Google Put Meta on a Token Diet, and the 600 Billion Dollar AI Giant Had to Ask Permission
2026-06-28
“Meta bet 600 billion on AI and still ended up rationing tokens like a student watching their mobile data.”
The Financial Times broke it and everyone piled on the same day: Google Cloud told Meta, around March, that it could not sell it the full Gemini capacity it wanted, and the shortfall quietly delayed some of Meta's internal AI work. So the company that talks about superintelligence like it is already in the building had to get in line behind everyone else for the actual brains.
The internal response is the part that stings. Meta is telling staff to be more efficient with their tokens, the little units that measure AI usage, and shifting work onto its own in-house model, Muse Spark, from its Superintelligence Labs. This is the same Meta that pledged roughly 600 billion in cloud investment over two years and cut 8,000 jobs in May while reshuffling 7,000 people into AI roles. Six hundred billion committed, and the message to the troops is please use fewer tokens.
The real lesson is bigger than the embarrassment. The bottleneck in this whole boom is not ideas, it is compute, and even the richest companies on earth are fighting over the same scarce silicon. Google itself is reportedly paying SpaceX around 920 million a month for GPU access, so nobody is comfortable. Meta just got caught looking least prepared.
- The Financial Times reported, with CNBC, Bloomberg and others following same day, that Google Cloud declined around March to sell Meta the full Gemini capacity it requested.
- The shortfall delayed some of Meta's internal AI projects, and Meta told staff to be more efficient with their token usage.
- Meta is shifting workloads onto its own in-house model, Muse Spark, built under its Superintelligence Labs division.
- This follows Meta's roughly 600 billion dollar cloud investment pledge over two years and its May cuts of 8,000 jobs with 7,000 reassigned to AI.
- Google Cloud posted 20 billion in Q1 revenue, admitted compute constraints capped growth, and reportedly agreed to pay SpaceX around 920 million a month for access to 110,000 Nvidia GPUs.
- 01
The squeeze is actually forcing Meta to ship its own model instead of renting someone else's brain, which is healthier long term. And the whole episode usefully exposes the real constraint of the AI era: compute, not vision, is what everyone is short of.
- 01
Meta, a trillion dollar company that publicly stakes 600 billion on AI and then has to ration tokens and ask a direct rival for capacity. When your AI ambition outruns your access to the chips that power it, the swagger starts to look like a press release with no plug socket.
Got something the world should see roasted? Drop it.
A full teardown from €2,99. No mercy.