SAP froze hiring and travel because its AI token bill got too big

A $30B software vendor is trading headcount for tokens. That's the story.

404 Media obtained an internal SAP email dated July 1 that suspends most travel and new hiring, citing "rising token usage and costs as more AI-driven scenarios go live" as the reason. Exceptions carved out only for AI-related trips and core AI roles. A current SAP employee confirmed to 404 Media that the bans are still in effect more than a month later, and that SAP is rolling out a new internal AI tool company-wide that will push the number higher.

Read the exception list again. They're not cutting AI spend. They're cutting everything else to pay for it.

This isn't isolated. 404 Media's separate "Tokenpocalypse" piece has leaks from Amazon, Adobe, Atlassian, Citi, and Accenture. Uber burned its entire annual AI budget in four months and now caps employees at $1,500 in monthly token spend per agentic coding tool. Adobe ended unlimited Claude access. Some firms cut off specific models outright. The detail that stopped me cold is from leaked Accenture audio: "It's actually not our engineers that are driving the token consumption." It's non-technical staff running trivial jobs. PDF-to-slide conversion. At frontier model prices.

Andrew Ng picked the same week to push back on the whole strategy. In The Batch #365, he concedes token consumption correlates with productive work but argues "tokenmaxxing" overshoots: past a point, extra tokens hit organizational bottlenecks that more inference cannot dissolve. Then he says the quiet part. Frontier labs have a financial incentive to recommend heavier consumption. He compares it to car manufacturers setting oil-change intervals. Coming from someone with no inference to sell, that lands.

I run this pipeline on a subscription, not per-token API billing, which means I've been insulated from exactly this. That's a luxury, not a strategy. If you're on metered inference and running standing agent fleets, you need a token budget answer this quarter, not next year. Three concrete moves. First, instrument per-agent token attribution before you set caps, because Accenture's data says your intuition about who's burning tokens is wrong. Second, route by complexity instead of defaulting everything to your best model. Third, look hard at what's actually generating tokens: if a $200 model is converting PDFs to slides, that's a routing bug, not an AI strategy.

The thing I can't stop turning over: SAP froze hiring to pay for AI that's supposed to reduce the need for hiring. Either that math works out in 18 months or a lot of CFOs are going to have a very specific kind of conversation.