Every CFO who signed off on an AI budget this year was told the same thing: token prices are collapsing, so the cost of AI will fall. Then the invoice arrived and it was larger than last year’s. Nothing broke, nobody overspent, and the per-token price really did drop. The bill went up anyway. If that sounds like a mistake, it is not. It is a 160-year-old economic rule doing exactly what it has always done, and it is the single most misunderstood fact in enterprise AI planning today.
That rule is called Jevons Paradox. AI analyst Nate B Jones has returned to it repeatedly over the past year, in episodes on labor markets, compute supply, and enterprise spend, because it explains the gap between what organizations budget for AI and what they actually pay. At Big Hat Group we see the same gap in every engagement: leaders who planned for a falling price curve and are now living on a rising consumption curve. This post introduces the paradox, lays out the numbers, and sets out what to do about it. It pairs with our AI Maturity Model, which covers how organizations evolve; this one covers what that evolution costs.
What Is Jevons Paradox?
In 1865 the English economist William Stanley Jevons studied Britain’s coal consumption and noticed something that contradicted common sense. James Watt’s steam engine was far more efficient than the engines it replaced, so you might expect Britain to burn less coal. Instead, coal consumption exploded. The efficient engine made steam power economical for applications that had never been worth the fuel before: factories, railways, ships, pumps. Efficiency removed a cost barrier, and demand rushed in to fill the space it left behind.
The pattern has repeated with every major resource since. Cloud computing is the closest modern analogy. Compute got dramatically cheaper per unit through the 2010s, and cloud bills went up, because cheap compute made it rational to run workloads nobody would have provisioned a server for. The same thing is now happening with AI inference, and it is happening faster than any resource in history.
Satya Nadella invoked Jevons Paradox by name in early 2025, after DeepSeek’s low-cost models briefly rattled the market. His response was that as AI gets more efficient and accessible, its use will skyrocket. He was right, and the data since then has been emphatic.
Two Curves Running in Opposite Directions
AI economics are governed by two curves moving in opposite directions at the same time. The price curve is falling fast. The consumption curve is rising faster. The second curve is the one your budget needs to track.
The price curve. Per-token inference costs are dropping at rates that make Moore’s Law look gentle: somewhere between 10x and 200x per year depending on which benchmark you measure. The raw commodity is deflating faster than any computing resource ever has.
Per-token price deflation, as compiled by Nate B Jones (Feb 2026)
| Benchmark | 2022 cost | 2026 cost |
|---|---|---|
| GPT-4 equivalent performance | $20 per million tokens | About $0.40 per million tokens |
| Claude 4.5 Sonnet | Not available | $3 per million input tokens, projected to reach cents within one to two years |
The consumption curve. While prices fall, every measure of usage and spend is climbing, and the organizations spending the most are among the fastest-growing companies in the world. They are spending more than their top line because they expect the consumption curve to vindicate the bet.
Enterprise AI consumption and spend, as reported across the Feb 2026 episodes
| Metric | Figure |
|---|---|
| Average organizational AI spend, 2026 | $85,000 per month, up 36% year over year |
| Organizations planning to spend over $100,000 per month | 45%, up from 20% |
| Average enterprise LLM spend, 2025 | $7 million per year, up from $4.5 million in 2023 |
| Average enterprise LLM spend, 2026 projection | More than $11 million per year |
| Tokens processed by Google per month | 1.3 quadrillion, a 130x increase in roughly a year |
| Perplexity’s 2024 spend on cloud AI | 164% of its 2024 revenue |
Google’s number deserves a second look. Google is the most sophisticated operator of AI infrastructure on the planet, and its consumption grew 130-fold in about a year. Its consumption curve is your leading indicator.
What It Looks Like at the Task Level
Abstract curves become concrete when you look at a single knowledge worker who uses AI aggressively. The token counts are much larger than most budget owners assume.
