The Utilization Assumption
The largest capital commitments in tech history rest on a number set in sprint planning, not at signing.
I. The Most Important Number in the Contract Is Not in the Contract
OpenAI has signed roughly $1.4 trillion in compute and infrastructure commitments. Its revenue barely crossed $20 billion in 2025. CoreWeave, the largest of the neoclouds, generated 98% of its revenue from take-or-pay contracts and closed the first quarter of 2026 with a contracted backlog of $99.4 billion. These are the largest capital commitments in the history of the technology industry, and they are being signed at a pace the industry has never seen.
Every one of these contracts prices three things.
Dollars per GPU-hour.
Term length.
Capacity delivered.
The negotiation happens over these three numbers, and the CFO who signs believes the quality of the deal lives in them. It does not. The fourth number, conveniently ignored: the fraction of committed capacity that converts into useful output. UTILIZATION.
Utilization is not set at signing. It is set afterward, in sprint planning meetings, by product teams who never saw the term sheet. I am going to call it THE UTILIZATION ASSUMPTION. Every multi-year compute commitment contains one, whether or not anyone wrote it down. It is a forecast of product decisions not yet made, and it determines whether the contract was a bargain or a liability.
In The Utilization Gap I argued that the scarce resource is not compute; it is the ability to use compute. In The Inference Tax I argued that inference economics silently shape which products get built. This essay closes the loop. Product decisions set compute economics, at the scale of the largest capital commitments in the market.
II. Anatomy of a Commitment
The contracts underneath the AI buildout are opaque to most people who write about AI. The structure is simple, and the structure is the story.
Start with take-or-pay. The buyer commits to paying for capacity over a fixed term whether or not it consumes that capacity. This is not the cloud model most people know. On-demand cloud converts compute into a variable cost; you pay for what you use. A take-or-pay commitment converts compute back into a fixed cost. The meter runs regardless.
The commitment usually comes with cash up front. CoreWeave’s filings show a weighted-average prepayment of 15% to 25% of total contract value across active contracts, wired before the customer receives access. Those prepayments are not a formality. CoreWeave management has said customer prepayments reached record levels and are being used to fund working capital. The buyer’s cash builds the seller’s data centers.
The terms are getting longer, not shorter. Meta’s agreement with CoreWeave is a five-year contract worth at least $27 billion, with $12 billion of committed capacity and up to $15 billion optional. Management noted average contract durations are extending alongside rising contract values.
Now read the backlog number correctly. A $99.4 billion backlog is a promise, not cash. Roughly 36% converts to revenue within two years and 75% within four, contingent on capacity being delivered. And the whole structure is levered. Neoclouds borrow against the hardware and the contracts themselves; Nvidia backstopped CoreWeave with a $2 billion equity investment and an $8.5 billion non-recourse delayed draw term loan. Follow the chain. The buyer’s utilization assumption is also the seller’s revenue assumption, which is also the lender’s collateral assumption. One number, three balance sheets.
This all sounds insane, but none of this structure is new. Take-or-pay is how pipelines, power plants, and telecom capacity have been financed for decades. Airlines run fixed fleets against variable demand. Hotels run fixed rooms. Fabs run fixed lines. Every one of those industries learned the same lesson the hard way: when costs are fixed, the entire game is yield. Airlines built revenue management into a science because an empty seat at takeoff is revenue destroyed forever. The AI industry imported the contract structure of capital-intensive industries. It has not yet imported the operating discipline.
An idle GPU-hour under a take-or-pay contract is an empty seat at takeoff. It expires worthless, and you already paid for it.
III. The Math That Flips the Deal
Run the napkin math on a $100 million, three-year commitment.
The contract buys a fixed pool of GPU-hours. The unit economics are not dollars per GPU-hour; that number was fixed at signing. The unit economics are dollars per useful token, and that number is the contract cost divided by what the capacity actually produces. At 20% utilization, every useful token carries five times its share of the committed cost. At 60% utilization, it carries 1.7 times. Same contract. Same negotiated price. A 3x difference in effective unit cost, decided entirely by what happens after the ink dries.
Now compare the two levers. A CFO who negotiates brilliantly might cut the price 20%. A product and engineering organization that moves utilization from 20% to 30% cuts effective unit cost by 33%. Moving it from 20% to 60% cuts it by two-thirds. The lever nobody negotiates is three times more powerful than the one everybody negotiates.
