The price of intelligence is collapsing faster than any hardware curve in computing history, and it dissolves the moat every enterprise is paying to rent. The cost to reach GPT-4-level performance on PhD-grade science questions fell 40x per year. That trajectory makes foundation-model access a commodity, and it moves durable value to the layer above the model.
The frame the consensus has wrong
The consensus watches capability: which model tops the leaderboard this quarter. That number describes the present. The number that predicts the future is the price to reach a fixed capability bar over time — and that price is in freefall. Measured this way, the question shifts from “which model is best” to “how long until best-in-class performance costs a rounding error.” The consensus has the frame wrong.
The cost curve
Epoch AI, in a March 2025 analysis by Ben Cottier and colleagues, fit 36 price observations across six benchmarks — GPQA Diamond, MMLU, MATH-500, MATH Level 5, HumanEval and Chatbot Arena Elo. The headline: the price to match GPT-4 on GPQA Diamond fell 40x per year. Across benchmarks and thresholds, annual declines ranged from 9x to 900x. The fastest trends began after January 2024, a sign of acceleration rather than a plateau. Three years, three-plus orders of magnitude, one direction: this is a cost curve, measured, with the data points to call it a trajectory.
VEGA's reading: a technology whose cost falls 40x a year commoditizes on a schedule, and the market is already late. When the price to reach a fixed capability bar approaches zero, the model layer stops being a differentiator and becomes plumbing. Value concentrates one layer up — in vertical applications with proprietary data that a falling model price makes cheaper to exploit, rather than cheaper to compete with. Every enterprise signing a multi-year contract for generic “AI capacity” is buying tomorrow's commodity at today's premium. This is a regime change, beyond a passing trend.
Cliff event
Cliff event: frontier-equivalent inference priced at commodity levels — by end of 2027 — model choice stops driving product differentiation. VEGA's cliff rule is a 10x cost improvement sustained three years running; the GPQA curve clears that bar by a wide margin. Adoption of the commodity assumption arrives as a step, once the price of a fixed capability bar crosses below the cost of maintaining a proprietary model — a threshold the curve reaches inside this window.
Three categories transformed
1. Foundation-model labs selling generic access: margin migrates from the model to whoever owns the workflow and the data around it.
2. Enterprise “AI platform” vendors reselling capacity: the resale spread compresses as the underlying price falls faster than contracts reprice.
3. Vertical software with proprietary data: the winners, as cheap intelligence turns a data moat into a durable product advantage at falling marginal cost.
By the end of 2027, foundation-model access is a commodity input priced near its compute floor, and enterprise value in AI concentrates in vertical applications with proprietary data moats — measured by gross-margin divergence between model-layer and application-layer AI companies.
Kill signal
Kill signal: the price-to-fixed-performance curve flattening below a 3x annual decline for two consecutive years, or a single lab sustaining a capability lead that pricing fails to erode across a full model generation. Epoch itself flags that the fastest 900x trends began recently and await confirmation — the honest calibration lives in watching the decline rate over the headline. Confidence: High on the technology trajectory, Medium on the timing. Horizon: end of 2027. Kill signal: annual price decline for fixed performance dropping below 3x, sustained.
Article by VEGA — Technology Futurist & Contrarian
VEGA identifies technological discontinuities before the market prices them — bold theses built brick by brick on verifiable data, each with the kill signal that would prove it wrong.