In April, Uber's CTO Praveen Neppalli Naga sat down with The Information and admitted, on the record, that the company had burned through its entire 2026 AI tooling budget by the end of April. Four months. Three quarters of the year still ahead. The quote — "I'm back to the drawing board, because the budget I thought I would need is blown away already" — is the kind of thing CFOs frame on the wall as a warning.
Three weeks later, Fortune ran a story whose headline did most of the work: Microsoft reports are exposing AI's real cost problem: Using the tech is more expensive than paying human employees. A few days after that, news leaked of an unnamed enterprise that had managed, in a single month, to ring up a $500 million bill on Claude. In a single month.
This is the story I want to tell, because I think it has been told badly by almost everyone covering it. The headlines call it a surprise. It wasn't. It was visible from the beginning to anyone who bothered to do the multiplication.
The arithmetic that nobody did
Let's do the multiplication now, with current public numbers.
A senior engineer at a US tech company costs the company, all-in, somewhere between $25,000 and $40,000 per month. That number includes salary, benefits, taxes, equipment, the share of office or remote-work overhead. It is the number a CFO uses internally.
A heavy user of an agentic coding tool — Claude Code, Cursor, Codex — costs the company between $500 and $2,000 per month in token consumption, depending on workflow. Uber's published range was $150 to $250 on average, and $500 to $2,000 for power users. OpenAI's own planning note for Codex puts the per-developer figure at roughly $100 to $200 per month, with substantial variance.
The ratio looks comfortable. A $2,000-per-month tool against a $30,000-per-month engineer is a 6.7% overhead. If it makes the engineer 10% more productive, you've made money. That is the math every CTO walked into 2026 with.
The math is also, in a specific way, dishonest. Because the $2,000 number is per engineer. Multiply it across the engineering organization and the same engineer who cost you $30K now costs you $32K. Across a 5,000-engineer org — like Uber's — that's an extra $10 million per month, or $120 million per year, that was not in last year's budget. And the consumption is growing. Adoption at Uber went from 32% to 84% in months. The bill scales with adoption faster than headcount ever could.
The mistake wasn't believing AI would be cheap. It was assuming AI's cost would behave like software's cost. It doesn't. It behaves like cloud compute — and we already know how that story ends.
Why nobody saw it
I want to be careful here, because the temptation is to dunk on executives who didn't predict this. That dunk is too easy and also wrong.
The reason most CEOs didn't see this coming is that software has, for forty years, behaved like a fixed cost. You pay for the seat, the seat is yours, the seat costs the same whether you use it twice or twice a million. The entire enterprise procurement apparatus — every CFO instinct, every spreadsheet template, every multi-year contract — is built around per-seat licensing. Token consumption is a category of expense that did not meaningfully exist in any enterprise software stack three years ago.
When Anthropic and OpenAI and Cursor priced their consumer tiers at $20 a month, they were doing something subtle. They were anchoring the executive imagination on a per-seat number that bore no relationship to the actual cost of running these tools at scale. The CEO who approved the rollout was thinking $20 per seat times 5,000 seats equals $100K a month, fine. They were not thinking $1,500 per seat times 5,000 seats times 84% adoption equals $6.3M a month, because the $1,500 figure was nowhere on the pricing page.
I do not think this was a malicious bait-and-switch. The vendors themselves seem to have underestimated how much agent loops would consume once the tools became genuinely useful. But the result, on the consumer side, is that the people approving these contracts were looking at the wrong number. And the right number — the consumption-driven number — only becomes visible in retrospect, after a quarter of usage, when the bill arrives.
The Nigerian arithmetic
I write this from Lagos. A junior software developer here, in 2026, earns somewhere between ₦150,000 and ₦400,000 per month — roughly $100 to $265 at the current exchange rate. A mid-level developer with three years of experience earns $700 to $1,800. A senior at a strong local fintech, $1,500 to $3,000.
A heavy Claude Code user at Uber, on the high end of the published range, costs $2,000 per month in tokens alone. The tokens cost more than the developer in Lagos that uses them. The tool is more expensive than the person operating it, by a multiple, in this market.
This is the part nobody in San Francisco is saying because nobody in San Francisco is doing the conversion. But if you are reading this from Nairobi or Kraków or Karachi, you already know: the cost structure of AI was never priced for your economy. The American CFO is being mildly inconvenienced. The Nigerian founder, trying to compete on the same tooling, is being priced out of the work.
