Google has refused to promote Meta all of the Gemini AI computing capability it wished, telling the social media large round March 2026 that it merely couldn’t meet the demand, based on a Monetary Occasions report. The restriction—which stays in place—has disrupted and delayed a few of Meta’s inner AI initiatives and pushed the corporate to inform workers to go straightforward on AI tokens, the items that measure AI utilization. Reuters, which couldn’t independently confirm the report, stated each firms declined to remark.The timing is what makes this sting. The cap landed weeks earlier than Alexandr Wang—Meta’s chief AI officer and broadly reported to be amongst its highest-paid hires after the blockbuster Scale AI deal—introduced on X {that a} new Muse Spark replace with “huge enhancements in coding and agentic capabilities” was coming quickly. Coding and inner workflows are exactly what Meta had been renting Gemini for. In impact, the world’s largest social media firm was caught relying on a direct rival for on a regular basis AI plumbing, acquired rationed for asking for an excessive amount of, and is now racing to construct its method out.
Why Meta was utilizing Google Gemini as an alternative of its personal Llama AI fashions
The dependence was a quiet admission of a functionality hole. Meta initially selected Gemini as a result of it carried out higher than the corporate’s personal open-source Llama fashions, the FT reported, citing individuals accustomed to the matter.Inside Meta, Google’s AI has been doing the unglamorous heavy lifting—automating security processes like rooting out scams and taking down dangerous content material, operating customer support and promoting assist chatbots, and aiding with coding and inner workflows, alongside different exterior fashions together with Anthropic’s Claude.A number of different Google shoppers have been affected by the capability restrictions too, however Meta was hit hardest as a result of its demand was exceptionally excessive, the report stated.
Meta’s AI token spending: From ‘tokenmaxxing’ leaderboards to utilization limits
The cap has collided with a spending downside Meta was already confronting. Earlier this yr, the corporate actively pushed workers to eat as many AI tokens as doable—a development the business dubbed “tokenmaxxing,” full with inner leaderboards celebrating prime customers.The New York Occasions reported in June that the period has flipped into its reverse, with Meta telling workers it might restrict AI use after seeing an exponential improve in prices, and directing engineers towards its inner coding assistant as an alternative of third-party instruments the place doable.By one SemiAnalysis estimate, Meta’s per-employee spending on AI tokens had reached roughly $50,000 yearly at record costs, with workers burning via over 60 trillion tokens in a single 30-day interval in early 2026.
Meta’s has a plan to finish its reliance on rival AI fashions
Meta’s escape route runs via its personal fashions. The corporate has begun prioritising Muse Spark—the primary mannequin from Meta Superintelligence Labs, launched in April—which is considered internally as extra aggressive with Gemini and reduces dependence on exterior fashions for some functions, per the FT.Wang has since stated the following replace, which Enterprise Insider reported is codenamed Watermelon and makes use of way more compute than its predecessor, has already caught up with OpenAI’s flagship GPT 5.5 in inner testing.The infrastructure wager is equally monumental: Meta has dedicated to investing $600 billion within the US by 2028 and plans capital expenditure of as much as $145 billion this yr, most of it on AI. Whether or not Muse Spark’s improve arrives quick sufficient to make the Gemini cap irrelevant—quite than simply embarrassing—is the query Meta now has to reply.