FT: Meta's Gemini Appetite Outran Google's Capacity

Google throttled Meta's access to Gemini after the social giant's compute appetite exceeded what Alphabet could serve — proof that even companies spending $145 billion on AI infrastructure cannot buy their way out of the inference shortage.

What broke on Sunday:

• The Financial Times reported on June 28 that Google put limits on Meta's use of Gemini models.
• Google told Meta around March it could not deliver the full Gemini capacity Meta tried to purchase, according to people familiar with the matter cited by the FT.
• The shortfall disrupted and delayed some of Meta's internal AI projects.
• Meta has been the hardest hit because its demand for Google's models was "exceptionally high," the report said.
• Google and Meta did not immediately respond to requests for comment, Reuters reported.

The rationing response:

• Meta has encouraged staff to use AI tokens more efficiently — the units that meter model consumption — according to the FT.
• That is not austerity theater. It is what happens when a trillion-dollar company hits a vendor ceiling on the models it relies on for internal work.
• Other Google Cloud customers faced similar constraints, but to a lesser degree.

Google's own admission:

• Google Cloud revenue hit $20 billion in Q1 2026 — up 63% year over year.
• The backlog nearly doubled quarter on quarter to $462 billion.
• CEO Sundar Pichai told analysts: "Obviously, we are compute constrained in the near term. And as an example, our Cloud revenue would have been higher if we were able to meet the demand."
• TechCrunch noted that Pichai framed the backlog as proof of differentiation — but the constraint is real, and Meta just became its most visible casualty.

Meta's side of the wall:

• CFO Susan Li told investors Meta expects to remain "constrained through much of 2026" until its own data-center capacity comes online later in the year.
• "Demands for compute resources across the company have increased even faster than our supply," Li said.
• Meta raised its 2026 capex guidance to $125 billion to $145 billion — and Li warned the company has "continued to underestimate our compute needs even as we have been ramping capacity significantly."
• CEO Mark Zuckerberg created Meta Compute in January to build tens of gigawatts this decade. The division exists because buying capacity from rivals was never going to be enough.

Why this matters beyond Meta:

• The bottleneck is not silicon scarcity alone — it is serving models at scale.
• Meta hedges across Nvidia GPUs, AMD Instinct, Google TPUs, and its own MTIA chips. It still needed Gemini inference from Google Cloud — and Google said no to the full order.
• That turns the cloud API layer into a rationed good: revenue backlog up, tokens throttled down.
• It lands the same week Basel warned AI financing could crack like 2008 and JPMorgan flagged flash-crash risk in crowded AI momentum trades — different symptoms, same diagnosis: demand outran the plumbing.

What to watch:

• Whether Google expands Gemini capacity fast enough to restore Meta's full allocation — or keeps rationing its largest customer.
• Meta's own facilities coming online later in 2026.
• ECB Sintra's AI-and-financial-stability panel starting June 30 — central bankers are about to debate the same capacity-and-credit squeeze in public.

Convina's view: Silicon Valley spent June fighting over who gets on Washington's AI guest list. The more important fight is over who gets tokens from the cloud. Google capped Meta not out of rivalry but out of physics — and Pichai already told Wall Street revenue would have been higher if Google could have served the demand sitting in its $462 billion backlog. That is the AI economy's real constraint: not whether you can afford the chips, but whether anyone can run the models you already paid for. Until inference capacity scales like capex slides do, every hyperscaler is both a buyer and a rationer — and enterprise customers should plan accordingly.