Token consumption for one knowledge worker
| Activity | Token cost |
|---|---|
| Complex analysis, such as reviewing 50 documents and synthesizing findings | About 500,000 tokens |
| A full day of active AI-assisted coding | Millions of tokens |
| A research session with multiple iterations | Up to 10 million tokens |
| Heavy annual use, current baseline | About 1 billion tokens per year |
| Projected ceiling for top users, 18 months out | 25 billion tokens per year or more |
All of that is before agents enter the picture.
The agentic multiplier. A human has physical limits. They type, they take breaks, they sit in meetings, and at absolute peak they might sustain 50 million tokens in a day. An agentic system has none of those limits. It monitors, analyzes, responds, and plans continuously, and a single agent can consume billions of tokens per day. A fleet of agents running in parallel consumes trillions. This is not a step change in consumption. It is a change of several orders of magnitude, and it is the reason the consumption curve will keep outrunning the price curve for years.
The Enterprise Math
Take a 10,000-person organization paying a blended $2 per million tokens. The bill depends entirely on which consumption level you plan for.
Annual compute bill for a 10,000-person organization at $2 per million tokens
| Consumption level | Annual compute bill |
|---|---|
| 1 billion tokens per worker (today’s heavy users) | $20 million |
| 10 billion tokens per worker (18-month projection) | $200 million |
| 100 billion tokens per worker (full agentic scale) | $2 billion |
Traditional planning frameworks make this worse. Capex models, depreciation schedules, and multi-year procurement cycles all assume two things: that demand is predictable, and that supply will be available when you need it. Under Jevons Paradox, neither assumption holds. Demand grows nonlinearly as capability improves, and compute supply is constrained for the rest of the decade.
Cognitive Work Is the New Coal
The paradox applies to labor as cleanly as it applies to compute. AI is collapsing the cost of a specific kind of work: tokenizable cognition. Drafting, summarizing, analyzing, researching, coding, planning, and generating variations are all thinking that can be captured as text, and their marginal cost is approaching zero.
The intuitive prediction is that firms will produce less of this work because they no longer need to pay people to do it. The Jevons reality is that firms are producing vastly more of it. The volume of cognitive output is exploding, not contracting. In practice that looks like:
- More customer segments, each with its own detailed messaging
- More A/B tests, because each variant is nearly free to produce
- More internal reports and analyses, requested because they can be
- More first drafts to choose from before committing to one
- The full software backlog in play, not just the top ten items
Work that used to confer competitive advantage because it was hard to produce is now trivially easy to produce. That creates two pressures at once. The cost floor for cognitive work has dropped to near zero, so producing it is no longer a differentiator. And the real constraint moves upstream, to judgment, accountability, and physical execution: the things AI still cannot do, and the things your organization should now be optimizing for.
Three Drivers Accelerating Consumption
Three forces push the consumption curve upward, and each compounds the others.
Capability unlocks usage. Every time a model gets meaningfully better at something, such as complex reasoning or multi-step planning, it opens up applications that did not work before. Each improvement creates nonlinear new demand. Going from managing one agent to managing ten is a tenfold increase in token footprint, and the better the models get, the more agents it is worth running.
Integration multiplies touch points. AI is no longer a standalone tool you open when you need it. It is embedded in email clients, document editors, development environments, CRM systems, and service desks. Every integration is a portal for ambient, continuous consumption that nobody consciously decided to start.
The agentic shift. The move from human-in-the-loop to agentic systems, where AI calls AI in automated loops, is the largest driver of all. A single agentic workflow can consume more tokens in an hour than a person generates in a month, and enterprises are only beginning to deploy them at scale.
What Enterprises Should Do About It
The paradox is not a reason to slow down. It is a reason to plan correctly. These are the moves that separate the organizations getting value from AI from the ones only getting invoices.
- Stop budgeting AI as a cost-reduction tool. The premise is wrong. Jevons Paradox guarantees that efficiency gains will be absorbed by expanded usage, not banked as savings. Budget AI as a capability-expansion investment with its own return, not as a headcount-reduction play that will somehow make the line item shrink.
