Google Pumps the Brakes on Meta’s Gemini Access as AI Demand Outstrips Supply

A Rare Look Inside the AI Capacity Crunch

Google has moved to limit how much of its Gemini AI technology Meta can buy, and in doing so it has exposed one of the industry’s most pressing problems: even the biggest players can’t get enough computing power. According to people familiar with the situation, Google decided around March that it simply could not hand Meta all the Gemini capacity the social media company requested.

Consequently, the cap has slowed and reshuffled several of Meta’s internal AI efforts. Rather than treating compute as an unlimited resource, Meta has since urged its engineers to spend AI “tokens” — the units that track model usage — far more carefully.

Why Meta Felt the Squeeze First

Notably, Meta isn’t the only Google customer affected, but it appears to be the hardest hit. Sources say a handful of other clients ran into similar limits, though to a smaller degree. Because Meta’s appetite for Google’s models runs so unusually high, it absorbed the brunt of the restrictions, which remain in force today.

This dynamic matters for a simple reason. When a supplier throttles one of its largest and most demanding customers, it signals that demand has outrun what the provider can realistically deliver. Indeed, the friction echoes a broader hardware scramble that PRISM has tracked closely — see NVIDIA Shares Dip as Meta Explores Buying Google’s TPU Chips, which captures how Meta has been hunting for alternatives to ease its compute bottleneck.

The Bottleneck Behind the Headlines

Across the industry, meanwhile, companies are pouring tens of billions of dollars into chips, data centers, and power — yet they still struggle to keep pace with surging demand for advanced models. Therefore, the Google–Meta episode reads less like a one-off dispute and more like a symptom of a broader shortage. For the wider landscape of the names racing to build this backbone, PRISM’s AI Datacenter and Infrastructure Index tracks the emerging players riding this wave.

A specific pressure point stands out: inference. Once a model finishes training, it still has to run, and those everyday workloads now pile up as businesses roll out chatbots, coding assistants, and AI agents at scale. As a result, inference has become one of the toughest challenges facing the entire sector.

How Google Is Scrambling to Catch Up

To close the gap, Google has moved aggressively to lock down additional capacity. Earlier this month, for instance, the company struck a deal worth roughly $920 million a month to lease computing power from Elon Musk’s SpaceX (SPCX). The scramble underscores just how quickly corporate demand — led by heavy users like Meta — has changed the math for cloud providers. Moreover, SpaceX’s rapid ascent as an infrastructure player is a story in itself, as PRISM detailed in SpaceX After the Pop: The $2 Trillion Story Hides a $2.58 Billion Loss.

The financial scoreboard tells a similar story. At its first-quarter earnings in April, Google reported that cloud revenue topped $20 billion for the first time, while its backlog of signed-but-undelivered cloud contracts nearly doubled from the prior quarter to more than $460 billion. Chief executive Sundar Pichai put it plainly, acknowledging that the company remains constrained on compute in the near term and that cloud revenue would have climbed even higher had Google been able to fully meet demand. Notably, Pichai has also stepped into the policy debate, as covered in PRISM’s Google CEO Calls for National AI Regulation as U.S.–China Competition Intensifies.

The Bigger Picture for Investors

Ultimately, the standoff highlights a tension that should interest anyone watching the AI trade. On one hand, providers like Google are booking record revenue and enormous order backlogs. On the other hand, they can’t always fulfill what customers want to buy — which both caps near-term upside and intensifies the race for infrastructure.

Tellingly, rivals are hedging the same way. Anthropic, the AI lab behind the Claude chatbot, recently signed a SpaceX compute arrangement that mirrors its existing deal with Google — a strategy consistent with its broader infrastructure push, including the $4 billion Amazon investment deepening its AWS collaboration. In short, the message across the industry is consistent: computing power, not ambition, has become the scarcest resource in AI — and the companies that secure it first stand to lead.

Both Google and Meta declined to comment.

Related PRISM Market View coverage

• NVIDIA Shares Dip as Meta Explores Buying Google’s TPU Chips
• Google CEO Calls for National AI Regulation as U.S.–China Competition Intensifies
• SpaceX Rockets 21% in Record Nasdaq Debut, Vaulting Past a $2 Trillion Valuation
• Anthropic’s Implied Valuation Tops $1 Trillion in Private Markets Ahead of Potential IPO
• OpenAI’s Spending Surged to $34bn in 2025 as IPO Looms
• PRISM Emerging AI Datacenter and Infrastructure Index