If you think real-world utilization sits comfortably high, the data says otherwise. When GPT-4 was trained on 25,000 A100s, the run reportedly achieved roughly 32% to 36% Model FLOPs Utilization. That is the frontier, with the best infrastructure teams on earth. Down the stack it gets worse. Cast AI’s 2026 report measured average GPU utilization of 5% across enterprise Kubernetes clusters before optimization.
One definition, because the dashboards lie. Hardware utilization measures whether the chip has work queued. Model FLOPs Utilization, MFU, measures what fraction of the chip’s theoretical compute performs useful math. A GPU can read 90% busy on a dashboard while doing 35% of its useful work, stalled on memory, waiting on other GPUs, or grinding through poorly batched requests. The headline number confirms activity. It does not confirm productivity. Cost per useful token is the only metric that survives contact with reality.
And the swing factors are software, not hardware. Continuous batching alone can raise utilization from under 20% to over 70%, cutting effective cost per token 3 to 4x with no change to the fleet. The signal I flagged in The Utilization Gap applies with full force here: the metric that matters is UTILIZATION PER DOLLAR of committed capacity. A multi-year commitment locks the denominator. Everything that happens to the numerator is a product and engineering decision.
That is the uncomfortable arithmetic underneath a trillion dollars of signed paper. The contracts fix the cost. The product sets the value. And the two functions rarely sit in the same room.
IV. Who Sets the Number
The utilization assumption is a forecast of decisions that are yet to be made. The decisions are:
Context window defaults. In The Inference Tax I showed that KV cache memory scales linearly with context length and concurrency; fifteen users at 32K context create the same memory pressure as sixty users at 8K. By setting a default context window, a product manager ends up setting the utilization curve of committed capacity.
Batching policy. Continuous batching, the scheduling technique that keeps a GPU full by interleaving requests instead of processing them one at a time, can move utilization from under 20% to over 70%. It is an engineering decision, made in a design review, invisible to finance.
Caching strategy. Semantic caching can eliminate a large share of redundant inference calls in high-repetition products like customer support. Every cached response is committed capacity freed for revenue-generating work.
Model routing. Sending a CSS autocomplete to a model that costs a fraction of a cent while reserving the frontier model for multi-file refactors is the difference between a fleet that serves ten times the traffic and one that drowns. Cursor built this into its product as Auto mode. Routing is product architecture.
Quantization. Running models at lower numerical precision, FP8 instead of FP16, let one team serve 1.8x the traffic on the same four-GPU footprint. A model quality decision with a capital efficiency consequence.
Agent design. A single agentic interaction triggers dozens of sequential inference calls with context that grows over the session. Agents change the shape of demand, from smooth request streams to bursty, long-running sessions. Demand shape determines how much of a fixed fleet can ever be utilized. The PM designing an agent workflow is designing the load profile of a nine-figure asset.
Do any of these levers appear in the contract? None. Yet, every one of them moves the number that determines whether the contract was a good idea. The utilization assumption, written down or not, is a bet that a product organization nobody consulted will make six categories of decisions well, for years, under feature pressure that pushes every one of them the wrong way.
V. The Depreciation Clock Runs on the Same Variable
When a company buys a GPU, the cash leaves immediately, but the expense hits the income statement gradually, spread over the asset’s estimated useful life. That estimate is a management judgment, not a fact. A longer life means lower annual expense and higher reported profit from identical hardware. If the estimate proves optimistic, the correction arrives later as an impairment, a write-down that lands all at once. Depreciation 101, applied to GPUs.
In November 2025, Michael Burry (yes, the guy from The Big Short) accused the hyperscalers of exploiting exactly this. His claim: they depreciate Nvidia hardware over five to six years while Nvidia ships new architectures annually, understating depreciation by roughly $176 billion between 2026 and 2028. The estimates have real spread. Amazon shortened the useful life of a subset of servers while Meta extended its estimate to 5.5 years, and Meta’s extension lowered its depreciation expense by $2.3 billion over nine months. Goldman’s sensitivity analysis found that moving the industry from five-year to three-year schedules would swing cumulative depreciation by roughly $1 trillion through 2031. Satya Nadella said it plainly: “I didn’t want to go get stuck with four or five years of depreciation on one generation.”
