Bill Arrives
Microsoft spent six months giving thousands of its employees free access to an AI coding tool called Claude Code and engineers loved it. So did product managers and designers, who aren't even supposed to be the primary audience. Then this month, Microsoft cancelled most of those subscriptions and told people to switch to a different tool by June 30, the last day of its financial year. The official reason is that Microsoft wants everyone on the same system. The real reason is that someone looked at the invoices.
The invoices were genuinely shocking. Uber's head of technology told a trade publication in April that his company spent its entire planned AI budget for 2026 in four months. I have heard PMs say that they delivered a year's roadmap in four months. If that is the math, then maybe its not so bad. Individual engineers were running up between $500 and $2,000 a month each, just in usage fees. The problem is that the tools work well enough that people use them constantly, and constant use is what makes the costs explode. Most software is priced like a gym membership, a flat monthly fee whether you show up or not. AI tools are priced more like a taxi, you pay for every mile the model thinks, and these tools think a lot. The more capable they get, the more they think per task, which means each new generation tends to cost more to run than the one before it.
The math that companies were hoping for, that AI would replace expensive engineers and save money, is turning out to be true in a narrower set of situations than the pitch decks suggested. A research team at MIT found that AI automation actually comes out cheaper than human labour for about a quarter of the jobs people expected it to displace. Gartner, which tracks enterprise technology spending, now says that only 28% of AI projects fully deliver on their business case, and that a quarter of planned AI budgets for this year will simply roll over into next year unspent, because the projects couldn't justify the cost. Nvidia, the company that makes the chips powering all of this, has said internally that what it spends on compute now exceeds what it spends on the employees using the tools. That's the chip company saying it.
All that said, the free-wheeling experimental phase is ending. This is where large companies were willing to hand out subscriptions freely and absorb whatever costs came back, on the basis that the learning was worth it. What replaces it will look more like how companies buy electricity: usage caps, tiered access depending on your role, and someone in finance who gets an alert before the bill doubles. The question companies are now working through isn't whether AI tools are useful. Most of them are. The question is whether the output is worth the running cost, measured honestly, task by task. For a lot of the use cases that generated the most excitement over the past two years, the answer is coming back as: not yet.
The invoices were genuinely shocking. Uber's head of technology told a trade publication in April that his company spent its entire planned AI budget for 2026 in four months. I have heard PMs say that they delivered a year's roadmap in four months. If that is the math, then maybe its not so bad. Individual engineers were running up between $500 and $2,000 a month each, just in usage fees. The problem is that the tools work well enough that people use them constantly, and constant use is what makes the costs explode. Most software is priced like a gym membership, a flat monthly fee whether you show up or not. AI tools are priced more like a taxi, you pay for every mile the model thinks, and these tools think a lot. The more capable they get, the more they think per task, which means each new generation tends to cost more to run than the one before it.
The math that companies were hoping for, that AI would replace expensive engineers and save money, is turning out to be true in a narrower set of situations than the pitch decks suggested. A research team at MIT found that AI automation actually comes out cheaper than human labour for about a quarter of the jobs people expected it to displace. Gartner, which tracks enterprise technology spending, now says that only 28% of AI projects fully deliver on their business case, and that a quarter of planned AI budgets for this year will simply roll over into next year unspent, because the projects couldn't justify the cost. Nvidia, the company that makes the chips powering all of this, has said internally that what it spends on compute now exceeds what it spends on the employees using the tools. That's the chip company saying it.
All that said, the free-wheeling experimental phase is ending. This is where large companies were willing to hand out subscriptions freely and absorb whatever costs came back, on the basis that the learning was worth it. What replaces it will look more like how companies buy electricity: usage caps, tiered access depending on your role, and someone in finance who gets an alert before the bill doubles. The question companies are now working through isn't whether AI tools are useful. Most of them are. The question is whether the output is worth the running cost, measured honestly, task by task. For a lot of the use cases that generated the most excitement over the past two years, the answer is coming back as: not yet.
Filed under:
Technology