Skip to content
Open Doors Partners

The Economics of Intelligence After Training: What Serious Capital Underwrites Once a Model Ships

A model’s capability is settled in the training run; its economics are settled afterwards, in the cost of serving a token.

A perspective from Open Doors Partners

_____________________________________________________________________________________________________________________________________

Artificial intelligence has been underwritten as a training story. Capital went into one large event, the run that produced the weights, on the assumption that the expensive part was then behind it. That assumption held while models were scarce and usage was thin. It has stopped holding.

The economics of intelligence are decided after training. What a finished model costs to run at scale, across billions of tokens a day and at a latency the market will accept, is answered in megawatts drawn, accelerators utilised, memory bandwidth consumed, and the distance between a query and the machine answering it. Open Doors Partners examines these companies at the point where those costs become legible, which is often later than the point at which they are discussed.

Training Is a Capital Event. Serving Is a Cost of Goods.

The distinction is not semantic. A training run is a discrete expenditure, amortized across whatever the model goes on to earn. Inference recurs with every unit sold, scaling with revenue rather than preceding it and compressing gross margin in proportion to use.

A business whose marginal cost rises with each additional customer is not a software business in the classical sense, whatever its interface suggests. It behaves closer to a utility: input costs, contracted supply, capacity planning, and the spread between what output costs to produce and what the market will pay.

In an inference-heavy business, adoption consumes capacity that must be procured, powered and cooled. Revenue growth and margin deterioration can arrive in the same quarter with nothing having gone wrong. What matters is whether cost per unit of useful output falls faster than the price the market will bear.

The Serving Stack Sets the Floor

Unit margin is set by engineering decisions that rarely surface in a commercial conversation. Throughput per accelerator turns on batch size, on the precision the weights are quantised to, on how much of the key-value cache a long context holds in high-bandwidth memory, and on whether a workload is dominated by prefill or by decode. Reasoning and agentic workloads change that arithmetic: a single request can consume output tokens an order of magnitude above a short completion, and the latency a customer notices sits at the tail. Serving to a strict tail-latency commitment costs materially more than serving on best efforts, because it requires headroom that is idle by design.

Two firms running comparable silicon on comparable weights can therefore report different gross margins, and the difference is operational rather than scientific. In this firm’s assessment it is the layer at which the sector’s cost curve is actually falling, as the hidden cost structure of AI sets out at greater length.

The Binding Constraint Is Physical

The limit on deployed intelligence is no longer principally algorithmic. It is electrical. Interconnection queues, transformer and switchgear lead times, substation capacity, direct-to-chip liquid cooling, and land with a viable power contract attached determine whether capacity exists at all. These are governed by permitting and multi-year procurement rather than engineering velocity, with grid connection timelines running in years rather than quarters (per Avanza Energy and Data Center Knowledge, 2026). Silicon has chokepoints of its own, advanced packaging and high-bandwidth memory among them, but both can be expanded faster than a grid connection can be secured (per TrendForce, April 2026). That asymmetry is taken up in the firm’s work on energy as the new constraint.

Scarcity of that kind takes years and regulatory consent to reproduce, a different quality of advantage from scarcity that lasts until the next model release. A siting decision made this year can set a cost structure for a decade, and it is made well before the capacity it produces appears in anyone’s revenue.

What Depreciation Conceals

Reported margin in an inference business rests on an assumption about how long an accelerator earns, and that assumption is a management estimate rather than an engineering fact. Between 2022 and 2024, Microsoft, Google, Meta and Oracle each extended the assumed useful life of server and networking equipment from roughly four years to five or six, with disclosed first-year benefits of more than ten billion dollars across the group (per company filings, compiled by Footnote Brief, 2026). In February 2025 Amazon moved the other way, shortening the assumed life of a subset of its servers and networking equipment to five years, citing the accelerating pace of development in artificial intelligence (per Amazon disclosure, February 2025). Two large operators of the same class of hardware reached opposite conclusions in the same period.

Hardware does not have to fail, or become obsolete, to stop being economic. It has only to become more expensive to run per token, all-in on power and rack space, than the newer accelerators available to a competitor. That point can arrive before the end of a fleet’s technical life, and before the end of the schedule financing it. Where the earning period and the depreciation period diverge, reported profitability rests in part on an accounting assumption rather than solely on the underlying economics. That is an observation about the assumption, not a conclusion about any particular company’s reporting.

