Google Capped Meta's Gemini Access, Forcing Reliance on Own AI Models

Weeks before Alexandr Wang teased a major Muse Spark update, Google capped Meta's Gemini access after the company sought more capacity than it could provide.

Google has refused to sell Meta all the Gemini AI computing capacity it wanted, telling the social media giant around March 2026 that it simply could not meet the demand, according to a Financial Times report. The restriction—which remains in place—has disrupted and delayed some of Meta's internal AI projects and pushed the company to tell employees to go easy on AI tokens, the units that measure AI usage. Reuters, which could not independently verify the report, said both companies declined to comment. The timing is what makes this sting. The cap landed weeks before Alexandr Wang—Meta's chief AI officer and widely reported to be among its highest-paid hires after the blockbuster Scale AI deal—announced on X that a new Muse Spark update with "big improvements in coding and agentic capabilities" was coming soon. Coding and internal workflows are precisely what Meta had been renting Gemini for. In effect, the world's largest social media company was caught depending on a direct rival for everyday AI plumbing, got rationed for asking for too much, and is now racing to build its way out.

Why Meta was using Google Gemini instead of its own Llama AI models

The dependence was a quiet admission of a capability gap. Meta initially chose Gemini because it performed better than the company's own open-source Llama models, the FT reported, citing people familiar with the matter. Inside Meta, Google's AI has been doing the unglamorous heavy lifting—automating safety processes like rooting out scams and taking down harmful content, running customer service and advertising help chatbots, and assisting with coding and internal workflows, alongside other external models including Anthropic's Claude. Several other Google clients were affected by the capacity restrictions too, but Meta was hit hardest because its demand was exceptionally high, the report said.

Meta's AI token spending: From 'tokenmaxxing' leaderboards to usage limits

The cap has collided with a spending problem Meta was already confronting. Earlier this year, the company actively pushed employees to consume as many AI tokens as possible—a trend the industry dubbed "tokenmaxxing," complete with internal leaderboards celebrating top users. The New York Times reported in June that the era has flipped into its opposite, with Meta telling employees it would limit AI use after seeing an exponential increase in costs, and directing engineers toward its internal coding assistant instead of third-party tools where possible. By one SemiAnalysis estimate, Meta's per-employee spending on AI tokens had reached roughly $50,000 annually at list prices, with employees burning through over 60 trillion tokens in a single 30-day period in early 2026.

Meta's has a plan to end its reliance on rival AI models

Meta's escape route runs through its own models. The company has begun prioritising Muse Spark—the first model from Meta Superintelligence Labs, launched in April—which is viewed internally as more competitive with Gemini and reduces dependence on external models for some applications, per the FT. Wang has since said the next update, which Business Insider reported is codenamed Watermelon and uses far more compute than its predecessor, has already caught up with OpenAI's flagship GPT 5.5 in internal testing. The infrastructure bet is equally enormous: Meta has committed to investing $600 billion in the US by 2028 and plans capital expenditure of up to $145 billion this year, most of it on AI. Whether Muse Spark's upgrade arrives fast enough to make the Gemini cap irrelevant—rather than just embarrassing—is the question Meta now has to answer.