The industry’s defense is the cascade model. Chips do not die when the frontier moves; they fall down the stack, from frontier training to inference to smaller-model serving, generating revenue for six years. It is credible: an H100 system resold in its third year at about 45% of new price.
Both sides are arguing about the wrong thing. The cascade defense is a utilization argument. A six-year useful life is defensible if and only if product workloads exist to fill years three through six of the chip’s life. Whether those workloads exist is not an accounting question. It is a product roadmap question. Every depreciation schedule in every hyperscaler 10-K is an implicit product roadmap, a bet that someone will build the inference workloads, the small-model serving, the batch pipelines that keep a 2024 chip earning in 2029. The accountants are making a product bet and calling it an estimate.
I lived in this world of inference accelerators. The datasheet tells you what a chip can do. The workloads that exist for it tell you what it will earn. Silicon roadmaps and deployment reality diverge constantly, and the gap between them is set by software and product decisions, never by the silicon itself. The economic life of a chip is a product outcome dressed in an accounting costume.
VI. The Organizational Gap
Now put the two halves of the company in the same frame.
The commitment is negotiated by finance and infrastructure. The utilization is produced by product and engineering. In most organizations, these functions do not share a model, a metric, or even a meeting. The CFO holds a spreadsheet with a utilization cell someone typed a number into. The product team holds a roadmap that will determine the real value of that cell. Neither knows what the other has assumed.
Enterprise clusters measured at 5% average utilization are what this looks like at the bottom of the market. Nobody owned the number, so the number found its natural level.
The fix is a role most org charts do not have yet: the UTILIZATION OWNER. Someone who models compute economics, diligences providers, and connects capital commitments to product decisions.
The forcing function is cash. Combined 2026 capital expenditures across the four largest hyperscalers have reached approximately $725 billion, up 77% from 2025, and free cash flow is compressing; Alphabet’s first-quarter free cash flow fell 47% year over year. When cash was abundant, low utilization was an invisible tax. When cash gets tight, utilization becomes the margin story, and someone has to own it.
VII. What This Means for Operators and Investors
For operators, the rule is simple to state. No multi-year compute commitment gets signed without a product-side utilization model attached. It cannot be just a cell in the CFO’s spreadsheet. It has to be a model that maps the product roadmap to a load profile: which features drive demand, what context and agent architecture they assume, what batching and routing infrastructure exists to serve them. Owned by the UTILIZATION OWNER. Reviewed quarterly against actuals, with one question on the agenda: which product decisions moved utilization this quarter, and in which direction.
For investors, I propose a simple rubric: ask any AI company carrying committed compute who owns the utilization assumption. If the answer is the CFO, the number is a guess. If the answer is a product leader with a model and a review cadence, the company understands its own cost structure. The tell is already visible in the market: companies running the same models on similar contracts are printing different gross margins, and the difference is not negotiation. It is INFERENCE LITERACY, the skill I named in The Inference Tax, operating at the scale of capital allocation instead of feature design.
The claim is falsifiable. Companies that fuse compute procurement with product planning will show it in gross margin within four quarters of signing. Watch the cohort of AI companies that signed major commitments in 2025 and 2026. The margin divergence among them will map to who owned the utilization assumption, not to who negotiated the better rate.
VIII. The Close
We constantly argue about the price of compute. Price per GPU-hour, price per token, price per gigawatt. The contracts got bigger, the terms got longer, and the negotiations got more sophisticated. All of that effort to optimize the number that matters least.
The contract is signed in a conference room. The economics are decided in the backlog.
Every take-or-pay commitment, every prepayment, every depreciation schedule, every GPU-backed loan in the AI buildout rests on the same buried variable, and that variable belongs to product. The utilization assumption never appears in the deal model. It lives in sprint planning, in context defaults, in routing tables, in agent architectures. Master the utilization assumption and the price negotiation becomes a footnote. The product makes the contract a bargain.
This is Part IV of the compute economics series. The Utilization Gap covered where infrastructure value accrues. The Inference Tax covered how inference economics shape products. The Concentration Thesis covered why inference concentrates. This essay closes the loop: product decisions are capital allocation.