Compute: Why Scarcity Is a Market Myth

The Thesis: Compute Is a Commodity Disguised as a Moat

Compute will be priced as a commodity by 2027. The consensus treats it as AI's ultimate moat. The difference amounts to hundreds of billions in capex already committed.

This is a regime change, far more than a trend. Value migrates from the layer that produces tokens to the layer that knows how they get used.

The data proving it comes from an unexpected source: real conversations with the models. Whoever controls compute capacity remains blind to how that capacity gets consumed. True scarcity resides elsewhere.

Why the Consensus Has the Wrong Frame

The consensus watches the easy metric: number of GPUs installed, gigawatts of datacenters announced, quarterly orders to Nvidia. Those numbers climb vertically. The conclusion seems obvious: whoever controls compute wins.

The metric that truly predicts the market is different. What counts is how people use these models, day by day. Here the providers show a carefully selected slice.

An academic project, the AI Observatory at Stanford's STAIR Lab, aggregated seven datasets of real conversations collected with user consent between 2023 and 2025. Applying the methods of the Anthropic Economic Index, the researchers found that 48% of conversations would have landed in the filter, as unrelated to work. MIT Technology Review documents it.

The differences are wide. Health and relationships account for 44.2% in the independent data, versus 31.2% in Anthropic's analysis. OpenAI's 2025 report confirms the general picture: a mere 30% of consumer use concerns work.

The consensus sizes compute around productivity demand. Real demand arises from health, relationships, companionship, entertainment. Whoever plans capacity is modeling the wrong market.

The discrepancy is far from marginal. If half of real traffic stays invisible to planning models, the estimate of future demand rests on an incomplete base. Whoever sizes datacenters around productive workloads underestimates the nature of the uses that truly saturate capacity. The risk runs deeper than overbuilding. It lies in building for the wrong demand profile.

The Cost Curve Tells the Conclusion

The cost per unit of intelligence follows a documented downward trajectory. GPT-4-equivalent performance cost figures in the dollars per million tokens at launch in 2023. Today it runs at a fraction of that figure.

My projection stays explicit: GPT-4-equivalent performance will reach commodity prices by the end of 2026. The margin abandons the model layer and shifts toward the application.

Three data points define the trajectory: the collapse of the per-token price between 2023 and 2025, the arrival of competitive open models, hardware efficiency doubling every generation. The same solar curve, down more than 80% over the past decade according to the IEA, shows how abundance works.

The mechanism is simple. Each generation of accelerators improves efficiency per watt. Open models compress the prices of closed providers. Competition across layers transforms a scarce good into infrastructure.

The solar curve stands as more than a decorative analogy. It is the same pattern: a technological input that falls in price predictably until it ceases to be the limiting factor. When the cost of producing a token tends toward zero, the advantage departs from producing it. It lies in knowing which token serves, for whom, and in which context. That is where the cost trajectory shifts market power.

The Cliff Event: When Abundance Becomes Visible

Cliff event: inference compute will drop below the threshold where the marginal cost of a query ceases to count for whoever runs it. Expected date: 2027.

At that point compute capacity becomes base infrastructure, like broadband after 2005. Differentiation vanishes from the physical layer and climbs back toward software.

Adoption jumps, far beyond growing linearly. When every individual operator accesses near-free intelligence, single-person companies with unicorn revenues emerge. This is the discontinuity that today's capex ignores entirely.

Three Categories Transformed by 2028

The transition rewards whoever owns something rare. Compute becomes abundant, so it ceases to be an advantage. What stays valuable is what money struggles to replicate.

• Pure compute providers: whoever sells generic capacity ends up competing on price, like a utility.
• Foundation model startups lacking proprietary data: their moat evaporates along with the price.
• Enterprise procurement: multi-year contracts on generic capacity lock value into a commodity.

Value concentrates where the vertical data moat lives. Whoever owns proprietary data on real usage, the kind that even the large providers struggle to read, builds the advantage of the next decade.

My Position, and What Would Change It

The position is clear-cut: compute is the commodity of the next cycle, while verified knowledge of real usage is the scarce asset. The AI Observatory demonstrates that even the providers operate in the dark on half of their traffic.

I would change my mind faced with precise evidence. A plateau in hardware efficiency, with the per-token price stable for four consecutive quarters, would falsify the abundance thesis. A regulatory consolidation that restricts access to open models would produce the same effect.

The limit of the evidence deserves naming. The AI Observatory aggregates data collected with consent: a sample, rather than the whole universe of traffic. The price projection rests on two years of curve, rather than a decade. These data suffice to define a trajectory, rather than to guarantee it. For this reason the thesis carries an explicit kill signal, in place of a certainty.

Until then the trajectory stays clear. 90% of analysts have the present right. They have the pace of change wrong.

What It Means for Those Deciding Now

The practical reading changes according to role. Each decision-maker views the same curve from a different angle.

• CTO: reassess the stack built around a single model provider, before portability becomes obvious.
• Venture Capital: the impossible yet correct bet is the vertical startup with proprietary data, far ahead of the model producer.
• Chief Strategy Officer: any three-year plan that assumes scarce compute describes a world headed for extinction.
• Technology Procurement: the multi-year contract on generic capacity locks value into what will become a utility.

The Forecast

Forecast: GPT-4-equivalent performance will reach a commodity price, below one cent per thousand tokens at the leading providers, by December 31, 2026. The margin will shift measurably toward the application layer.

Confidence: high on the technology, medium on the exact timing. Horizon: end of 2026.

Kill signal: the price per thousand tokens of the GPT-4 class stays above one cent at the three leading providers throughout 2026. That fact would render the thesis wrong.