AI Compute Cost Curve: The Collapse Markets Misprice

The thesis, stated as fact

The AI compute cost curve bends toward zero faster than any capital plan assumes. This is a trajectory, the documented decline in the price of machine intelligence per token across three consecutive years.

The consensus prices compute as a scarce, appreciating asset. That frame expires in 2026, and the repricing will look sudden to everyone reading capex headlines.

Confidence on the technology: high. Confidence on the exact month of the market repricing: medium. The direction admits one answer, and the slope has been visible for anyone reading the right series.

Where the consensus frame breaks

Analysts read the capital-expenditure headlines and conclude that intelligence grows dearer. They watch the wrong number.

Capex measures the race to build capacity. The metric that predicts the future is cost per token delivered, and that figure falls each quarter. Rising spend and falling unit cost coexist, because volume explodes while price craters.

Ninety percent of analysts have the present right. They read the pace of change wrong.

They extrapolate a supply crunch across a decade, when the observed curve has broken every scarcity forecast to date. The frame error is structural, and it will cost slow readers a full cycle.

The cost curve: three anchor points

Call something a trajectory when three points align. The curve of frontier inference offers them cleanly.

Industry data shows the price to serve a given quality of model output has fallen by roughly an order of magnitude per year since 2023. GPT-4 class performance that commanded premium pricing at launch now sells at a small fraction of that early rate.

The precedent sits in solar photovoltaics. The IEA documented a 90% decline in module cost across 2010 to 2020.

Compute follows the same learning-rate mechanics: cumulative volume drives cost down a predictable slope. Each doubling of installed capacity buys a fixed percentage reduction, and accelerator volume doubles fast.

The ASML tell

Watch the picks-and-shovels layer to time the shift. ASML sits at the base of every advanced AI accelerator produced today.

The company repurchased shares across late July 2026 at weighted average prices between roughly €1,375 and €1,525, according to its buyback disclosure filed on 3 August 2026. A firm buying its own stock at that scale signals conviction that demand for lithography compounds for years.

Read the signal correctly. The supply chain expects volume to keep climbing, and volume climbing is the exact mechanism that drives unit cost down the curve.

The hardware layer thrives while the intelligence it enables commoditizes. Those two facts feel contradictory to the consensus, yet they describe the same engine.

The cliff event

Adoption rarely climbs a smooth ramp. It sits flat, then jumps when price crosses a psychological floor.

Cliff event: frontier-equivalent inference reaches roughly $0.001 per token by the close of 2026. At that price, embedding a capable model into every workflow costs less than the human minute spent reading the output.

Marginal reasoning becomes effectively free. When a resource crosses into abundance, architecture changes.

Teams stop rationing calls to the model and start wrapping entire products around continuous inference. That is a regime change, and it arrives on a date, reflecting mechanics rather than sentiment.

Three categories that dissolve

Three categories will vanish in their current form by 2027:

• Generic "AI capacity" resellers, whose margin evaporates as the underlying token price approaches zero.
• Foundation-model pure-plays that lack a proprietary data moat, squeezed between falling prices and rising training bills.
• Enterprise procurement desks locked into multi-year compute contracts priced at 2025 rates.

The value migrates upward, into vertical applications that own proprietary data. Those firms rent commoditized intelligence and sell an outcome that rivals cannot replicate.

The moat lives in data nobody else holds, in workflow depth, in distribution. Generic tokens buy none of that.

A CTO reading this should audit every contract that assumes compute stays expensive. A growth investor should fund the data moat, and treat generic model access as a line item rather than a thesis.

Who gets hurt, and who compounds

The repricing sorts winners from laggards along one axis: proximity to proprietary data.

A Chief Strategy Officer running a three-year plan built on expensive inference is planning for a world that dissolves mid-plan. Rewrite the assumptions now, while the change stays cheap to absorb.

Procurement teams face the sharpest edge. A multi-year compute contract signed at 2025 rates locks a premium into the balance sheet exactly as the market price collapses beneath it.

Venture capital gets the inverted lesson. The bet that looks expensive today, a vertical firm with a narrow data moat, prices correctly once intelligence turns into a utility.

The position, and what would change my mind

My standing position: foundation models commoditize within 18 to 24 months, and the entire margin migrates to the application layer.

Buying generic capacity today means buying tomorrow's commodity at yesterday's price. I hold this with high conviction, and I keep the falsifier explicit.

Kill signal: a sustained rise in cost per token across two consecutive quarters, driven by a genuine supply ceiling in advanced packaging or lithography. That pattern would break the learning-rate mechanics and validate the scarcity camp. I watch ASML order books for the first sign.

The prediction

Prediction: frontier-equivalent inference will sell at or below $0.001 per token from at least one major provider before 31 December 2026.

Confidence: 78%. Horizon: the end of 2026. The mechanism is cumulative volume driving the learning curve, the same force that took solar down 90% across a decade.

Kill signal: two consecutive quarters of rising per-token cost tied to a hardware supply ceiling. Absent that pattern, the curve holds. Read the full argument, then check the dates against reality on the blog.