AI Is Too Expensive To Replace Us — Uber’s $3.4B AI Budget Just Proved It

The Uber CTO burned the entire 2026 AI budget in four months. Bryan Catanzaro at Nvidia says compute now costs more than his team. The “AI will replace workers” story is colliding with finance reality — and what comes next is more interesting than either the doomers or the boosters predicted.

TL;DR: The narrative that AI is making workers obsolete is starting to collide with the actual cost of running it. Uber’s CTO burned the company’s full-year 2026 AI budget in four months on Claude Code. Nvidia’s own VP of applied deep learning says compute now costs more than his team’s salaries. Token-based billing has created a class of enterprise cost that finance teams haven’t modelled. The result, at least for now, isn’t AI replacing humans — it’s AI making humans more expensive to keep, more strategically valuable to govern, and more central to the actual delivery of work than the replacement narrative would have you believe.

I was sitting in our team’s Tuesday planning session in Shoreditch a few weeks back when one of our engineering leads said something that’s stuck with me. We were talking, vaguely, about AI tooling budgets — the usual conversation, except he’d just come back from a finance review. “Mo,” he said, “the cost of one of our senior engineers running Claude Code for a month is now approaching the cost of a junior contractor.” Half-joke. Half not.

I went home that evening and started reading. And what I found was that the story I’d been internalising — the one every LinkedIn thinkfluencer has been pushing for two years, about AI replacing knowledge workers en masse — was quietly falling apart in the most ironic way possible. The tools work. They work too well. They’re being adopted faster than anyone forecasted. And the bill that’s arriving on CFOs’ desks is making the replace-the-worker math look very, very different from how it was being sold.

This is the article I needed to write to think it through properly.

The Uber Story Nobody Outside Tech Twitter Has Properly Reckoned With

In April 2026, Uber’s Chief Technology Officer Praveen Neppalli Naga sat down with The Information and admitted, in plain English, that the company had burned through its full-year 2026 AI budget in four months. “I’m back to the drawing board because the budget I thought I would need is blown away already,” he said.

The driver wasn’t infrastructure. It wasn’t a failed pilot. It was Claude Code — Anthropic’s terminal-native agentic coding tool — spreading through Uber’s 5,000-engineer organisation faster than anyone in finance had modelled. Adoption jumped from 32% to 84% between December 2025 and March 2026. By April, 95% of Uber engineers were using AI tools at least monthly. Roughly 70% of committed code at Uber now originates from AI. About 11% of live backend updates are written by AI agents with no human in the loop.

From a productivity perspective, this is a triumph. From a finance perspective, it’s a disaster.

The cost mechanics are what make this story matter. Claude Code isn’t priced like a SaaS seat — it’s priced on token consumption. Monthly costs per engineer at Uber reportedly range from $500 to $2,000, with averages around $150-$250 but power users spiking far higher. Uber’s overall AI-related costs have risen approximately 6x since 2024. R&D expenses hit $3.4 billion in 2025, up 9% year-over-year, with AI a key driver.

And here’s the bit that should stop you in your tracks if you’ve been told AI is going to make engineers obsolete: Naga said hiring hasn’t slowed. He’s still recruiting software engineers. The tools are too good to abandon, too expensive to sustain at current throttle, and somehow the humans are still in the loop and still needed.

Nvidia Admitted The Same Thing — Just More Quietly

This isn’t an isolated Uber problem. Bryan Catanzaro, vice president of applied deep learning at Nvidia — which is to say, the company literally selling the picks and shovels of the AI gold rush — told Axios in April 2026 that “for my team, the cost of compute is far beyond the costs of the employees.”

Read that sentence again. The company building AI infrastructure is publicly stating that the compute its team runs costs more than the team itself.

If you’re an executive listening to consultants tell you that you can swap headcount for AI agents and come out ahead, that quote should be doing something to your stomach. Because if Nvidia — the most cost-efficient AI operator on the planet, with deep technical knowledge of how to optimise inference — is saying compute outruns salary cost, the maths is not going to work in your favour at a typical enterprise either.

Worldwide IT spending is expected to hit $6.31 trillion in 2026, up 13.5% from 2025, with AI infrastructure driving the bulk of that increase. The “Premium Reckoning,” as some analysts are calling it, has arrived. Boards are demanding ROI proof. CFOs are looking at token bills and asking the question they should have been asking 18 months ago: what is this actually delivering, and is it cheaper than what it replaced?

The honest answer, in a lot of cases, is no.

Why The “AI Will Replace You” Math Was Always Sloppier Than It Looked

Here’s the part I’ve been thinking about as a PM, because I think there’s a real product lesson in this. The original case for AI replacing knowledge workers was built on three assumptions that are now all visibly cracking.

Assumption one: inference costs would fall faster than usage scaled. They did fall — the cost per million tokens for GPT-4-level performance has dropped by over 98% since early 2024. But usage scaled faster. When you give an engineer a tool that can refactor a whole codebase autonomously, they don’t use fewer tokens. They run agents in parallel. They delegate entire workflows. Token consumption per active user can 10x in a quarter. The deflation in unit cost is real and the inflation in unit consumption is bigger.

