Cheaper Parts, Faster Engines

Token prices are collapsing. AI bills are exploding. Both are true, and neither is the number that matters.

Two charts are circulating through boardrooms right now. One shows the price of a token down 90 percent. The other shows enterprise AI bills tripling. Board members forward the first. CFOs live inside the second. Each side thinks the other can’t read a chart.

Underneath the chart war sits a sharper question, usually pointed at anyone selling AI capacity into a labor budget: if the input costs are collapsing, why isn’t the price?

The first question is easy. The second one is the point of this essay.

Both sides are right

Two weeks ago the debate got its first real dataset. Exponential View published its State of the AI Economy on June 25, a bottom-up, deduplicated reconstruction of the generative AI market built on company-level financial models and Epoch AI capability data. It is the best measurement of AI demand anyone has produced, and it ends the argument by splitting it in two.

Token prices collapsed. The blended price of a million tokens fell from about $17 to $2. Ramp’s enterprise spend data shows the realized version of the same move: average cost per million tokens across major providers down from roughly $10 to $2.50 in a year. And the next leg is already scheduled. Capital is 89 percent of the cost of running AI infrastructure, so deflation from here is a hardware story, and Nvidia’s Rubin platform targets a 10x cut in inference cost versus Blackwell.

Consumption grew faster. Tokens processed per output token tripled, from 12 to 36, because reasoning models spend tokens thinking and agents spend tokens planning, calling tools, and checking their own work. A chat turn was one model call. An agentic task is ten or twenty. Global volume passed 30 quadrillion tokens a month, up 14x in a year. Ramp’s June index puts numbers on the enterprise side: token consumption among the heaviest adopters up 1,001 percent year over year, total AI spend across its 70,000 businesses up 497 percent, with median month-over-month swings of 58 percent against SaaS variance of roughly zero. Uber’s CTO confirmed the company burned its entire 2026 AI budget in four months as Claude Code adoption jumped from 32 to 84 percent of its 5,000 engineers.

The mechanism connecting the two is elasticity. Exponential View measures it at 1.2 to 1.8: a 10 percent price cut produces 12 to 18 percent more consumption. It’s a time series, price and volume trend together, so hold the precision loosely. The direction is not in doubt. Google cut prices 97 percent and volume rose 50x. Every price cut raises total spend.

Unit price falls, units per task rise, the bill grows. Gartner’s May forecast has worldwide AI spending hitting $2.59 trillion in 2026, up 47 percent, in the same year it predicts inference for a given model gets 90 percent cheaper by 2030. The FinOps Foundation found 73 percent of enterprises exceeded their AI cost projections. Everyone’s chart is correct.

None of this is new physics. Jevons watched it with coal in 1865: better steam engines meant more coal burned, not less. Nordhaus ran the numbers on artificial light and found it got 99.97 percent cheaper per lumen-hour between 1800 and 1992, and instead of pocketing the savings the world lit itself. When the price of a general-purpose input collapses, total spending on it goes up. Intelligence is following the script at a faster clock speed.

A billing metric, not a unit of value

The report’s most useful claim isn’t a number. The token, it argues, is AI’s billing metric, not its unit of value.

Every general-purpose technology starts here. Edison billed his first customers per lamp installed. Metering came later. The early web sold pageviews, and the ad market stayed small and stupid until pay-per-click tied spend to outcomes. Mobile sold megabytes before anyone understood that a daily active user was the thing to count. Industries bill in the unit they can meter, not the unit the customer values, and markets stay confused until someone closes the gap.

A CFO can price a kilowatt-hour. A CFO can price an hour of labor. No CFO can price a quadrillion tokens. Nothing pegs to it. No benchmark, no budget line, no way to bundle it into value.

