The Token Reckoning
For every dollar in AI spend, eighty-two cents never makes it to production, according to the startup EntelligenceAI. That figure captures a crisis that has nothing to do with frivolous leaderboard-chasing or rogue employees running up token bills for fun. The real problem is structural: bug fixes, code rewrites, review loops, and rework are burning through most of the spend before a single shipped feature reaches a customer. Companies are blowing through annual AI budgets in months, triggering panic about runaway costs and cancelled initiatives. But the panic is misplaced. The waste isn't happening because teams are profligate. It's happening because the economic model underneath—how infrastructure is priced, how tokens are generated, and how value is measured—was built for a different era.
The shift from training-heavy to inference-heavy AI workloads is reshaping the entire cloud stack. New data sharpens the unit-economics question into something starker: the pricing model hyperscalers built for training is breaking under the weight of inference, and the companies that survive won't be those with the cheapest compute but those who solve the conversion problem—the eighty-two cents that vanish before production.
The infrastructure is priced by time, but the value flows through tokens. The gap is a thousand-to-one mismatch.
The mismatch starts at the infrastructure layer. Hyperscalers still rent GPUs by the hour, the same way they've done for a decade. You go to AWS, CoreWeave, or Lambda and pay two to four dollars per GPU-hour. That made sense when the workload was training: a big batch job, a cost center, something you spin up for days or weeks and then shut down. But inference is different. A single GB200 NVL72 rack—the latest generation of Nvidia's infrastructure—generates over one million tokens per second, or roughly four billion tokens per hour. At current hyperscaler pricing, renting the equivalent GPU capacity in that rack costs somewhere between one hundred fifty and three hundred dollars per hour. That's seventy-two GPUs at two to four dollars each. Now price those tokens instead.
Even at rock-bottom commodity rates—say, fifty cents per million tokens—that same rack is generating two thousand dollars of token value per hour. The infrastructure is priced by time, but the value flows through tokens. The gap is a thousand-to-one mismatch, and it's creating an arbitrage opportunity for anyone willing to move to per-token or credit-based models.
Jensen Huang spent much of his GTC keynote this year hammering this point home. He showed what he called a Pareto frontier, a chart mapping the tradeoff between throughput—how many total tokens you can pump out—and latency, how fast each individual user gets their response. The insight is that depending on where you sit on that curve, the economic value of a GPU hour changes dramatically. If you're optimizing for throughput, you can pack more inference requests into the same hardware and generate far more tokens per hour. If you're optimizing for latency, you're reserving capacity to guarantee sub-second responses, which means lower utilization but higher perceived value. Either way, the old model of charging by the clock no longer reflects what customers are actually buying.
They're buying tokens, and increasingly they're buying the orchestration, memory, and tooling that turns those tokens into agents that can do work.
The neoclouds and enterprise teams have noticed. CoreWeave's backlog ballooned to nearly one hundred billion dollars in the first quarter of twenty twenty-six, up from just under twenty-six billion a year earlier. Nebius raised their twenty twenty-six capex guidance from sixteen to twenty billion dollars to between twenty and twenty-five billion, explicitly citing higher demand. Both companies report that materially more than half of their compute is now being used for inference, not training. And both are starting to shift pricing models. The old seat-based or hourly rental approach is giving way to credit systems where credits effectively buy you a certain amount of consumption across any product the platform offers. It's an obfuscated per-token model, but it aligns pricing with actual token delivery in a way that hourly rental never could.
The problem is that generating tokens cheaply is only half the battle. If eighty-two percent of the tokens you generate never make it to production, you haven't solved the economics. You've just made the waste cheaper. That's the real reckoning. Uber's COO reportedly said that it was tough to justify all the AI spending because it wasn't leading directly to productivity. The comment went viral because it named what many enterprises are seeing but few want to admit: the ROI question remains unsolved. Token waste in service of productive use, not waste for the sake of waste, is the biggest problem in AI right now. Tokenmaxxing can be fixed by removing leaderboards or monitoring budgets closely. Unproductive uses of the technology can only be solved by finding productive uses.
The companies that do solve it will have a structural advantage that compounds over time. When you tightly align price with value delivered—charging more when your product does more, scaling with outcomes rather than headcount—the ceiling on what you can extract per customer goes way up. A traditional SaaS company selling seats was always capped by how many humans work at the company. An AI-native company selling on usage, outcomes, or work completed has a theoretically limitless TAM within each account. That's the flip side of the token economy: getting pricing wrong is fatal, but getting it right is a superpower.
The infrastructure build required to reach that point, though, is so expensive that only the companies with sustainable token economics will survive when the subsidy era ends. Google sold eighty-five billion dollars of stock, upsizing a planned eighty-billion-dollar raise after demand poured in. Berkshire Hathaway anchored it with a ten-billion-dollar investment, buying shares at a six to eight percent discount. The company sits on well over one hundred twenty billion dollars in cash and short-term investments, generated one hundred sixty-five billion in operating cash flow last year, and can borrow at rates few companies on earth can match. Firms in that position almost never sell equity. Diluting your own shareholders is the most expensive money there is—a slice of every future dollar of profit, forever.
But Google is doing it anyway because the AI buildout has grown so large it requires every available funding tool. The company expects to spend one hundred ninety billion dollars in twenty twenty-six alone on capex, and that figure could rise by year-end.
Amazon's numbers reportedly tell the same story from a different angle. The company is reportedly on track for roughly two hundred billion dollars in capital expenditures in twenty twenty-six, up from a trailing-twelve-month figure of one hundred forty-seven billion. Meanwhile, trailing-twelve-month operating cash flow grew to one hundred forty-nine billion at the end of March. A few years ago, that would have funded the whole thing with room to spare. It no longer does. Amazon is now spending essentially every dollar it generates. Free cash flow reportedly collapsed to just one point two billion over the past twelve months. Andy Jassy was explicit on the first-quarter call: the faster AWS grows, the more short-term capex they'll spend. Growth and free cash flow are pulling in opposite directions. Self-funding is a position of strength—you don't need anyone's permission to build.
The catch is that it caps your spend at what the business produces today. The AI bill is now testing that ceiling everywhere.
Equity dilution and cash flow destruction show the infrastructure build now requires every available funding tool: cash on hand, debt issuance, and selling pieces of the company. That's not a sign of weakness. It's a sign of how large the opportunity is perceived to be. But it also raises the stakes. If the eighty-two percent waste problem doesn't get solved, these companies are spending hundreds of billions of dollars to build infrastructure that generates tokens no one can afford to use at scale. The winners won't be the ones with the most GPUs or the lowest cost per token. They'll be the ones who've aligned pricing with actual token delivery and eliminated the conversion loss between spend and shipped product. The token reckoning isn't about how cheap intelligence can get.
It's about whether companies can turn that intelligence into something customers will pay for, consistently, at a margin that justifies the capital required to build it. The math only works if the eighty-two cents stops vanishing.