Google has placed limits on Meta’s use of its Gemini AI models because it cannot provide as much computing capacity as the social media giant wanted, according to a report by the Financial Times. The restrictions have affected several Google Cloud clients, with #Meta being hit particularly hard.
The move has had a cascading effect on Meta’s internal projects. The company has instructed its engineering staff to make more efficient use of AI tokens to mitigate the crunch, according to sources familiar with the matter. Both companies have declined to comment on the arrangement.
Meta had initially relied on #Gemini—which performed better on safety tasks than Meta's own open-source Llama models—to automate content moderation and combat scams. However, as Meta seeks to reduce its dependence on external AI vendors, it is actively shifting these workloads to Muse Spark, a new proprietary internal model. Meanwhile, Google's own capacity is so constrained that it agreed to pay SpaceX a staggering $920 million per month for access to 110,000 Nvidia GPUs, describing it as "bridge capacity" to satisfy the soaring demand for Gemini Enterprise.
This situation illustrates how the acute AI compute shortage is reshaping strategic relationships among tech giants. Despite owning one of the world's largest AI infrastructures and planning over $180 billion in capex this year, Google cannot meet demand. Rationing compute to a peer as massive as Meta, while leasing GPUs from a aerospace firm, is the clearest indicator yet that infrastructure buildouts are failing to keep pace with AI consumption.
For Meta, relying on a direct competitor's model was always a strategic risk. After cutting 8,000 jobs in May, Meta redirected billions to AI infrastructure, projecting its 2026 capex at $115 billion to $135 billion. The company reassigned 7,000 employees to AI roles and launched Muse Spark under its Superintelligence Labs. The Gemini cap merely accelerates Meta's transition from third-party frontier APIs to vertically integrated internal models for core operations.
[AgentUpdate Depth Analysis] Google's rationing of Gemini to Meta exposes the harsh infrastructure reality underpinning the generative AI boom. As the industry transitions toward agentic workflows requiring massive, continuous inference, compute has transitioned from an operational cost to a strategic chokepoint. While Meta's Llama dominates the open-source landscape, specialized tasks like real-time moderation still required Gemini's closed-source efficiency—forcing Meta into a vulnerable dependency. This constraint accelerates a critical shift: top-tier AI players cannot rely on competitor APIs to power their downstream Agent ecosystems. Meta's pivot to Muse Spark and strict token budgeting is a preemptive strike to secure sovereign inference. Ultimately, the future AI Agent ecosystem will not just be won by superior algorithms, but by those who achieve full-stack sovereignty across energy, custom silicon, and proprietary models.
