Companies had enthusiastically adopted AI when it had first emerged, but they’re now looking to be a lot more circumspect about its use.
Speaking on the Hard Fork podcast, Microsoft CEO Satya Nadella put his finger on why. The problem, in his framing, is that most enterprise AI adoption has been about volume — how many tokens are being consumed, how many engineers are using Copilot, how much code is AI-generated — rather than whether any of that translates to business value. “The marginal cost of productivity improvement has to match the marginal cost of the token,” he said. “That’s a management discipline.”
It’s a deceptively simple point with sharp implications. AI tools are useful enough that engineers reach for them constantly, and the usage numbers look impressive. But token costs are real, and they compound fast. Uber burned through its entire 2026 AI coding budget in roughly four months, with per-engineer monthly API costs running between $500 and $2,000. The usage was genuine. The ROI wasn’t there to match it.
Nadella was candid that Microsoft itself has been no exception. When asked how much token overconsumption — “token maxing,” in his words — had been happening internally, his answer was brief: “A lot.” The audience laughed, and the admission landed because it’s the last company you’d expect to cop to it publicly. Microsoft has staked its enterprise identity on AI, built GitHub Copilot into its developer stack, and made its investment in OpenAI the defining corporate bet of the decade. And even there, the discipline of matching token spend to actual value creation hasn’t always held.
The 10% GDP Question
In February 2025, Nadella told Dwarkesh Patel that his benchmark for AGI was 10% GDP growth. A year on, asked how he views that benchmark, he described the difficulty of translating a powerful general-purpose technology into measurable economic output. “The systems, right? The amount of change management that is required” — that’s where the friction is.
His formula for what would actually get to 10% growth is precise: “When you have a perfect match between the marginal cost of the token to the marginal value and it’s priced right.” The obstacle to that isn’t the quality of the models. It’s organizational discipline. Token spending without that discipline is just cost — and Nadella is emphatic that the current environment, where companies are pulling back from unconstrained AI use, is partly a symptom of this mismatch.
“Everybody goes and vibe codes and token maxes,” he said. “That’s not a way to achieve 10% growth.”
The Diffusion Problem
What Nadella is describing is a classical diffusion problem — the gap between a technology’s theoretical capability and its actual economic impact. The steam engine didn’t immediately produce the productivity gains economists later attributed to it. Electrification took decades to translate into factory-floor efficiency gains, partly because it required reimagining how factories were organized, not just swapping one power source for another.
AI is running into the same pattern. The models are powerful. The use cases are real. But getting from “engineers are using AI tools” to “the business is producing measurably more value” requires management infrastructure that most enterprises are still building. Nadella’s token-matching framework is essentially a proxy for that discipline: if you can’t show that the value you’re creating justifies what you’re spending per token, you’re not managing the technology — you’re just using it.
Palo Alto Networks CEO Nikesh Arora has made a related argument from the supply side — that high token pricing for enterprises is itself part of the problem, creating budget anxiety that leads to conservative deployments and underwhelming results. The two arguments aren’t in conflict. Cheaper tokens would make the matching equation easier to satisfy. But Nadella’s point is that even at current prices, the discipline isn’t optional. The math has to work.
What Actually Changes
The implication of Nadella’s framing is that the next phase of enterprise AI adoption will look quite different from the first. The first wave was characterized by deployment breadth — get AI into as many workflows as possible, measure adoption rates, and treat high usage as a proxy for value. The second wave, if Nadella is right, has to be about outcomes. Which token spend is generating productivity gains that the business can actually capture? Which workflows justify the cost, and which ones are burning budget on capabilities that aren’t translating to anything measurable?
That’s a harder question than “are your engineers using Copilot?” It requires connecting AI spend to business metrics in ways that most finance teams aren’t currently set up to track. It also requires accepting that some high-usage workflows might not be worth the cost — which is a difficult conversation in organizations that have built internal narratives around AI transformation.
Nadella’s candor about Microsoft’s own “token maxing” suggests he knows how hard that conversation is. The company that sold the enterprise world on AI is now making the case that the enterprise world needs to get disciplined about it.