An open market for commoditized intelligence

01 A new commodity

Intelligence is becoming a commodity.

Inference has surpassed training to become AI’s dominant cost, approaching two-thirds of all compute, up from roughly a third in 202311Deloitte finds inference is roughly two-thirds of all compute in 2026, up from a third in 2023 and half in 2025. Deloitte, “More compute for AI, not less,” Nov 2025.

This explosion is driven by the extraordinary growth in token consumption volume, which is doubling every year and on track to grow more than 20× by 203022Goldman Sachs projects token consumption multiplying 24× to 120 quadrillion tokens per month between 2026 and 2030, about 2.2× per year. Goldman Sachs Research, May 2026. By revenue, the inference market is already over $100 billion, and is set to more than double, reaching $255 billion by the end of the decade33The global AI inference market is projected to grow from $106.15 billion in 2025 to $254.98 billion by 2030, a 19.2% CAGR. MarketsandMarkets, AI Inference Market 2025–2030.

This shift turns the output of a model into something an economist would recognize as a commodity. A unit of inference, one million tokens, is now bought and sold much like a barrel of oil or a kilowatt-hour of electricity.

We are standing at the cusp of commoditized intelligence.

02 The problem

The missing instrument.

Here is the problem. Every company that builds on AI now carries a large, growing, and unavoidable inference bill. When the cost gets big enough, predictability becomes a necessity.

In every other commodity market, that exact situation calls for a futures market: a way to agree today on the price you will pay tomorrow. For oil, electricity, and carbon, such markets exist. For intelligence, none does.

04 Why prices move

Why the price fell, and why that ends.

Since 2023, the price of a fixed level of intelligence has fallen by roughly tenfold a year, one of the steepest cost declines in the history of computing. Four forces drove it down.

• Labs subsidize. The leading labs price tokens below cost to win users and lock in market share.
• Chips got faster. Each new generation of inference hardware produces more tokens per dollar, but slowly: performance per dollar improves only about 1.3× a year, doubling roughly every two years.
• Models got leaner. Distillation and smaller architectures cut the compute behind a given answer by roughly 3× a year, halving it every eight months. This is the fastest of the four forces, and also a finite one.
• Competition intensified. Open-weight challengers undercut the incumbents hard (DeepSeek launched at roughly 90% below prevailing prices), and each new frontier model turns last year’s premium capability into this year’s commodity.

Two of these are pure supply mechanics, captured in a simple identity:

Faster chips raise hardware efficiency and leaner models raise algorithmic efficiency, and both multiply the tokens a fixed amount of capital and energy can produce. Subsidy and competition then push the market price below even what those mechanics require.

But every one of these forces is near its limit, and a fifth has begun to push the other way.

• Subsidy ends. Below-cost pricing cannot outlast the race for share. Headline prices are already decelerating toward roughly 40% a year, with some services nudging back up.
• The buildout slows. Chip stock can double every seven months on paper, but it is gated by power, which grows only about 15% a year and takes five to seven years to bring online.
• Chips run short. That 1.3× a year is already near the physical floor, and a supply crunch (HBM sold out for 2026, year-long GPU lead times) can push prices the other way.
• Good enough arrives. Open models are already good enough for most work, so chasing an ever-higher frontier stops driving prices down. Buyers settle on what works, and the capability they actually use stops getting cheaper.
• Demand explodes. Above all, autonomous agents consume tokens at a scale no human ever did: business token use grew more than tenfold in the sixteen months to early 2026, a single agent burns 5 to 30 times the tokens of a chat, and demand is forecast to multiply another 24× by 2030. Their appetite is inelastic: they cannot drop to a weaker model when prices rise.

Line them up and the cross looks inevitable: efficiency buys about 3× a year and is slowing toward its 1.3× hardware floor, while demand has lately grown closer to 8× a year. The downward forces are decelerating; the upward one is not.

05 No inventory

A commodity with no inventory.

And because a token cannot be stored, nothing absorbs the gap. It is produced and consumed in the same instant, with no warehouse to draw down when demand spikes and none to fill when it falls. Supply has to meet demand in real time, so any mismatch shows up at once as a price swing. This is why electricity, which also cannot be stored, is one of the most volatile commodities on Earth. Intelligence is about to join it, and the shift comes in three phases:

Demand moves at the speed of software. Supply moves at the speed of construction. The distance between those two speeds is where volatility lives.

06 The evidence

Does the hedge actually work?

A hedge is only worth building if it measurably lowers risk. Calibrate a standard price model to the dynamics we expect, a downward trend, reversion toward a moving average, and occasional sharp upward jumps, and the answer is clear. A buyer who hedges cuts the volatility of their procurement costs by roughly 62 to 78 percent in every scenario tested. The same model shows why the need is real: about 15 percent of simulated price paths contain a spike of 100 percent or more within three years. A risk that large, with no instrument to manage it, is the gap this market fills.

07 Roadmap

How to build it.

A market like this cannot be declared into existence, and the sequence is the whole thesis. Each stage is cheap to run, proves out the next, and leaves behind an asset the eventual exchange is built on.