Google throttles Meta's Gemini access as AI compute demand outstrips supply

Google Throttles Meta's Gemini Access as AI Compute Demand Outstrips Supply

Google capped Meta's access to Gemini AI models around March after Meta sought more computing capacity than Google could fulfill, disrupting and delaying some of Meta's internal AI projects. The restriction remains in place.

The constraint reflects a broader reality: computing power has become the tech industry's scarcest commodity as demand for advanced models surges across the sector. Google told Meta it could not provide all the Gemini capacity the company wanted to purchase, according to three people familiar with the matter. Several other Google clients faced similar restrictions, though Meta was particularly hard hit because of its exceptionally high demand for Google's models.

In response to the cap and a broader push to streamline AI costs, Meta encouraged staff to be more efficient with AI token usage.

Why the numbers matter

Google's capital intensity underscores the scale of the constraint. The company spent $200 billion on capex, yet still cannot meet demand from a single customer. That frames the bottleneck not as a temporary shortage but as a structural imbalance between compute supply and enterprise appetite.

Meta's token spend across Google's systems is opaque. Google has not disclosed Gemini revenue, so investors cannot see what fraction of Meta's AI budget flows to Gemini versus other providers like Anthropic. The exposure is likely substantial: Gemini powers Google's search overviews (which operate in the quadrillion-token range), YouTube's video chat feature, the Gemini app, and integrated workflows across Meta's own platforms including Instagram and the Meta app, which collectively reach billions of users. Even low token density per user, scaled across Instagram's roughly 2 billion users, generates enormous aggregate demand.

Meta is also running internal Llama and Musac Spark workloads, but the company's use of third-party capacity appears to exceed its own model infrastructure, at least for certain workloads.

The cap is bullish for Google's cloud business and suggests demand destruction for Meta's projects is real—not speculative. What remains unknown is whether Google's throttling reflects pure capacity constraints or strategic choices about which customers get priority.