The old edge is expiring
For a decade the winning move in compute was to buy more. More GPUs, more data centers, more capacity than the competition could afford. It worked because capacity was scarce and access was the bottleneck.
Both halves of that have broken. Raw inference prices have fallen roughly tenfold a year, with GPT-4-class capability dropping from around 20 dollars per million tokens in late 2022 to under one dollar now.¹² When the price of the thing you were hoarding collapses by an order of magnitude annually, owning more of it is not a moat. And the easy declines are running out: the cheap price cuts have largely been taken, and the next savings come from better use, task-aware routing, and serving designs that lift throughput without lifting error rates.³ The market still rewards efficiency. It has stopped rewarding accumulation.
The capex numbers confirm the buyers know it. Even the operators spending the most are not behaving like firms that have solved the economics. Alphabet's capex is projected to nearly double in a single year, toward 175 to 185 billion dollars, and Google moved its consumer AI from unlimited access to metered credits, a concession that flat-rate token consumption is unsustainable even at hyperscaler margins.⁴ Spending more is not winning. It is treading water at greater cost.
The new edge is throughput per dollar of capacity
When accumulation stops working, the question changes from how much did you buy to how much work did you get out of it. That is the efficiency frontier, and the metrics that define it are not the ones the old game used.
The winning numbers are throughput per watt, per rack, and per dollar of capital.³ How many useful tokens does a megawatt produce. How much work comes off a rack before the next one has to be powered, which in a market where the next megawatt is years out is the only capacity that arrives on demand. The unit of analysis moves too, from dollars per raw token to dollars per successful workflow step, because caching and reuse make the raw token count a poor measure of what a job really costs.⁵ An operator that measures GPU-hours held is counting the old thing. An operator that measures work delivered per dollar of installed capacity is counting the new one.
This is why fit is the competitive variable. Two operators can run the same hardware at the same price and deliver wildly different amounts of work, because one fits the work to the resource and the other holds expensive accelerators idle. At a fleet utilization near 5 percent, the gap between a good operator and an average one is not a few points of margin. It is multiples of output on the same capital.
How the market reorganizes
Follow that to its conclusion and the structure of the market changes in three ways.
Capex stops being a moat and starts being a liability if it is not used. A competitor who runs a smaller fleet at high fit can underprice and outproduce one who bought more and uses less, because the second operator is carrying depreciation on idle hardware that the first is not. The balance sheet that looked like strength becomes weight.
The discipline becomes routing, not buying. Teams already converge on tiered routing, a cheap model for most traffic, a mid-tier for moderate complexity, a flagship reserved for the hard tasks, because matching the work to the cheapest resource that can do it is the dominant cost lever once token prices stop falling on their own.⁶ Routing by task difficulty at the model layer is the same move as routing by resource fit at the execution layer. Both are fit. Both are how an efficient operator pulls ahead.
Efficiency compounds where accumulation plateaus. Buying hits a physical wall, power, transformers, interconnect queues, and a financial one, depreciation on assets that may not earn out. Fit hits neither. Every point of utilization recovered is output on hardware already paid for, and the operator who is better at it stays better as the fleet grows and fragments, because the advantage is a capability, not a purchase. One can be copied with a checkbook. The other cannot.
Who wins
The winner in this world is not the operator with the most hardware. It is the one who turns the most of its hardware into work, measured on its own bill, and keeps doing so as the fleet gets messier. That operator spends less to deliver the same result, holds less idle capital, and can price below a larger competitor while earning more per dollar deployed. It competes on how well it executes, not on how much it spent.
This is the quiet inversion the build-out has been hiding. The era that rewarded the biggest buyers is ending because the thing they bought stopped being scarce. The era that rewards the best operators is starting because the waste they can recover is enormous and the capacity they would otherwise buy is physically out of reach. Capacity was the last decade's game. Execution is this one's.
The point
Competing on execution means the edge comes from fit, not capex. Token prices have fallen far enough that owning more capacity is no longer a moat, the metrics that matter have shifted to work delivered per watt, per rack, and per dollar of capital, and the market reorganizes around operators who extract more from what they run rather than buyers who acquire more than they use. Capex becomes a liability when idle. Routing becomes the discipline. Efficiency compounds while accumulation plateaus against physics. The teams that fit work to machines will spend less, deliver more, and win on execution. That is the world this whole series has been describing, and it is the one arriving now.
A note on the numbers
Figures are reported with their source and what they measure. The roughly tenfold annual decline in token prices and the GPT-4-class drop from about 20 dollars to under one dollar per million tokens are reported trends current to Q1 and Q2 2026, and the precise figures vary by model and provider. The throughput-per-watt and per-rack framing is qualitative, drawn from how efficiency-focused analysts describe the competitive frontier, not a single benchmark. The capex projection and the metered-pricing shift are from public company guidance and reporting. The 5 percent fleet utilization figure is the same fleet-level average used across the series, and the claim that fit yields multiples of output is a direct consequence of recovering from a 5 percent base, not an independent measurement. As across the series, the argument does not rest on any single number. It rests on the shape of all of them together: cheap, capped capacity on one side, and a widening gap between operators who use it well and operators who do not on the other.
References
- Inference cost has declined roughly 10 to 100x from 2023 to 2026, driven by model efficiency, hardware advances, inference optimizations, and competitive pricing; it is the operational cost that determines AI application unit economics. Startups.com inference cost lexicon (Jun 2026). Startups.com
- Token prices have fallen about 10x per year since 2021; GPT-4-class capability that cost roughly 20 dollars per million tokens in late 2022 now runs closer to 0.40 dollars. DeepInfra inference economics. DeepInfra
- Once easy price cuts are exhausted, the competitive edge shifts to whoever delivers more throughput per watt, per rack, and per dollar of capital; further savings come from better use, task-difficulty routing, and serving designs that lift throughput without raising error rates. Sesame Disk on 2026 inference cost trends. Sesame Disk
- Alphabet capex projected to rise from 75 billion dollars toward 175 to 185 billion in a single year, mostly AI infrastructure; Google shifting consumer AI from unlimited access to metered AI Credits, signaling that flat-rate token consumption is unsustainable even at hyperscaler margins. Artefact on the token cost illusion. Artefact
- The better unit of analysis is dollars per successful workflow step rather than dollars per raw token, because context reuse and caching make the raw token count a poor measure of what a request costs. Sesame Disk. Sesame Disk
- Teams converge on tiered routing, a cheap fast model for most requests, a mid-tier for moderate complexity, and a flagship reserved for the hardest tasks, as the dominant cost lever; matching the work to the cheapest capable resource. DeepInfra inference economics. DeepInfra