's energization date. It says the equity book targets "AI model makers, infrastructure financiers, and other private companies." And it says that under some AI-cloud agreements Nvidia will "participate in revenue share generated by the AI clouds from their third-party customers." The filing leaves the clouds unnamed. If Lambda is one, Nvidia takes a sixth chair once Anthropic's rent clears the floor: a cut of Anthropic's payments to Lambda, on chips it sold Lambda, in a building it leases from Hut 8.

Unlock a focused set of AI strategies built to streamline your work and maximize impact. This guide delivers the practical tactics and tools marketers need to start seeing results right away:

Stay ahead of the curve with these top strategies AI helped develop for marketers, built for real-world results.

Download the Free Report

The price is a transfer price.

So what is $35 billion the price of?

A price carries information when the two sides want different things. Anthropic wants cheap compute, Lambda wants margin, Hut 8 wants rent, Morgan Stanley's syndicate wants its coupon. When one balance sheet sits behind three of those four, the number stops being a market-clearing price for compute and becomes something closer to a transfer price: what a company charges itself to move goods between divisions. Nvidia's compute is worth what Nvidia's ecosystem says, and in this deal the ecosystem is mostly Nvidia.

My diligence template has a line for customer concentration: anyone above about 20% of revenue gets a paragraph. It has a line for related-party revenue: sales to your investor's portfolio companies get a haircut, because the buyer's judgment came pre-installed. It has no line for a counterparty that is the investor, the supplier, the customer, and the landlord at once, because until this year I had never met one.

Frankly, the other side of this is strong, and older than software. Western Electric built every phone in the Bell System. IBM leased its mainframes rather than sell them. When your customers are supply constrained and borrow at worse rates than you do, lending them your balance sheet is the rational move, and shareholders pay you 75% margins for the privilege. The structure works right up until somebody outside the group has to price the asset: the syndicate when Lambda's loan amortizes, or the S-1 reader deciding what $80 billion of compute contracts signed in a single week are worth as a liability. Each of them needs an independent price, and every price in the room was set with Nvidia on the other side of the table.

I added the missing line to my template on Monday. Then I went looking for what I'd missed, and every fact in this piece was public before Sunday: the leaseback last September, the Series E in November, the Hut 8 lease in May and again in July, the loan on August 10. I had read most of them. I filed each one under a different company, and the only name in all five folders never got a folder of its own.

So, an assignment. The next time one of these numbers crosses your feed, skip the dollar figure and count the chairs. Investor. Supplier. Customer. Landlord. Lender. Guarantor. Write a name in each. If one name fills three or more, read the headline the way you'd read a benchmark the vendor ran on its own hardware: real, and comparable only to itself.

— SWEdonym