They told us more tokens equals more productivity. Six months in, a lot of companies are staring at the bill and feeling a little sick.
Uber burned through its entire annual AI budget in four months. Not a typo — the CTO said it in an interview, back to the drawing board. Harvey, the legal AI shop, jumped to roughly 12 trillion tokens a month, a 12X increase. Databricks had an engineer who casually blew $7,000 on tokens alone. There were leaderboards for this. Real ones. Companies tracking each employee's token spend like it was a gym challenge, which honestly tells you everything about how weird this whole thing got.
Nvidia's Jensen Huang said something that basically became the corporate motto: if your $500,000 engineer isn't burning at least $250,000 worth of tokens, he'd be deeply alarmed. To be fair, the man sells the GPUs. Take that quote with the whole salt shaker.
The interesting part is the mechanical reasons costs explode. Model selection alone can move the needle 5-10x — Opus-class frontier models versus the optimized Haiku tier on the same provider. And context windows are the quiet killer: double the window, and compute can roughly quadruple, because every token attends to every other token. Feed one session a full novel and the same question gets dramatically more expensive. From my perspective, that's the biggest lever most teams ignore while they're busy arguing about which frontier model is "the best."
The predictable backlash has started. Microsoft killed standalone Claude code licenses and folded everyone into Copilot. Meta's talking about per-employee token caps. A few companies are quietly rationing AI by task type. The race is real, but so is the hangover.
And while we were all burning tokens, something quieter was happening to the people doing the burning. MIT strapped EEG monitors on writers, some with chatbot help, some without. The assisted writers produced faster drafts — but their brains showed markedly lower engagement while writing, and once the tool was removed, they performed worse than the people who never leaned on it. They call it cognitive debt: a deficit that accrues slowly and comes due only when the crutch disappears.
Dr. Vishal Kapoor, a researcher who works in banking, measured his own drop on a brain-training tool — roughly 40-50 points on a 1,000-point scale — after heavy AI-assisted reasoning. He calls the fix "AI-fed, human-led": think before prompting, never outsource the first question or the final decision. It's a soft claim, honestly, self-reported on brain-training apps, and I'm not about to overstate EEG science from one study. But the pattern keeps showing up in the research, and it's worth keeping in mind the next time you let a model do all the thinking for a week straight.
Privacy is the other conversation nobody wants to have. DuckDuckGo ran a survey and found about a third of people share secrets with chatbots they wouldn't tell friends, family, or their doctor. Among self-described AI enthusiasts it's 56%. Some of that is a feature — the chatbot doesn't judge you — but here's the ugly underbelly: unless you opt out, your chats get recorded and stored, can be used for training, subpoenaed, and last year hundreds of Claude transcripts ended up showing in Google search results. Your employer can read your work chats, too. The chatbot is a confidante the way a tape recorder is.
On the brighter side of the funding world, Instinct — the viral consumer AI assistant that runs your calendar and emails from Gmail — is reportedly raising $250M at a $2.5B valuation, co-led by Index Ventures and Benchmark. One early adopter used it to buy a house. To be fair, the terms of service let the company train on your data, which has some users squirming, and the founders say they're working on it. The invite-only rollout is basically Superhuman's playbook, and we've seen how that ends.
Quick add-on note: Thomson Reuters quietly launched its own in-house LLM, trained on decades of Westlaw and Reuters content, built on an open-source foundation. Fully controlled, domain-specific, zero dependence on someone else's frontier model. That's the quiet trend underneath all the noise — companies deciding they'd rather own the model than rent it.
Honestly, the past six months have been one big lesson in incentives. Vendors want you to burn more tokens. Your brain wants the easy win. The companies that figure out how to ration AI by task, keep the human in the loop, and stop chasing token leaderboards are the ones that'll still be standing when the bill comes due.
Anyway, that's where things stand. If you're in the middle of figuring out your own AI budget, or just trying to do the math on what your team actually consumes, a Math Calculator never hurts — numbers beat vibes when the CFO starts asking questions.
Top comments (0)