The market has already voted on this. On June 4, Ramp raised at a $44 billion valuation on the thesis that AI tokens are a third pillar of business spend, after payroll and procurement, that finance teams cannot track, allocate, or control. It shipped token spend management as a product in April. Companies in its dataset now run a median of nine AI models each, an average of 16.5, and nobody has added up the invoices. An entire FinOps industry is being capitalized to translate the billing unit into something a CFO can govern. A good unit does not need a translation industry.

Exponential View proposes quality-adjusted output tokens instead: strip out input and reasoning tokens as production costs, count user-facing output, weight it by model capability. Better. It also proves the point: the people who measure this for a living agree the raw token tells you almost nothing about value. Any diligence question that starts from token prices has already picked the wrong unit.

The floor under the deflation

Today’s prices also sit on a capital structure that hasn’t been tested.

The report tallies $2 trillion of cumulative hyperscaler and neocloud capex through 2026, with roughly $850 billion landing this year alone and about $160 billion of it financed by new debt. The 2026 depreciation charge approaches $111 billion. Q4 2025 was the first quarter AI revenue cleared quarterly depreciation, and Q1 2026 headroom at the infrastructure layer is 19 percent, before power, staff, and everything else.

The unit economics put a floor under prices. A gigawatt of AI capacity costs $7.9 billion a year to own and operate. Divide by token output and an open-weight token needs 25 to 50 cents per million to earn a real gross margin. A closed frontier token with licensing needs $1.05 to $2.10. The blended market price today is about $2. The market is near the level the balance sheets can support, not far above it. And demand is pressing on that floor from below: inference is now roughly two-thirds of global AI compute demand, up from a third in 2023, and compute prices have already bounced once. H100 rentals went from $3.05 at launch scarcity to $1.70 at peak overbuild fear, back to $2.40 on inference demand.

Frontier prices are set by private companies burning venture money and debt for market share. Several analysts argue the labs price inference below cost, a floor that normalizes upward when capital discipline arrives. Once these companies are public, the income statement votes. A plan that requires permanent frontier deflation is a bet that discipline never shows up.

Trailing intelligence deflates toward zero. Frontier intelligence holds. Open weights commoditize last year’s frontier within months, and on OpenRouter the top three labs’ token share fell from 72 to 33 percent. But demand keeps re-anchoring to the current frontier, because frontier agents do work trailing models can’t. Anthropic reportedly grew from $4.8 billion in Q1 revenue to a projected $10.9 billion in Q2. Nobody buys last year’s level. Gartner said it flatly this spring, warning product leaders not to confuse the deflation of commodity tokens with the democratization of frontier reasoning.

Where margin goes when an input collapses

Back to the diligence question. If the cost of intelligence is collapsing, why doesn’t the price of AI work collapse with it?

It does, for anyone reselling the input. A markup on API calls falls with the API price, and if it doesn’t, competition fixes it shortly. For token resellers the question answers itself.

For everything else, a century of economic history gives the answer: when an input’s price collapses, value migrates to its complements. Cheap steam built fortunes in railways and textiles, not in engines. Cheap electricity is the cleaner case. Paul David studied the dynamo and found that factory productivity barely moved for decades after electric power got cheap, because the gains required rebuilding the factory around the motor: individual machines with their own drives, new floor layouts, new workflows. The fortune went to whoever reorganized production, not to whoever sold the current.

Tokens are the electrons. The labs sell parts. The margin work is the engine: selecting the parts, assembling them into a system, dropping that system into an organization’s actual operations, and compounding what it learns. The CRM nobody documented. The approval flow that lives in one person’s head. The compliance regime, the entitlements, the exceptions that make up half of real work. The memory that makes month twelve better than month one. Watching, correcting, proving. None of that is in the token price. It never was.

And the engine gets rebuilt every season. No F1 team runs last year’s car because engines got cheaper. New parts ship, better models get swapped in, the system gets re-engineered around them, and the car gets faster. Stop and it falls behind in one cycle. In a market moving this fast, perpetual rebuild is both the cost of the business and the moat around it.

