The labs are leaving their landlords

Regulatory filings are not written to be read for their prepositions, which is why the one that matters in Broadcom's 8-K of April 6th slipped past almost everyone. Anthropic, the filing said, would "access through Broadcom" roughly 3.5 gigawatts of next-generation TPU capacity beginning in 2027: Google's chip design, assembled into Broadcom's racks, delivered to a lab that Google had funded and that Google Cloud was supposed to serve, with the cloud itself appearing nowhere in the sentence. Hock Tan, Broadcom's chief executive, had already described the earlier tranche of the same relationship in December with a verb his customer's landlords might have preferred he avoid, telling analysts that Broadcom had received "a $10 billion order to sell the latest TPU Ironwood racks to Anthropic," and that an $11 billion order had followed it. The word was sell, and the buyer was a company whose compute had, until quite recently, arrived as a monthly invoice from someone else's data center.

That is a long way from how the frontier labs were financed, and it is worth remembering how recently the old arrangement held. Between 2019 and 2024 the money came from the clouds on terms the clouds called partnership and that behaved, in every clause that counted, like a lien: Microsoft's $13 billion in OpenAI carried exclusive cloud rights, a right of first refusal on every gigawatt the lab needed, an IP license running all the way up to AGI and a 20% share of OpenAI's revenue, while Amazon's $8 billion in Anthropic bought primary-cloud status and a training partner whose models were being built on Trainium. The cloud was landlord, lender and shareholder at once, collecting rent on the racks, interest on the runway and a slice of whatever the lab became, and for five years the lab's economics and the cloud's economics were, for practical purposes, the same set of numbers.

Eighteen months later the labs are converting those partnerships into procurement, and the hyperscalers' claim on them is decaying from equity-like to rent-like. The backlog dollars the clouds report are real. The strategic premium embedded in those dollars is smaller than the figures suggest, and the contracts signed since April say so in writing.

Take the sequence at Anthropic. In November it committed $50 billion to data centers in Texas and New York built for it by Fluidstack, a neocloud, facilities "custom built for Anthropic" and owned by nobody in the cloud oligopoly. On April 20th it signed for up to five gigawatts of Trainium under a ten-year, $100 billion-plus agreement with Amazon that includes an option to purchase future generations of Amazon's silicon outright. Two weeks earlier it had signed the Broadcom paper, and on May 5th The Information reported the Google side of that arrangement at roughly $200 billion over five years, more than 40% of Google Cloud's disclosed backlog. On May 6th it leased the whole of SpaceX's Colossus 1 in Memphis, 220,000-plus Nvidia GPUs and 300 megawatts, priced in SpaceX's S-1 at $1.25 billion a month through May 2029, more than $40 billion over the term, with a clause allowing either party to walk on 90 days' notice. Five deals, four counterparties, two of them clouds.

Nvidia's ledger records the migration. When the company split its data-center revenue in May into "hyperscale" and "ACIE" (AI clouds, industrial and enterprise), the two halves were nearly equal in size and unequal in speed: hyperscale grew 12% sequentially, ACIE 31%. In the quarter reported on August 26th hyperscale grew 13% to $48.7 billion, while ACIE grew 25% to $40.3 billion, up 138% on the year, and Colette Kress, Nvidia's finance chief, guided that ACIE would drive most of the next quarter's growth. ACIE is where a lab's spending lands when it routes around the public clouds: to CoreWeave, to Fluidstack, to a Memphis warehouse owned by a rocket company. Nvidia itself notes that part of ACIE demand is hyperscalers renting from neoclouds, which softens the point without reversing it. The clouds are now sometimes tenants too.

Rent control

The clearest record of the conversion is the contract Microsoft rewrote on April 27th. OpenAI's October recapitalization had already stripped Microsoft's right of first refusal on compute in exchange for an incremental $250 billion Azure commitment. April went further. Microsoft stopped paying OpenAI a share of Azure's revenue from its models; OpenAI's 20% payment to Microsoft became "subject to a total cap" and runs only to 2030; and OpenAI won the right to sell its full product line on AWS and Google Cloud. An uncapped participation in the upside was exchanged for a capped royalty and a purchase order. That is the difference between a shareholder and a vendor, and it is now in the term sheet. The SpaceX contract, two weeks later, wrote the same thing from the other side: $15 billion a year of compute that the tenant can leave in a quarter.