The residual market for used accelerators is being tested in public rather than settled, with reported resale figures pointing in both directions (per Introl, 2026, and National Law Review, 2026). None of this is visible in a demonstration. It is visible in financing terms, contract length, counterparty quality, and the proportion of capacity committed under take-or-pay agreements. It matters more as the marginal dollar of capacity moves towards vendor financing and private credit secured against the hardware (per BlackRock commentary reported by Yahoo Finance, 2026).

Where Durable Value Settles

Frontier capability has a short half-life as an advantage: contested, expensive to hold, reset with each cycle.

Harder to reproduce quickly is power secured under long contract, capacity sited where it can be energised, distribution into workflows where switching carries real cost, and the discipline to keep expensive assets fully used. An underused facility loses money at close to the rate a full one earns it, and the gap between them is scheduling rather than technology.

None of these characteristics is protective on its own. Contracted power becomes a liability if demand does not materialise, and a build cycle running ahead of adoption produces stranded assets on long-dated obligations. These are risks to be underwritten rather than characteristics to be assumed.

Much of this work is being done while these companies remain private; the capital intensity that makes them consequential is the same intensity that makes public-market patience difficult to secure. Underwriting them there is a matter of examining cost structure rather than capability, which is the practice described in how the firm works.

Capability is decided on the run. Everything that determines whether it becomes a business is decided after it.

_____________________________________________________________________________________________________________________________________

Frequently Asked Questions

What separates training cost from inference cost Training is a discrete capital expenditure that produces a set of model weights. Inference is the recurring cost of running those weights in production, incurred with every request served. Training is amortized across future earnings. Inference behaves as a cost of goods sold.

Why inference cost affects valuation It sits in gross margin rather than in a prior period. A business whose serving costs scale with token volume carries a margin profile closer to infrastructure than to software, which is a consideration in how it is valued.

What determines cost per token Utilization of the accelerator fleet, batch size, quantization, key-value cache pressure from long context windows, the balance between prefill and decode, and the latency commitment being served. The all-in power and cooling cost of the site sits underneath all of it.

Why power is treated as the binding constraint Capacity cannot be energized faster than interconnection queues, transformer supply and permitting allow, and those timelines run in years. Accelerators, packaging and memory can each be procured considerably faster than the electricity required to run them.

Why accelerator depreciation schedules matter Reported profitability depends on the assumed earning life of the hardware, which is a management estimate that has moved in both directions across the sector since 2022. Where newer silicon makes an existing fleet uneconomic per token before the schedule ends, reported margin reflects an accounting assumption as much as the underlying economics.

Which assets have been harder to reproduce Contracted power, capacity sited where it can be energized, distribution into workflows with real switching costs, and sustained high utilization. Model capability has reset with each cycle to date. These inputs have been slower to reproduce, though none of them removes the risk of overbuilding into demand that does not arrive.

Sources

  • Footnote Brief, “The $200 Billion Question Hiding in Big Tech’s AI Spending,” 2026, citing Microsoft,
  • Alphabet, Meta and Oracle 10-K property and equipment disclosures
  • Amazon.com, Inc., disclosed change to assumed useful life of a subset of servers and networking equipment, February 2025
  • Introl, “Secondary GPU Markets: Buying and Selling Used AI Hardware,” 2026
  • National Law Review / Deep Quarry, “Useful Lives of GPUs: Key Considerations,” 2026
  • Avanza Energy, “Record Demand, Declining Construction: The Grid Bottleneck Reshaping America’s Data Center Map”
  • Data Center Knowledge, “AI Data Center Boom Rewires US Power Supply Chain”
  • TrendForce, “Semiconductor packaging and testing capacity race intensifies as ASE, Samsung and Amkor announce new moves,” April 2026
  • Yahoo Finance, “BlackRock sees private credit taking a bigger role in AI infrastructure”

—

Open Doors Partners LLC | Exempt Reporting Adviser
Informational only. Not investment advice and not an offer or solicitation to buy or sell any security. Private investments are speculative and illiquid and may result in loss of the entire investment. Forward looking statements may prove incorrect. Third party data has not been independently verified.
Read full disclosures here.