For most of the generative-AI era, enterprise strategy has been organised around scarcity. The best models were expensive, access was constrained, compute was difficult to secure and only a handful of companies could afford to build frontier systems. That assumption is beginning to break. The emerging market for AI inference now looks less like a luxury market and increasingly like a commodity market — more suppliers, faster turnover, relentless price compression and far less loyalty to any single model.

The most striking evidence comes from a 2026 paper in the Journal of Economic Perspectives by Mert Demirer, Andrey Fradkin and Nadav Tadelis. Using OpenRouter market data, the authors document a remarkable transition: the market price of intelligence has fallen roughly a thousandfold, while open-source models now cost about 90 percent less than comparable closed-source alternatives. At the same time, the identity of the “best” model changes frequently and no single provider dominates every use case.

Stanford’s AI Index had already shown the direction of travel. Between November 2022 and October 2024, the cost of querying a model with GPT‑3.5-level benchmark performance fell from roughly $20 per million tokens to $0.07 — a decline of more than 280 times in less than two years. Depending on the task, Stanford observed annual inference-price declines ranging from 9× to 900×.

AI Did Not Get Slightly Cheaper. Its Economics Were Rewritten.

This matters because price compression changes behaviour. When a high-quality model call costs dollars, organisations ration it. When the same class of capability costs cents — and eventually fractions of cents — entirely new categories of automation become economically viable. Documents that were once sampled can be analysed continuously. Customer interactions can be evaluated in real time. Software repositories can be reviewed after every change. AI can move from a tool a human occasionally invokes to infrastructure that runs continuously in the background.

Why the Price of Intelligence Keeps Falling

There is no single cause. The collapse is the combined effect of smaller capable models, better accelerators, improved serving software, quantisation, batching, caching, competition among inference providers, rapidly improving open-weight systems and a market that can route workloads across multiple suppliers rather than accepting one vendor’s price.

The Stanford AI Index captured one particularly important change: the smallest model able to cross a 60 percent MMLU threshold shrank from Google’s 540-billion-parameter PaLM in 2022 to Microsoft’s 3.8-billion-parameter Phi‑3-mini in 2024 — a 142-fold reduction in model size for that benchmark threshold. The same report found the gap between leading open-weight and closed systems narrowing to just 1.7 percentage points on some benchmarks.

That does not mean a small open model is equivalent to the strongest frontier model on every task. It means an enterprise no longer needs frontier capability for every task — and increasingly has no economic reason to pay frontier prices when it does not.

If the Model Stops Being the Moat, Value Has to Move Somewhere Else

The early generative-AI market encouraged a simple assumption: the company with the smartest model would capture most of the value. Commodity economics challenge that assumption. If high-quality intelligence can be purchased from dozens of suppliers, if model leadership rotates rapidly and if switching becomes easier through common APIs and routing layers, then the model itself begins to resemble a replaceable input.

The moats that weaken — and the ones that strengthen

The One-Model Enterprise Is Starting to Look Economically Irrational

A procurement model built around choosing one strategic AI provider made sense when switching costs were high and performance differences were enormous. It makes less sense when models specialise, prices diverge by an order of magnitude and the best provider for coding may not be the best provider for extraction, reasoning, customer service or low-latency classification.

The natural response is model routing: applications decide which model should handle each request based on cost, latency, capability, privacy requirements and risk. The model becomes an execution layer behind an orchestration system rather than a brand the employee consciously chooses.

This changes vendor power. A company that can route around a price increase has bargaining leverage. A company whose entire AI architecture is deeply tied to one proprietary model does not.

Five Assets That Become More Important as Intelligence Becomes Abundant

Asset	Why scarcity increases	Strategic value
Proprietary data	Generic models are shared. High-quality organisational context is not.	Very high
Distribution	When model supply explodes, owning the customer relationship becomes more valuable.	Very high
Workflow integration	Useful AI must act inside real systems, not merely generate text.	Very high
Evaluation & governance	More model choice makes reliability, security and benchmarking harder.	Rising
Compute efficiency	At massive scale, small cost differences compound into material economics.	Rising

Data is the most obvious answer, but not simply because “data is the new oil.” Generic data has also become abundant. The valuable asset is context that competitors cannot easily acquire: transaction histories, operational telemetry, expert decisions, customer outcomes, failure modes, domain-specific documents and the organisational semantics required to interpret them.

This is why companies with large proprietary datasets increasingly talk about open-weight models as a strategic opportunity rather than a compromise. A cheaper model adapted to unique internal data can outperform a more expensive generic system on the task that actually matters to the business.

What Happens to the Frontier Labs If Their Core Product Deflates?

Commodity dynamics do not imply that OpenAI, Anthropic, Google, Meta or leading Chinese labs disappear. Commodity markets can still support enormous businesses. Cloud compute, semiconductors, storage and telecom bandwidth all exhibit intense price pressure while sustaining valuable suppliers.

But the basis of competition changes. Frontier labs increasingly need more than raw intelligence. They need enterprise distribution, integrated developer ecosystems, agent platforms, proprietary tools, trusted security controls, hardware advantages, consumer networks or vertical applications that give customers reasons not to treat the underlying model as interchangeable.

The frontier itself may also remain scarce. State-of-the-art reasoning, very long context, multimodal capability, advanced coding and autonomous-agent performance can command a premium even while yesterday’s frontier becomes tomorrow’s commodity. In that sense, the market may split into two layers: premium intelligence at the frontier and commodity intelligence everywhere else.

Six Enterprise Moves for the Commodity-Intelligence Era