Assumption two: you could just substitute AI agents for headcount one-for-one. This is the assumption that’s looking most foolish in hindsight. Real-world enterprise AI deployments don’t replace a worker — they replace a task, while creating new tasks around governance, prompt engineering, evaluation, security review, and integration. Uber didn’t fire 4,000 engineers when it rolled out Claude Code. It made 5,000 engineers more productive and more expensive to support.

Assumption three: codified knowledge could be automated cleanly and that was most of the work. The Dallas Fed published research in February 2026 making exactly this distinction — AI can replicate codified knowledge (textbook stuff) but struggles with tacit knowledge (experiential, contextual, judgement-based). Their conclusion: AI tends to substitute for entry-level workers but augment experienced ones, which is showing up in wage data as rising salaries for experienced workers in AI-exposed occupations.

That last point is, honestly, the one I find most underrated in the broader discourse. AI isn’t a flat replacement layer. It’s a leverage multiplier for senior people and a competitive pressure on junior people. Which means the workforce ends up shaped differently — but not smaller in the way the narrative suggested.

What This Means If You’re A PM Right Now

I’ve been trying to figure out what to take from this in my own work. A few things have crystallised for me.

The first is that “AI ROI” needs to be a first-class metric in every PM’s roadmap, not an afterthought. If your team is shipping AI features, you need to be tracking the unit economics of each feature — not just engagement and retention. What does it cost per query? Per session? Per active user? Most PMs I know cannot answer this for their own products. We’ve been so focused on capability that we’ve ignored economics. That gap is about to bite.

The second is that the governance layer is where the next decade of PM jobs is going to be created. Only 43% of organisations have formal AI governance policies right now. Only 21% have mature agentic AI governance models. That’s a massive gap, and someone has to close it. The PMs who can build the internal tooling, guardrails, evaluation frameworks, and cost-control mechanisms around AI deployment are going to be more valuable, not less, than they were before AI showed up.

The third — and this is the one that took me a while to admit — is that the productivity gains from AI tooling are mostly accruing to the people who already know how to do the work. I’ve been using Claude Code daily for about eight months. It’s made me dramatically faster at certain things. But it’s made me faster at things I already knew how to do. The same engineer-lead I quoted at the start of this article used to spend three hours debugging a thorny backend issue. With Claude Code he might do it in 45 minutes. That’s a productivity gain on top of a person who already had the context and judgement to know what to debug.

That same tool, in the hands of someone fresh out of bootcamp without that context, doesn’t compress the work — it amplifies the confusion. And the cost is the same.

The Honest Version Of What’s Probably Happening

Here’s where I think this is actually heading, and I’m hedging because honestly, anyone who tells you they know with certainty what AI does to the labour market over the next five years is selling something.

Goldman Sachs Research, in April 2026, estimated AI is reducing monthly payroll growth by roughly 16,000 jobs and raising the unemployment rate by 0.1 percentage point — measurable, but a modest net drag, not the apocalypse. Citadel Securities tracked S&P 500 earnings calls and found 58.4% of Q4 2025 calls discussed AI in a workforce context, up from 17.1% in Q1 2021 — so the talk about AI replacing workers is enormous, but the actual workforce data shows a far more nuanced picture.

The likeliest scenario, if I’m being honest, is something like this. AI does displace specific job categories — particularly entry-level cognitive work — meaningfully but not catastrophically. AI also creates a new layer of work around governance, integration, and oversight that requires senior judgement. And AI quietly drives up the total cost of doing knowledge work for at least the next two to three years before the unit economics catch up, because the tools work well enough that adoption outpaces optimisation.

Which means, for now, the AI-will-replace-you story has it backwards. AI is currently too expensive to replace us. It’s making knowledge workers more productive and more expensive to support. It’s reshaping which skills get rewarded but not eliminating the skills layer wholesale. And the executives who bet the farm on workforce-reduction-via-AI in 2024 and 2025 are increasingly the ones explaining to boards why their token bill is bigger than their old payroll bill.

I don’t think this is permanent. The token deflation will eventually outpace usage growth. The economics will eventually catch up. Some categories of work will eventually be replaceable in a way that the unit economics actually support.

But “eventually” is doing a lot of work in that sentence. And in the meantime, the version of AI we’re actually living through is one where the humans are still here, still essential, and increasingly the ones being asked to figure out how the company is going to afford the tools they’ve been told would replace them.

The irony is almost too neat. Worth writing down before someone tries to rewrite the history of this moment.

What’s your read on this — is the cost story a short-term blip on the way to genuine replacement, or is there something more structural going on about why humans stay in the loop? Genuinely curious where other PMs are landing on this one.

If this resonated with you — or even just made you pause and think — I’d really appreciate a clap or two. It genuinely helps the article reach other product managers who’d find it useful.

And if you’re not already following along, I write weekly-ish about product management, AI tooling, and what’s actually working (and what isn’t) in my day-to-day as a PM in London. Hit follow, and you’ll catch the next one.

If you’ve got value from my writing and fancy supporting the work directly, you can buy me a coffee here:

Either way — thanks for reading.