Why Cheap Models Struggle to Make Money: Enterprise AI Procurement Pricing

    Enterprise AI spend is shifting toward more expensive frontier models even as inference prices fall. This article explains the data, why intelligence is priced more like talent than electricity, and how task success rate—not token unit price—drives total cost.

    A Counterintuitive Industry Pattern

    Over the past year, the most visible change in the large-model market has been price cuts. Inference prices have kept falling, and the average cost per million tokens has dropped sharply; in some scenarios it is almost negligible. Classic supply-and-demand logic would suggest that lower prices should expand demand, and cheap models should capture the profits.

    Surveys of enterprise decision-makers point the other way. Budgets are accelerating toward more expensive flagship closed-source models. Spend on frontier models has risen from roughly 30% to more than 60%, and their share of token consumption has nearly doubled. Small closed-source models, even after price cuts of more than 80%, have seen their token share shrink. Open-source models may grow in call volume, but extremely low prices mean that high usage still produces limited revenue.

    “Selling more” and “earning more” have diverged. That gap is the subject of this article.

    Two Data Points That Cross-Check Each Other

    Open-source model revenue structure

    One class of surveys tracks what enterprises spend on open-source models. Most of that spend goes to a handful of leading open-source providers; well-known Western open-source models take only a small share. The important caveat is the metric: it is open-source model spend, not all enterprise AI spend. Open-source models are a modest slice of the overall AI budget, and unit prices are extremely low. Winning inside the open-source lane still leaves them a supporting role in industry-wide revenue.

    A simple illustration:

    1. Assume a company has a monthly AI budget of $1,000,000:
    2. - Flagship closed-source models: $700,000 (high unit price, concentrated usage)
    3. - Small/mid closed-source models: $150,000
    4. - Open-source / self-hosted models: $150,000 (very low unit price)
    5. A single model holds 60% of the open-source lane
    6. → Actual revenue = $150,000 × 60% = $90,000
    7. → Only 9% of the company’s total budget

    “61% of the open-source lane” is not the same as “61% of industry revenue.” Much of the public confusion comes from mixing those two definitions.

    Prices fell, share fell

    Small closed-source models are even more telling: prices dropped more than 80%, yet token share fell from nearly 30% to a little over 10%. Discounting did not buy share. That rarely happens for ordinary goods unless the buyer’s marginal value of the product is itself declining.

    Intelligence Is Not Electricity—It Is Talent

    Why the “intelligence as electricity” story fails

    A popular claim is that intelligence will become a metered public good like electricity, and the cheapest supplier will win. The analogy sounds clean; enterprise buying behavior does not support it.

    If intelligence were electricity, procurement lists would collapse onto a “cheap line”: every workload would move to the lowest-cost compute, and price would be the only dimension of competition. In practice, willingness to pay a premium for stronger models is rising, and selection is shifting from value tiers to flagship tiers.

    What enterprises are actually buying

    A better analogy is the labor market. When companies buy models, they are not buying a standardized fuel called “tokens.” They are buying the ability behind those tokens—whether a complex job can be completed end to end:

    • A long agent chain that fails halfway wastes all prior spend.
    • A failed code test costs more in retries and human time than the saved API fee.
    • A legal or compliance search that misses a critical clause can wipe out a year of “savings.”

    The unit of procurement is not the token; it is task success rate. Strong models are senior specialists; weaker models are interns. An intern’s hourly rate can be very low, but if core work always comes back for rework, the intern still does not get the critical projects. That is why flagship models can get more expensive and still gain share: they sell “getting it right the first time,” not cheapness.

    A failure-cost model

    A schematic cost model:

    1. Total cost per task = model call cost + rework cost + opportunity cost
    2. Model A (cheap):
    3. Call cost 0.1, success rate 70%
    4. Expected cost = 0.1 + 0.3 × (rework cost R)
    5. Model B (expensive):
    6. Call cost 1.0, success rate 95%
    7. Expected cost = 1.0 + 0.05 × R
    8. When R > 3.6, Model B has the lower expected total cost

    The more critical the task and the longer the chain, the larger R becomes, and the easier it is to cross the threshold where the expensive model is cheaper overall. That is why price cuts do not rescue weaker models: failure rates keep their expected total cost high.

    Three Implications for Enterprises and Developers

    1. Cost by task, not by token. When selecting models, do not stop at price per million tokens. Estimate end-to-end task cost: retries, human review, and rework. In evaluation, set a baseline per business task, record first-pass success and average retries, and decide on expected total cost.

    2. Layered routing is the real cost optimization. Not every call needs a flagship model. Grade by risk: draft generation, format conversion, and simple summarization can go to cheaper models; decision-making, reasoning, and code generation should go to stronger models with automated checks. Routing can be rules or a light classifier—modest engineering, significant savings.

    3. Capability-per-dollar still matters in open-source and cloud toolchains. Survey data indicate that Chinese models already account for a leading share of global open-source model spend, which suggests competitiveness on capability versus price—including cloud-hosted model APIs and supporting routing, tuning, and evaluation tooling. Include those options in the candidate set and decide from real workload evaluations, not parameter sheets.

    Conclusion

    Cheap models fail to make money not because the price war “lost,” but because intelligence is not priced as unit price × volume. It is priced as capability × task value. As the core demand shifts from “it can run” to “it runs correctly and to completion,” paying for certainty outweighs paying for cheapness. For vendors, the moat is task success rate. For users, real savings come from matching each class of task to the right capability tier: spend where it must be spent, and save the rest with architecture.