July 7, 2026
The Price Is Not Right
This post is part of my Medium blog.
Two CNBC interviews, one week apart, and they're both saying the same thing from opposite directions.
Ed Zitron went on CNBC and made the bear case against generative AI — OpenAI's questionable finances, the lack of ROI, the whole thing as a symptom of a tech industry that's run out of hypergrowth ideas. A few days earlier, Alex Karp was on the same network saying "something has gone completely wrong" with how AI is sold, that enterprises are paying for "tokens that create no value," and that the token model used by OpenAI and Anthropic is broken.
I don't generally find myself agreeing with Alex Karp. But he's not wrong.
Here's what's obvious if you watch these companies for more than five minutes: they are designed to encourage waste. Not as a side effect. As the business model.
Jensen Huang said it out loud at GTC 2026. His thought experiment: you have a software engineer making $500,000 a year, and that engineer should be consuming at least $250,000 worth of tokens annually. When asked if Nvidia is spending around $2 billion a year on tokens for its engineering team, he said "we're trying to." This is the CEO of the company that sells the GPUs telling you that the engineers who use those GPUs should be burning half their salary in inference costs. He sells the shovels. Of course he wants you digging.
Then there's Steve Yegge's Gas Town, the multi-agent orchestrator that runs dozens of coding agents simultaneously and burns through thousands of dollars a month in API costs. It's a token disaster dressed up as innovation, and the same companies that want you to sacrifice your budget to tokens celebrate it. Maggie Appleton called it "utterly unhinged" and "a serious indication of how agents will change the nature of software development." She was being generous. It's a serious indication of how much money you can burn before anyone asks whether the output was worth the input.
And then there's tokenmaxxing — the practice, covered by the New York Times, of treating token consumption as a productivity proxy. Internal leaderboards at Meta and Microsoft ranking employees by how many tokens they burn. Engineers at Microsoft admitting they're "tokenmaxxing" not to get on the leaderboard, but because they don't want to be seen as using too few tokens. This is lines-of-code metrics from the 1990s, resurrected as a status game, and nobody in leadership is stopping it because the companies selling the tokens are the same ones benefiting from the culture.
I started to suspect something was up when Anthropic put so much effort behind Skills. The strategy of turning everything into inference tokens — wrapping every routine task in a model call, converting every workflow into a prompt chain — that started there. Skills aren't about making you more productive. They're about making you consume more tokens per task. The skill isn't the product. The token flow is the product.
Here's the reality of where this is going. Companies pay for outcomes, not tokens. And a lot of people are starting to realize this the hard way. You get a quick hit of dopamine generating a business report with Opus 4.8 xhigh or Fable 5, and it feels like the future. You do it every day and you start to notice the surprise bill — much larger than you realized, much larger than anyone warned you. If you're a small or medium-sized business, you also start to realize something worse: you don't have the weight to negotiate a legal agreement that protects your intellectual property. You're sending your data to a model provider who may be training your competition on your inputs. You don't know. You can't know. You don't have the leverage to ask.
The industry is moving toward open source. That's not a prediction — it's happening. The companies banking on token consumption as a permanent revenue stream are betting that people won't notice the alternatives, or won't care, or will be too locked in to switch. But the same dynamics that made cloud cost management a discipline — the slow realization that nobody was watching the bill, that the defaults were designed to spend, that the convenience tax was real — are showing up in AI spend right now.
And there's a consequence the token sellers aren't pricing in. Industries designed to encourage waste have a way of attracting regulators — and when the waste is the business model, regulators don't stay polite for long. They get creative: efficiency standards, mandated disclosure of compute per outcome, certification regimes that decide what gets deployed and where. I've spent a lot of time thinking about where that road leads — it's the world of The Condition Set, where the hidden cost of everything eventually gets a bureaucracy to match. That's fiction, for now. But the pattern is as old as industry itself: waste at scale invites intervention at scale.
So for the record: Karp's right. The token model is broken. The companies selling tokens are designed to maximize your consumption, not your outcomes. And the people starting to question their devotion to the biggest models aren't being contrarian. They're doing math.
In Essential, the third book in The Condition Set trilogy, the law that requires a human to remain in the decision loop is ninety days from expiring. Rob Coleman runs the agency that oversees every AI system in Canada. The question isn't whether the machines work. The question is whether anyone can tell when they stop working in the public's interest.