Key takeaways
- Inference, not training, is becoming the dominant economic bottleneck for most enterprise AI deployments once usage scales.
- Today's pricing environment reflects strategic competition as much as sustainable economics. Many buyers are validating use cases on pricing that was set to acquire market share, not to reflect steady-state infrastructure economics. This is a material planning risk.
- The Jevons paradox is not a theoretical concern in AI — it is the observed pattern. Per-token costs have fallen dramatically over the past two years. Enterprise AI budgets have not fallen. They have risen, in many cases substantially, as cheaper access triggered more use cases, wider access, and deeper agentic integration.
- Falling per-token costs do not guarantee lower budgets. The correct financial question is not "what does a token cost?" but "what does a business outcome cost, and what happens to that number as the system scales?" Most enterprises cannot answer this question.
- Enterprise buyers who have not stress-tested their AI economics against 2-3x higher inference costs, or against demand volumes 5-10x current levels, are operating with a single-scenario plan for a multi-scenario problem.
The real bottleneck is moving
For the last two years, much of the AI cost conversation has been organised around training. That is understandable. Training large models is capital-intensive, technically demanding, and highly visible. But for most enterprise buyers, training is not the economic bottleneck they will live with. Inference is.
Inference is where the recurring cost sits. It is what the organisation pays each time a customer query is handled, a workflow is assisted, a document is summarised, or an agentic system takes the next step in a chain of actions. Training may define the frontier economics of model creation. Inference defines the day-to-day economics of enterprise use.
That distinction matters because enterprises often buy and approve AI capability using the easiest commercial surface available today. If the current pricing environment is unusually aggressive, then many organisations are testing demand on economics that may not last.
Why today's prices may mislead buyers
One of the less comfortable realities in the current market is that pricing is still shaped by strategic competition as much as by steady-state economics. The market is trying to establish dominance, scale, and developer dependence at the same time that it is trying to monetise expensive infrastructure. That is not a stable condition.
Public market commentary and infrastructure economics point in the same direction. OpenAI's economics have been widely discussed as heavily investment-intensive. Model providers are racing to secure massive compute capacity. Nvidia, meanwhile, continues to benefit from a market that still prices accelerated compute as a scarce strategic resource. The result is a value chain where buyers enjoy impressive capability today, but should not assume today's inference prices reflect mature market equilibrium.
That does not mean prices can only rise. Technical improvements are real. Nvidia has claimed large future inference efficiency gains through next-generation systems. Sparse-attention and related architectural improvements promise meaningful per-token compute reductions on long contexts. These advances matter. But they do not erase the portfolio risk for buyers. Lower unit cost can coexist with much higher total spend if demand expands faster than efficiency improves.
The Jevons problem in enterprise AI
This is where the Jevons paradox becomes strategically important. When a resource becomes cheaper or easier to use, total consumption can rise rather than fall. AI is an unusually fertile environment for that effect.
If model usage becomes cheaper, organisations rarely respond by keeping demand fixed and enjoying the savings. They add new use cases, widen internal access, deepen context windows, increase automation steps, and experiment with more ambitious workflows. Agentic systems intensify this pattern because a single business action may require multiple model calls, tool invocations, and validation passes. If a simple chatbot interaction already creates meaningful cost, an agentic workflow can create several times more.
That means enterprise buyers should not ask only whether model pricing is falling. They should ask what happens to total spend if their successful use cases become much more heavily used, more deeply embedded, and more autonomous.
Why agentic systems change the economics
The move from chat assistance to agentic execution changes the economic denominator. A human asking one question is one thing. A workflow that orchestrates planning, retrieval, reasoning, tool use, verification, and follow-up actions is another.
This has several consequences. First, the cost unit needs to move from token and prompt thinking toward cost per action and cost per outcome. Second, system design matters more because inefficient branching, overlong context, and indiscriminate premium-model routing become multiplicative rather than marginal. Third, observability becomes a financial control issue, not just an engineering one.
The primary cost problem is therefore no longer "what does one request cost?" It is "what does one useful business action cost once the whole chain of decisions is visible?"
What enterprise buyers often miss
Enterprise buying teams often miss four things.