Run the token debate through this structure and it stops mattering, in both directions.

Tokens deflate: cheaper parts make engines faster at the same price. Buyers get more work per unit of capacity every year without renegotiating anything.

Frontier tokens hold or rise: the optimization becomes the margin. Frontier models for the hard reasoning, lighter models for the volume, the right part for each job, evolving as the parts do. One analysis of 2.4 billion enterprise API calls found firms routing across a model portfolio paid $2.31 blended per million tokens while firms defaulting everything to frontier paid $18.40. Ramp’s June data shows the same spread a layer down: 60 percent of businesses on its token product run cache hit rates above 80 percent, 13 percent sit below 20. Same list price, radically different effective cost, and the difference lives in technical decisions finance never sees. An 87 percent spread from architecture alone is not a rounding error. It is where the industry’s margin actually lives.

The input moves. The output holds. That is the definition of selling an outcome instead of an ingredient.

The right unit

If the token is the wrong unit, the right one is the one every buyer already prices: a unit of completed work. Labor is the oldest unit of account for output. An FTE of work is legible, it maps to a budget line that already exists, and it lets a buyer evaluate the trade in one move. Nobody has to learn what a quadrillion is.

The industry is moving the other way. GitHub moved Copilot from flat subscriptions to metered credits on June 1, and one developer’s projected bill went from 67 euros a month to 966. Provider by provider, the meter is being pushed onto the buyer, which is how 73 percent of enterprises ended up over budget. The older allocation is better: price the output and keep consumption risk with the seller, because the seller is the only party positioned to reduce it. Every routing improvement, every cache, every model swap accrues to whoever owns the meter, and the meter should belong to the party that can move it. Priced that way, the buyer funds a return on a compounding asset, not a rent on a metered input.

Generality against specificity

The threat to the engine builders is in the same report. The labs are shipping vertical products in law and code, and in one of Exponential View’s scenarios the frontier labs absorb the integration work and generic wrapper pricing power collapses.

Thin shells over an API die in that scenario, and they should.

Models will keep improving. Everything above assumes it. The constraint was never model capability. It is organizational legibility. Organizations are illegible: undocumented processes, tribal knowledge, entitlements, exceptions, regulators, a decade of workarounds nobody wrote down. No company’s actual operations exist in a document a model can read. Making work legible to machines, inside real systems, under a real compliance regime, is specific work. Labs win by being general. Engine work is valuable because it is specific, and it compounds because what one rebuild teaches transfers to the next. Someone has to survey the territory before the map is worth anything.

The freshest data says this is exactly where the market sits. SemiAnalysis went customer by customer at the end of June and found the spend distribution wildly bimodal: 99th percentile firms spend near $90,000 per employee per year on AI, the median Fortune 500 under $100. Tokens are effectively free at the median and the work still isn’t getting done. That gap is not a pricing problem. It is a deployment problem, and deployment is the engine.

When the next chart crosses your desk: both sides are right, and neither is the number to watch. Watch cost per unit of completed work, quality-adjusted. On the report’s own version of that measure, useful output grew roughly 30x in fifteen months while every capability index climbed alongside it. The intelligence actually delivered is compounding faster than any price series, in either direction.

Cheaper parts make engines faster. Pricier parts get optimized around. The rebuild happens every year regardless.

The labs sell parts. Someone has to build the engine. It gets faster every year.

Sources: Exponential View, The State of the AI Economy (June 25, 2026); Ramp AI Index and token cost analysis (June 2026); SemiAnalysis, TokenBudgeting (June 2026); Gartner AI spending forecast (May 2026) and inference cost prediction (March 2026); The Information and WSJ reporting on lab revenues and Uber (May 2026); Optimum Partners, enterprise API cost analysis (May 2026); FinOps Foundation, State of FinOps 2026; Epoch AI price dataset; Paul David, “The Dynamo and the Computer” (1990).