AI Compute Cost Curve: The Collapse the Market Is Ignoring

The thesis the market refuses to price

The cost of AI compute is collapsing along a brutal deflation curve. This is the documented trajectory of the last three years, hardly a reckless forecast.

The per-token price of foundation models has fallen by more than an order of magnitude from 2023 to today. The direction remains clear. Speed is the variable that consensus underestimates.

My thesis is blunt: foundation models will become a commodity within 18-24 months. Value will migrate entirely toward vertical applications with proprietary data moats. The market still prices the exact opposite.

Consensus treats this deflation as cyclical noise. I read it as a structural signal, the signature of a learning curve that is accelerating.

Consensus has the wrong frame

90% of analysts are right about the present. They are wrong about the pace of change.

Consensus fixes its gaze on data center capex and concludes that compute will remain scarce and expensive for a long time. That figure describes the input. It ignores the output.

The metric that matters is the cost per unit of useful intelligence, the effective price per token. Hardware scarcity coexists with the collapse of that price. The two dynamics run in parallel. A company can pay more for the GPU and less for the token in the same quarter. They are two distinct counters.

Anyone who confuses aggregate spend with unit price reads the curve backwards. It is the central analytical error of this cycle. Observers extrapolate the present and miss the inflection, while the cost curve remains the only reliable predictor of the pace. The consequence for the decision-maker is concrete. Anyone who anchors their plan to today's price is building on a level that no longer exists twelve months from now.

The cost curve: three points trace the route

A trajectory requires at least three points. AI compute offers crisp ones.

In 2020, generating a million tokens with leading models cost several dollars. By 2023 the price had fallen by roughly an order of magnitude. In 2025 open-weight models pushed the marginal cost toward zero.

The same dynamic governed solar photovoltaics. The International Energy Agency documented a roughly 90% cost drop over the 2010-2020 decade. The learning curve of digital hardware runs steeper still.

Three points, one direction. This is anything but a passing trend: it is a regime change in costs. The limit of this evidence must be stated clearly. Three points describe a trajectory; they do not guarantee it will continue. That is why the thesis carries an explicit kill signal, further down.

The cliff event: when adoption jumps

Technology adoption rarely grows in a linear way. It jumps abruptly at a precise price threshold.

The cliff event arrives when the cost per token equivalent to GPT-4 touches roughly 0.001 dollars. At that point, embedding intelligence into every workflow becomes a trivial accounting decision. I place that threshold by the end of 2026.

Below that level the calculation changes in nature. Intelligence stops being a premium feature and becomes default infrastructure, like broadband or electricity. No one today assesses the cost of a single query on a search engine. The same will happen to a model call.

The causal mechanism stays explicit: marginal price near zero, therefore pervasive diffusion, therefore margin abandoning the model layer. Each step follows from the previous one. If the price does not collapse, the chain breaks at the first link.

Look at the precedent of cloud computing. The price of storage collapsed, and value migrated toward the managed services built on top. AI repeats the same script, at double the speed.

Three categories that will change shape

Three categories of company that will disappear in their current form by 2027:

• Resellers of generic AI capacity
• Providers of static-signature security software
• Horizontal SaaS platforms without a data moat

Resellers of generic capacity sell what becomes a commodity. Their margin evaporates along with the per-token price. They stay crushed between hyperscalers upstream and vertical applications downstream.

Security already shows the pattern. ZDNET notes in its August 2026 update that modern antivirus goes beyond signature scanning, integrating real-time monitoring and VPN. Value shifts toward the data and the service, the commodity stays below.

Horizontal platforms without proprietary data offer workflows that an agent rebuilds on the fly. The moat becomes the vertical data, hardly the interface.

Concrete names populate each category, from GPU brokers to generalist productivity suites. Margin pressure will hit first those who sell undifferentiated volume.

My position, and what would falsify it

The founding position remains this: AI margin lives in the vertical application layer, hardly in the model. Companies buying generic capacity today are purchasing tomorrow's commodity.

Those building vertical data moats are laying down the advantage of the next decade. The economic logic converges on this outcome with remarkable force. You will find the desk's founding positions in the analysis archive.

It is worth voicing the counter-thesis the piece leaves implicit. A foundation model provider can defend its margin with a capacity advantage that open weights do not close. If that performance gap stays stable over time, pricing power returns upstream and the migration of value stops. It is the scenario I keep under observation.

What would change my thesis. A plateau in the per-token price for four consecutive quarters would indicate that the learning curve has burned out. A stable return of pricing power to foundation model providers would falsify the entire architecture of the reasoning.

Until then, the curve speaks clearly.

What it means for those deciding now

For the CTO: reassess every generic compute contract today before commoditization makes it obvious. Lock in flexibility, avoid multi-year lock-in on prices destined to collapse.

For venture capital: the bet that seems impossible is the vertical with a narrow data moat, hardly yet another model layer. The curve's data says where the margin will end up.

For the Chief Strategy Officer: any three-year plan that assumes scarce and expensive compute describes a world in dissolution. Rewrite it assuming near-free intelligence.

For technology procurement: the vendor you are about to lock into a long contract most likely sells a technology near obsolescence. Negotiate exit clauses anchored to the market price.

The forecast, with its kill signal

Here is the explicit, verifiable and dated forecast.

Claim: the cost per token of a model equivalent to GPT-4 will fall toward 0.001 dollars by 31 December 2027. In parallel, the gross margins of generic foundation model providers will fall below those of vertical applications with a data moat.

Confidence: high, 75 out of 100. Horizon: 31 December 2027. The mechanism remains the learning curve of hardware and the competition of open-weight models.

Kill signal: a list price per token of a model equivalent to GPT-4 above 0.005 dollars at the end of 2027 falsifies this thesis. It is a public number, verifiable by anyone.