In a revealing turn of events behind Meta’s recent AI announcements, the spotlight has fallen on the company’s unexpected reliance on Google’s Gemini models. As the industry’s competition escalates, the cracks in Meta’s AI self-sufficiency narrative became evident around March 2026. According to the Financial Times, Google informed Meta it could not provide the computational power that the tech giant sought, raising questions about its operational autonomy.
Meta’s Heavy Dependence on Google’s AI
Meta’s internal AI framework appears to be intricately woven with Google’s Gemini, a fact that has been more profound than public knowledge suggested.
As reported, Meta’s dependency on Gemini sheds light on the company’s operational tactics. Internal sources disclosed that Gemini was preferred for specific tasks because it consistently outperformed Meta’s own models. Moreover, Meta employed other models, such as Anthropic’s Claude and Llama, as part of a strategy to diversify its AI resources rather than tying the company to a single external provider.
Insights into this complex relationship reveal:
- Gemini underpinned a range of functions including customer service and content moderation
- After requesting additional compute resources, Google placed capacity limits on Meta
- Meta’s usage demand was noted to be atypically high compared to other Google Cloud users
- Employees received directives to limit their usage of “AI tokens”, a measure of computational capacity
- In light of these constraints, Meta began transitioning some tasks to Muse Spark, its in-house model tailored for content moderation
The Growing Token Crunch
The global demand for AI compute resources continues to surge, outpacing even the most robust infrastructure providers.
This situation is not indicative of Google penalising Meta; rather, it’s a reflection of the structural constraints within the industry. The demand for Google Cloud’s AI services has skyrocketed, driving the backlog to a staggering $460 billion, as noted in Alphabet’s recent earnings report. The processing capabilities of its models have increased by 60% quarter-over-quarter, reaching over 16 billion tokens per minute. Despite substantial investments in data centers, the industry is unable to keep pace with demand, akin to the GPU shortage witnessed during the crypto boom.
Faced with these challenges, Meta has acted swiftly, committing to invest up to $600 billion in data center expansion by 2028, alongside the development of custom MTIA chips in collaboration with Broadcom. This initiative mirrors the ambitions of the Stargate Project, underscoring an accelerated race for control over AI infrastructure.
In a striking shift, per CNBC, Meta has begun phasing out human moderation providers such as Accenture and Concentrix, allowing Muse Spark to take over essential moderation roles previously managed by Gemini.
The uncomfortable truth behind the “AI-native” label suggests a veneer of marketing sophistication. Amidst grand presentations, the reality is that leading platforms are purchasing critical AI capabilities from competitors, illustrating the fragile interdependencies within the space. Meta may be striving for self-sufficiency, but its recent communication regarding token usage could be likened to discovering that a beloved establishment sources its signature offerings from a neighboring business. The future of this dynamic rests on who ultimately commands the crucial AI infrastructure—and that list remains alarmingly limited.
