When AI Costs More

Now a Nvidia VP is telling us his AI compute bill runs higher than his entire team's payroll. Uber burned through its annual AI tools budget before spring. And somehow this is still being framed as a cost-reduction strategy.

Bryan Catanzaro, VP of Applied Deep Learning at Nvidia — a company whose entire business model depends on you believing AI is inevitable — said the quiet part out loud: "the cost of compute is far beyond the costs of the employees." This is not a skeptic. This is the guy whose employer sells the GPUs making all of this possible. When Nvidia tells you AI is expensive, you might want to update your spreadsheet.

Meanwhile, Uber's CTO let slip that their AI coding tools budget had been "blown away already" — and we're talking about a company that processes millions of transactions daily with infrastructure budgets most enterprises dream about. If Uber's AI bill is running hot in April, imagine what Q4 looks like when every engineering team has agents running on every pull request.

But here's what the press releases leave out: this was always going to happen. The economic logic of "AI replaces workers therefore costs drop" was always a demo-environment calculation. In production, the math gets messier. Intensive developer workflows — code generation, test cycles, architectural review — burn tokens at a rate that makes your enterprise subscription look quaint. Every agent loop, every context window reload, every parallel execution: that's compute, and compute costs money.

MIT ran the numbers in 2024. AI automation is economically viable in roughly 23% of roles — specifically those reliant on computer vision tasks like image inspection and visual quality control. For the other 77%, humans are still cheaper. The industry responded to this finding by announcing $740 billion in capital expenditures for 2026 — a 69% increase from the year before. AI software licensing fees climbed 20-37%. And companies quietly continued laying off the workers whose salaries were supposedly being replaced by cheaper automation.

The cynical reading: the workers got cut anyway, the AI costs arrived anyway, and now CFOs are explaining both to shareholders simultaneously. The optimistic reading: this is a temporary mismatch, compute gets cheaper, returns materialize, everything works out. Tech history suggests which reading has the better track record.

What's genuinely interesting — not good, but interesting — is that we're watching the hype cycle hit the ledger in real time. The promise of AI-driven cost reduction is colliding directly with the reality of AI-driven cost centers. Most enterprise technology eventually has to justify itself against an actual budget. AI just managed to get very expensive very fast before anyone had solid productivity numbers to point at.

The question nobody wants to field on an earnings call: what exactly is the ROI on $740 billion in capex when your own VP is telling you compute costs more than the people it was supposed to replace?

Start the timer. Somewhere in a boardroom, the pivot story is already being drafted.