Google Throttled Meta's Gemini Access Over Token Limits

Meta told Wall Street on July 1 it has excess AI compute to sell — three days after Google throttled Meta's Gemini access for eating every token Alphabet could serve, and while Meta still rents tens of billions in capacity from the neoclouds whose stocks crashed on the news.

What broke Tuesday:

• Bloomberg reported on July 1 that Meta is building a cloud infrastructure business to sell AI computing power and model access.
• CNBC confirmed Meta will sell excess computing power to outside customers; Jim Cramer verified the report on air.
• Meta shares closed up nearly 9%. CoreWeave and Nebius each sank about 12%, CNBC reported; Axios put Nebius down 17%.
• Meta declined to comment. Bloomberg's sources asked not to be named.

Two business models on the table:

• Model hosting: Sell developer access to Meta's AI models — including closed-weight Muse Spark — on Meta's own chips, similar to AWS Bedrock.
• Raw compute: Rent GPU cycles directly, mimicking neoclouds like CoreWeave.
• Both paths run through Meta Compute, led by infrastructure head Santosh Janardhan, Superintelligence Labs leader Daniel Gross, and president Dina Powell McCormick, per Bloomberg and TechCrunch.

The timing problem:

• On June 28, the Financial Times reported Google throttled Meta's Gemini access after demand exceeded what Alphabet could deliver — disrupting internal AI projects.
• Google CEO Sundar Pichai told analysts Alphabet is "compute constrained in the near term" and Cloud revenue would have been higher if Google could meet demand.
• Meta CFO Susan Li told investors Meta expects to remain "constrained through much of 2026" until its own data centers come online.
• Meta signed a $27 billion Nebius deal in March and a $21 billion CoreWeave expansion in April — then neocloud stocks cratered when Meta said it might become their competitor.

Zuckerberg telegraphed it:

• At Meta's May shareholder meeting, Zuckerberg told investors selling compute is "definitely on the table."
• His exact words, via Axios: "Almost every week there are different companies that come to us from outside asking us to both stand up an API service, or asking if we have compute that they could buy from us at some premium to what we've bought it at."
• He added: "We haven't done that yet, because we think that we have a use for the compute, but obviously if we get to a point where we feel that we have overbuilt, then that is an option that we have."
• TechCrunch noted Meta had committed $182.9 billion to AI infrastructure as of Q1, including Louisiana and Ohio projects Zuckerberg compared to the size of Manhattan.
• Meta plans up to $145 billion in 2026 capex on data centers and GPUs, CNBC reported.

Why Wall Street bought it anyway:

• Meta is copying Elon Musk's playbook: SpaceX via xAI started selling excess capacity in May, signing deals with Anthropic at $1.25 billion per month and Google at $920 million per month, per CNBC.
• Bernstein analyst Madison Rezaei wrote, via Axios: "One can absolutely debate the will-they-won't-they, but we do know that Meta is sitting on one of the largest data center footprints in the world."
• Unlike Google and OpenAI, TechCrunch noted Meta doesn't break out meaningful revenue from Meta AI or Llama — making infrastructure monetization the clearest path to justify the spend.

The market read-through:

• The neocloud selloff is the tell: investors read Meta's "excess capacity" story as deflation for GPU rental pricing — even before Meta confirms a launch date or pricing.
• If Meta becomes both CoreWeave's biggest customer and its biggest competitor, the AI infrastructure trade stops being about who ships the best model and starts being about who owns the megawatts.
• Meta's cloud pivot landed the same week JPMorgan flagged a dot-com-style divergence between surging AI hardware stocks and struggling hyperscaler spenders — a warning that capex stories need revenue lines, not just bigger buildouts.

What to watch:

• Whether Meta officially confirms Meta Compute and sets pricing.
• If neocloud contracts survive Meta entering the rental market.
• Whether Google restores Meta's full Gemini allocation — or keeps rationing the company now pitching spare GPUs to strangers.

Convina's view: Meta did not discover surplus compute — it discovered a surplus narrative. Google capped Meta's tokens on Sunday, Meta rents $48 billion in neocloud capacity on long-term contracts, and Susan Li still says the company is constrained through 2026. The July 1 cloud story worked because investors needed a reason to stop punishing Meta's capex — not because Zuckerberg suddenly has idle GPUs. Treat "excess capacity" the way you treat any hyperscaler press release: a financing instrument first, a physics report second. The winners of this cycle will not be the labs with the best benchmarks. They will be the companies that can turn megawatts into recurring revenue before the debt matures.