Meta’s AI Push Hits Google’s Compute Limits and Government Scrutiny

Google has started rationing access to its Gemini AI model for Meta. The reason: a shortage of compute resources. Despite spending $920 million a month to lease 110,000 Nvidia GPUs from SpaceX, Google can’t keep up.

Meta is pushing hard to replace human content reviewers with AI. About half of review requests are already handled by large language models this year. The company aims to automate over 90% of content reviews for some categories by the end of 2026.

Meta’s internal AI model, Muse Spark, launched in April but faced delays. Its developer API was supposed to roll out in April but slipped to June with no firm launch date. Muse Spark offers “Instant” and “Thinking” modes, signaling new AI interaction styles.

On the model front, Meta released Llama 3.3 in December 2024, a 70-billion parameter text-only model tuned for multilingual dialogue. Llama 4 followed in April 2025 with a Mixture-of-Experts design and native multimodal capabilities. Both models come with a strict 700 million monthly active user cap under Meta’s Llama Community Licence.

Meta is playing a dual strategy. It pushes self-hosted Llama models as open-weight alternatives while promoting its proprietary Business Agent across WhatsApp, Messenger, and Instagram globally. The Business Agent is a key AI product tying Meta’s ecosystem together.

Meanwhile, the US government is stepping up pressure on Meta. On June 25, 2026, officials urged the company to submit its AI models for evaluation amid safety and security concerns. Meta remains the only major AI developer refusing voluntary model review, unlike OpenAI and Anthropic, who collaborate with the government.

Other AI firms like Google, xAI, and Microsoft agreed to share early access to their models with the Center for AI Standards and Innovation. This agency, led by Commerce Secretary Howard Lutnick, enforces new AI oversight frameworks under the Biden administration.

Anthropic faced government orders to suspend access to its Mythos 5 and Fable 5 models for foreign nationals. This highlights an increasing regulatory clampdown on AI exports and access worldwide.

Meta’s internal message is clear: “Make more efficient use of AI tokens.” The company knows compute is scarce and expensive. Its 2026 capital expenditure guidance sits between $115 and $135 billion, reflecting the massive investment AI demands.

Google’s rationing of Gemini access underscores a broader issue. Even tech giants struggle to supply enough compute for cutting-edge AI. It’s a costly bottleneck that will shape innovation and partnerships in the AI race for years to come.