The most interesting analog for the GPU economy isn’t software-as-a-service. It’s project-financed oil and gas.
When CoreWeave raised a $2.3 billion debt facility against Microsoft contract revenue in 2023, the structure was textbook offtake-backed project finance — a familiar shape to anyone who has financed a refinery, a pipeline, or an LNG terminal. A long-dated take-or-pay contract from a creditworthy buyer is collateralized against the physical asset that fulfills it. The lender doesn’t underwrite the operator’s balance sheet; the lender underwrites the contract.
This is not how SaaS gets financed. SaaS is underwritten on metrics — net retention, ARR, gross margin — that decouple revenue from any single hard contract. GPU cloud revenue is structurally different. The revenue is a contract, dated, sized, often with a named investment-grade counterparty. That structure is what makes the offtake analogy work.
But the analogy breaks in three specific places, and the gap between where it holds and where it breaks is the thesis.
Where the analogy holds
Take-or-pay structures, capacity commitments, dedicated infrastructure, long-duration contracts. The mental model from oil & gas project finance — debt service coverage ratios, sponsor support, contract-level credit enhancement — maps cleanly onto a hyperscale GPU contract. A $1.5 billion facility against a $5 billion contract is sensible if you can verify the counterparty.
Where it breaks — three places
First, asset durability. An LNG terminal lasts 40 years. A rig lasts 30. An H100 lasts five, maybe seven in best case — and the depreciation curve isn’t linear. It is dictated by what Nvidia ships next. When the H200 ramped, the residual value of H100s shifted in ways no traditional infrastructure depreciation schedule contemplates. Lenders who underwrote 7-year amortization against H100 collateral are now staring at residual-value risk that looks more like a tech-product cycle than a steel-and-concrete asset.
Second, technology obsolescence is path-dependent. Oil from the ground is fungible. Compute from an H100 is not exactly fungible with compute from a B200. Customers are increasingly specifying hardware in their contracts — and as inference workloads consolidate on optimized chips (Groq, Cerebras, Google TPUs, Amazon Trainium), the question of whose hardware sits behind the contract becomes a credit consideration, not just a technical one. This is unfamiliar territory for project-finance underwriters.
Third, the operator matters. In a pipeline project finance, the operator is a commodity service. Three years into the GPU cloud era, operator execution is wildly variable. A CoreWeave or a Crusoe with the right cap-ex discipline and customer mix performs very differently from a leveraged also-ran. Financing structures coming out of 2024–2026 increasingly require operator-level guarantees in addition to contract-level collateral — closer to corporate credit than pure project finance.
Why this matters now
The financing layer is institutionalizing. KKR, Blackstone and Apollo are writing tickets in the space. Bank syndicates are forming around the largest deals. But the long tail — mid-tier GPU clouds, inference providers without a Microsoft anchor, regional players from Singapore to Stockholm — is still pricing capital through ad-hoc, asymmetric arrangements. That is where the spreads live, and where an investor who can model the layer correctly has a real edge.
It is also where the first credit losses will surface. A GPU cloud financed on an H100 amortization schedule against a customer who didn’t extend their contract is the 2026 version of every distressed project-finance story you have ever read. The lenders who survive will be the ones who priced obsolescence risk explicitly.
What I’m watching
Three questions over the next two quarters:
One. How many of the H100-collateralized facilities written in 2023–2024 actually amortize cleanly on their original schedules. Stress points: 2026–2027 refinancings, customer contract renewals, residual-value mark-downs.
Two. Whether the mid-tier GPU clouds — Lambda, Together, Crusoe — can refinance into more permanent capital structures, or whether the bridge debt taken in 2024 starts to come due against a softer pricing environment. This is the inflection point for who survives the next 24 months.
Three. The first telemetry-collateralized credit product. Someone — probably not a bank — will write a working-capital line against API usage and inference revenue, treating the telemetry as the collateral and the runtime as the asset. When that product is structured cleanly, it changes how AI-native companies finance themselves. I expect it to show up in 2026.
If you finance, build, or back this layer, I want to know what you’re seeing. Reply with one thing I have wrong.
— Carol