The first is that the model invoice is only one layer of recurring inference economics. Retrieval, guardrails, observability, fallback calls, support, and human oversight can all expand with usage.
The second is that a low-cost pilot can create a high-cost production pattern if its workflow design is inefficient.
The third is that pricing normalisation can turn today's acceptable unit economics into tomorrow's portfolio issue if the use case has already become business-critical.
The fourth is that the right comparison is rarely between provider rate cards alone. It is between alternative workflow designs, model-routing strategies, and deployment approaches under multiple demand scenarios.
How buyers should prepare now
There are five practical moves buyers should make before the market forces them to.
1. Measure cost at the workflow level
Cost per token is useful, but insufficient. Buyers need cost per inference, cost per action, and eventually cost per outcome if they want the economics to remain decision-useful.
2. Design for routing and caching from the start
Premium models should not become the default for every step of every workflow. Simpler tasks often deserve smaller models, cached outputs, or batched processing.
3. Stress-test the business case under less favourable pricing
If provider prices rose materially or if agentic depth increased, would the use case still be worth scaling? A resilient case should survive more than the most favourable current scenario.
4. Separate shared capability from local demand
Foundational platform and observability investment may be rational. But it should be distinguished from the economics of one workflow so buyers do not misread success or failure.
5. Connect inference economics to portfolio governance
Inference cost is not only an engineering concern. It should inform funding, sequencing, and stop or scale decisions at portfolio level.
What this means for different buyers
For CFOs, the inference cost crisis is a visibility and timing problem. Use cases may be approved under one economic assumption and scaled under another.
For CIOs and Heads of Engineering, it is an architecture problem. Routing, context policy, and system design can determine whether a useful AI capability becomes economically resilient or brittle.
For FinOps leaders, it is a live demand-governance problem. The goal is not just to report spend. It is to influence how the workload is behaving before the pattern hardens.
For portfolio and strategy leaders, it is a sequencing problem. Expensive recurring demand should not crowd out more defensible or strategically useful AI bets without being compared explicitly.
Stress-testing your own case: three scenarios to run now
Before treating any AI investment as economically validated, organisations should run three scenarios that most current AI economics analyses do not include.
Scenario one: provider pricing normalises upward by 50-100%. This is not an extreme scenario. It is a return toward economics that cover infrastructure costs without strategic subsidy. If the use case is not viable at twice current inference prices, the business case is priced on market structure rather than on fundamental value. That market structure may or may not persist.
Scenario two: demand doubles, then doubles again. When an AI capability works and demonstrates value, the organisation's natural response is to expand it — more users, more use cases, deeper integration. If a workflow was economically marginal at pilot scale, it may be economically unsustainable at production scale. Run the unit economics forward at 2x and 4x current demand before assuming the pilot economics represent a stable baseline.
Scenario three: the capability becomes operationally embedded before pricing changes. This is the most dangerous scenario for enterprise buyers. A capability is deployed, adopted, and integrated into core workflows. Dependencies form. Switching costs accumulate. At that point, a provider pricing revision — or simply a reversion to sustainability economics — finds a buyer with limited ability to respond. The governance question is not just whether the economics work today, but whether the organisation has enough architectural and commercial flexibility to respond if they change.
If the business case does not survive all three scenarios at acceptable return levels, that is not necessarily a reason to abandon the use case. It is a reason to either redesign the workflow to be more efficient, build in more commercial flexibility, or be explicit that the current justification depends on favourable conditions that may not persist.
The practical conclusion
The inference cost challenge is not a prediction of collapse. It is a warning about where enterprise AI economics will become most contested as adoption deepens and pricing matures. The organisations most exposed are not those using AI aggressively — it is those using AI aggressively on a single pricing and demand scenario without stress-testing the economics.
The Jevons paradox is already playing out across enterprise AI portfolios. Total spend is rising even as unit costs fall, because demand is rising faster than efficiency is improving. At some point, rising total spend encounters an ROI scrutiny test. The organisations that pass that test will be the ones that measured workflow economics carefully, designed for efficiency from the start, and governed AI use as a portfolio of recurring operating commitments. The ones that fail it will be the ones that confused a favourable pricing environment with a validated business model.