Google restricted Meta’s access to its Gemini AI models earlier this year because it simply could not supply the compute Meta needed. According to reporting by the Financial Times picked up by CNBC on June 28, 2026, Google told Meta around March 2026 that it could not meet the volume of AI capacity the social media giant had sought to purchase.

The shortfall disrupted and delayed some of Meta’s internal AI projects. Meta responded by encouraging its engineering teams to be more efficient with AI tokens. Google reportedly extended similar restrictions to other enterprise clients as well, not just Meta.

That last detail matters. This is not a story about Meta being unusual in its demand. It is a story about the entire industry running into a wall.

The Numbers Behind the Crunch

To compensate for the gap between its own compute capacity and what customers wanted to buy, Google struck a deal to rent 110,000 Nvidia GPUs from Elon Musk’s SpaceX, at a reported cost of $920 million per month. To be clear: Google, a company that has invested more than $180 billion in AI infrastructure, had to go to a competitor’s rocket company to keep its cloud AI business running.

Meta’s own AI ambitions give a sense of why demand is so extreme. The company laid off 8,000 employees in May 2026, redirected thousands more into AI-focused roles, and set capital expenditure guidance of $115 to $135 billion for the year. All of that spending creates enormous demand for AI inference and training capacity. Google could not absorb it.

What This Means for Business

If you are running an AI strategy at a mid-sized company, there is a tempting assumption that compute problems are someone else’s issue. Google and Meta have entire infrastructure engineering teams. You are just using the API.

But this story reveals something important about how AI infrastructure actually works right now. Even the best-resourced companies in the world are operating at the edge of what the hardware supply chain can deliver. GPUs are still constrained. Data centers take years to build. Demand is growing faster than supply.

For business leaders evaluating AI investments, this has practical implications:

AI costs are not yet stable. When even a hyperscaler like Google cannot meet customer demand and is renting compute from SpaceX, it signals that the underlying cost structure of AI inference has not settled. Businesses building AI workflows on top of foundation model APIs should expect pricing volatility and, occasionally, capacity constraints.

Token efficiency matters more than most teams think. Meta instructing its engineers to be more efficient with AI tokens was not just a cost-cutting measure. It was a response to actual supply constraints filtering down from Google. Teams that build sloppy AI workflows, sending enormous context windows for simple tasks, will face both cost and reliability risks as the market tightens.

Vendor concentration creates hidden risk. Meta was heavily reliant on Google’s Gemini models for internal use. When Google could not supply enough, Meta’s projects stalled. This is a vendor concentration risk that many enterprise AI teams are building into their systems without fully accounting for. Diversifying across AI providers, or using orchestration layers that can reroute between providers, is no longer just a procurement strategy. It is operational risk management.

The infrastructure buildout is real and the lag is real. The fact that Google is paying $920 million per month to SpaceX for GPUs it does not own tells you that even with nearly unlimited capital, you cannot bring new compute online fast enough to meet demand today. That lag has real consequences for every company betting on AI cost curves falling quickly.

The Larger Picture

The AI compute crunch has been a background story for two years. What this development makes visible is that it is affecting production systems at the most sophisticated AI users in the world. The FT’s reporting puts a real company name and a real disruption on what has mostly been an abstraction.

For Enterprise DNA’s clients thinking about AI transformation, the message is this: the infrastructure layer is still a genuine constraint, not a solved problem. Building AI workflows that are resilient to supply limitations, that use compute efficiently, and that do not concentrate risk in a single provider is as strategically important as choosing the right model for the task.

The companies that treat AI as pure software and ignore the hardware underneath are building on assumptions that the industry’s largest players just proved wrong.