Here is the contradiction most readers miss. Jensen Huang stands before a half-trillion-dollar underwriting syndicate and calls GPUs "productive, long-lived, fungible and flexible" assets — then, in the same breath, concedes the whole thing turns on whether "GPUs really retain their value when newer generations of the chips emerge."
Both statements are true. That is the problem.
Reportedly $1.65 trillion in hidden borrowing per Fortune, with off-balance-sheet leases treated as debt-equivalent liabilities per Moody's. This week the wrinkle sharpens: the leases have grown teeth, stretching to 15, 16, 25 years, and a new species of financing vehicle has appeared in which the tenant owns the landlord.
The leases have grown teeth, stretching to 15, 16, 25 years — and a new species of financing vehicle has appeared in which the tenant owns the landlord.
Start with one lease and do the arithmetic out loud.
On August 10, Fermi Inc. signed its first binding tenant: TensorWave, for a facility in the Texas Panhandle. The number is $6.5 billion in contract revenue over an initial 15-year term, for 222 megawatts starting to deliver in the second half of 2027.
That facility is designed to house tens of thousands of next-generation AMD Instinct GPUs.
Now hold the industry's own cadence against it. Chip generations reportedly turn over roughly every three years. A GPU installed in 2027 would, by the midpoint of that 15-year lease, reportedly be four or five generations obsolete — running inference workloads at a fraction of the efficiency of whatever ships in 2033.
The renewal options do not soften this. They extend it: two five-year extensions push the commitment to 25 years. Someone is being asked to underwrite a quarter-century of rent against silicon with what could be characterized as a three-year half-life.
And the backstop? Fermi expects "some obligations under the lease to be guaranteed by a global AI leading enterprise, whose identity has not been disclosed."
Read that again. The confidence holding up a 25-year bet rests on a guarantor Fermi will not name.
This is the physical shape of the mismatch. The building lasts decades. The power lasts decades. The chips inside — the only thing generating the revenue that services the lease — do not.
Huang's asset-class reframing exists to bridge exactly this gap. His pitch to Apollo, BlackRock, Blackstone, Brookfield, Goldman and KKR is that compute is no longer a depreciating box but infrastructure "like electricity, like the internet." KKR's co-CEOs bought it, invoking "delivery, not ambition" as the hard part.
The reframing is elegant because it changes the question. If compute is depreciating hardware, a 15-year lease is madness. If compute is infrastructure, a 15-year lease is prudent. Same asset, opposite conclusion — and the only thing that decides which is true is whether demand for that specific compute holds for fifteen years.
Meanwhile the demand side keeps writing checks. The four largest hyperscalers — Alphabet, Microsoft, Meta, Amazon — now carry roughly $2 trillion in purchase commitments, Alphabet alone near $811 billion, per analyst estimates. Moody's puts hyperscaler lease commitments at $1.2 trillion, more than $820 billion of it for data centers that do not yet exist.
The scale is meant to reassure. It should do the opposite, because the commitments are not independent bets. They are the same bet, placed repeatedly.
Watch how the guarantees actually route. Look at Bitdeer, the Bitcoin miner that pivoted to AI hosting, and its 16-year, $4.7 billion lease to a customer called Volta — reportedly Anthropic — for 121 megawatts of Nvidia chips in Norway.
The revenue is backstopped by roughly $1.3 billion of letters of credit arranged by JPMorgan affiliates and another major financial institution. Solid, until you read the fine print: those guarantees are "subject to customary conditions," and Volta can walk away without a fee after year ten.
So a 16-year lease has firm credit support for a fraction of its value, escapable at year ten, on a build whose first half must begin operating by December 31, 2026. The headline says 16 years. The actual enforceable commitment is far shorter and far softer.
Now the new species. On the same day as the Fermi lease, Anthropic, Macquarie and Singapore's GIC launched Theseus Infrastructure — a platform to "develop, operate and lease data center infrastructure at scale to Anthropic under long-term agreements."
Read the structure carefully. Macquarie and GIC fund the majority of the equity. Anthropic is the anchor tenant. The platform's entire purpose is to build sites for which Anthropic is the customer.
The tenant is, in effect, anchoring its own landlord. The lab is the demand, the lab is the reason the equity gets deployed, and the lab's growth is the sole assumption that makes any of it pencil out. Theseus is a lab-owned SPV in everything but name.
This is the sleight of hand. Every backstop, traced to its source, points back into the same closed loop of counterparties.
Nvidia is reportedly in talks to backstop up to $250 billion of OpenAI's debt for its Ohio campus — lending OpenAI its credit so OpenAI can lease the buildings that will house Nvidia's chips. Nvidia guarantees the tenant; the tenant buys from Nvidia. Sharon AI signs a six-year, $4.9 billion "collaboration" with Nvidia that gives it access to as many as 40,000 GB300 GPUs and, Sharon says, "helps de-risk the hardware buildout." The de-risker is the seller.
Moody's named this plainly: hyperscalers invest billions into labs like OpenAI and Anthropic, which spend those billions on cloud from the same hyperscalers — a "circular AI ecosystem" in which "many of the industry's biggest companies are increasingly dependent on the same AI customers and the same assumptions about future demand."
That last clause is the whole game. The same assumptions about future demand.
Huang's phrase for his syndicate was that these capital providers would "independently underwrite AI infrastructure." But independence requires that the parties fail for different reasons. Here they do not.
Trace the Bitdeer lease: the tenant is Anthropic, the chips are Nvidia, the credit is JPMorgan's — but the thing being underwritten is Anthropic's demand for Claude compute in 2035. Trace Theseus: Macquarie's and GIC's equity, Anthropic's tenancy — underwriting the same demand. Trace Fermi: TensorWave's lease, AMD's silicon, an unnamed "global AI leader" as guarantor — same demand, one layer removed.
When the guarantor, the tenant, the chip supplier and the equity sponsor are all exposed to whether a handful of labs keep growing, the letters of credit and the guarantees are not diversification. They are the same risk wearing four costumes.
This is why the murky unit economics matter — not because we can solve them here, but because their murkiness makes the duration risk unforecastable. Nobody signing these 15-to-25-year terms can tell you what a 2027-vintage GPU rents for in 2035, or whether the inference market absorbs five generations of stranded silicon, or whether open-weight Chinese models undercut the pricing power that OpenAI's near-trillion-dollar valuation assumes.
What the structures do accomplish is to move the question somewhere you cannot ask it. Off the hyperscalers' balance sheets, into SPVs and neoclouds and pivoted miners. Onto guarantors who will not print their names. Behind letters of credit "subject to customary conditions" that lapse at year ten.
A demand correction does not disappear in these structures. It just becomes impossible to locate in advance — because every backstop resolves to a counterparty exposed to the identical assumption, and a chain of guarantees that all break at once is not a chain at all.
The 15-year contract feels like safety. It is not a backstop. It is a longer fuse.