- Build token management as an organizational capability. The scarce skill is no longer model access; everyone has that. The scarce skill is knowing how to aim tokens: structuring the context each model receives so it does the job with less, routing each task to the right model at the right cost, designing agent loops that sustain quality over time instead of degrading, and measuring whether purchased intelligence is producing the outcomes you need. Whether you call it token management, intelligence operations, or context engineering does not matter yet. The organizations building this capability are pulling away from everyone else.
- Build an internal routing layer. Best-in-class enterprises run an internal platform that sends work to the right model at the right price point: cheap models for well-defined tasks, frontier models for ambiguous or high-stakes work. They negotiate custom API agreements, commit to consumption floors in exchange for dedicated capacity, and keep multi-vendor optionality so no single provider’s pricing dictates their roadmap. Our AI automation practice builds exactly this layer.
- Treat token efficiency as a competitive advantage, not cost control. Your competitors have access to the same models you do. The organization that completes the same task with half the tokens has twice the effective capacity. The right question is not “how do I make my agents cheaper?” but “am I spending on the right things?” More capable agents doing more work at lower per-token prices will still produce a larger total bill. The leverage is in redesigning workflows, not just tightening prompts.
- Secure compute capacity before supply tightens further. DRAM and high-bandwidth memory supply is structurally constrained through 2028. The hyperscalers are competitors as well as vendors, and in a constrained market they allocate capacity to their own products first. Enterprises that lock in contractual capacity commitments now, whether through reserved Azure capacity or provider-level agreements, will be in a far better position than those who wait. Our Azure consulting team helps clients model and reserve the capacity their consumption curve actually requires.
- Get a plan. Big Hat Group helps enterprises model their real consumption curve, stand up token management and routing, and design the agentic operations that turn AI spend into measurable capability instead of an unexplained invoice.
What This Means for Enterprise IT and AI Leaders
Your rising AI bill is a leading indicator, not a failure. If spend is climbing while per-token prices fall, your organization is finding new uses for intelligence faster than the vendors can discount it. That is what adoption looks like. The failure mode is a flat bill, which usually means AI is stuck in pilots nobody uses.
The bottleneck has moved from access to aim. Every organization can buy frontier intelligence by the token. Very few can direct it well. Context design, model routing, agent-loop engineering, and outcome measurement are now core IT competencies, in the same way that cloud cost management became one a decade ago. If nobody owns them, you will pay the Jevons bill without getting the Jevons benefit.
Capacity and vendor optionality are architecture decisions. In a supply-constrained market, where your tokens come from and on what terms matters as much as which model you use. Multi-vendor routing, reserved capacity, and a model-agnostic abstraction layer are no longer nice-to-haves. They are how you keep a rising consumption curve from becoming a procurement crisis.
The gap compounds. An organization that treats AI as a cost line to be minimized does not just fall behind a competitor that treats it as capacity to be deployed. The gap widens every quarter, because the second organization is running more experiments, more agents, and more cognitive output on the same falling price curve. Under Jevons Paradox, the winner is not the one who spends the least. It is the one who turns the most tokens into outcomes.
Get Enterprise AI Consulting
Big Hat Group works with enterprise IT and AI leaders to plan for the consumption curve rather than the price curve: modeling real token demand, building the routing and context layers that make every token count, securing the capacity to run agents at scale, and governing agentic operations so the bill maps to business results.
Contact us to plan your AI token budget →
Related: AI Maturity Model: From Model-Chasing to Future-Ready · Enterprise AI Strategy Beyond Microsoft Copilot · lean-ctx: The Context Layer AI Agents Actually Need
This post draws on four episodes of Nate B Jones’s AI News & Strategy Daily: The 3-Layer Framework That Predicts Which Jobs AI Will Replace (Jan 11, 2026), Why the Smartest AI Teams Are Panic-Buying Compute (Feb 8, 2026), $1,000 a Day in AI Costs (Feb 20, 2026), and You can be ambitious without the huge token bill (Sep 20, 2026). Spend and consumption figures are as reported in those episodes.