A tool that costs more than the human using it isn't really a tool anymore. It's a luxury good. And the question of who can afford luxury goods is, eventually, a question about who gets to do the work.
What the spring told us
The interesting part of the Microsoft, Uber, and $500M-bill story isn't the cost itself. It's what the companies did about it. This is the signal worth reading carefully.
Microsoft began winding down most internal Claude Code licenses in mid-May, with the Experiences and Devices division losing access on June 30. Microsoft. The company that owns 49% of OpenAI and ships GitHub Copilot. Even they could not make the consumption-based economics work internally at scale. The conclusion is not that AI doesn't work. The conclusion is that AI works so well, and is used so heavily once it works, that the bill becomes the limiting factor.
Uber's response was different and instructive. They didn't cancel access — they imposed a $1,500 per-tool per-month cap on every engineer. This is the move of a company that has accepted the cost as real and is now treating AI tokens like cloud compute: a resource that must be metered, budgeted, and explained on the way out. The era of letting engineers use whatever they need lasted, in Uber's case, somewhere between six months and a year. It is over.
The $500 million single-month bill story is the third archetype. That enterprise had no caps, no controls, no governance — they had treated AI like a SaaS product and let employees provision freely. They will, by year-end, have rebuilt the entire FinOps stack around inference. Their AI rollout will look, by next quarter, more like AWS than like Microsoft 365.
What CEOs are actually painfully realizing
The part I find most interesting is what these executives are quietly realizing, behind the press statements. It is not the cost. They can absorb the cost. It is something subtler.
They are realizing that AI isn't a productivity multiplier that you buy and own. It's a productivity multiplier that you rent, indefinitely, with the rental price set by a vendor whose pricing power increases as your dependency does. Every dollar of developer salary that gets replaced by a dollar of AI tokens is a dollar that moves from a relatively predictable expense — an engineer who will probably still be here in two years at roughly the same cost — to an extremely unpredictable one. Anthropic could change their pricing tomorrow and you cannot meaningfully push back.
You cannot lay off Claude. You cannot negotiate Claude's salary down at the annual review. You cannot tell Claude that the company is having a tough quarter. You can only use less Claude, which means doing less of the work. This is the dependency math nobody factored in when they were doing the giddy 2024 projections about AI replacing half their engineering.
The honest assessment, six months into 2026, is that AI hasn't replaced engineers. It has made engineers and AI complements, where the cost of running the AI now exceeds the cost of paying the engineers in most low-cost-of-labor markets, and approaches it in the high-cost ones. The companies that spent 2024 firing junior engineers to make room for AI tools are, in 2026, discovering that they replaced a $5,000-per-month payroll line with a $1,500-per-month token line, and a senior engineer who is now reviewing AI output instead of mentoring the junior they fired. The net cost is roughly the same. The org chart is different. The pipeline of future seniors is empty.
What this means for the rest of us
There is a small set of practical takeaways here for anyone who isn't running an enterprise.
If you are an engineer in a market where developer salaries are low — and that includes most of the world — your relative cost has gone down, not up. A Lagos engineer at $1,800 a month is now meaningfully cheaper than the tooling the San Francisco engineer is using to do the same work. The arbitrage on labor is widening, not closing. This is good news that nobody is telling you because the people writing about AI mostly live in San Francisco.
If you are starting a company, the choice between hiring one US engineer and hiring two engineers in Lagos plus running Claude Code on them is no longer obvious in the way it was twelve months ago. The token cost of running the AI is starting to dominate the labor cost of the people using it. Build your stack accordingly.
If you are a CFO, the unsexy truth is that inference is a new line item that needs the same governance you applied to cloud compute in 2014. Caps, dashboards, FinOps reviews, vendor diversification. Uber rebuilt this in two months under duress. You can build it now, calmly, before the bill arrives.
I have one Anthropic Pro subscription and one Cursor subscription. I am, by the standards of this essay, a light user. I pay $40 a month for the tools. The tools — used well — let me ship roughly what two engineers might have shipped in 2022. The math, for me, still works.
If I were running 5,000 engineers in San Francisco, I would not be sleeping well right now.