Google restricts Meta's access to Gemini computing capacity

Google has restricted Meta's use of its Gemini models because it does not have enough computing capacity available to meet the full demand from the social media giant. That was reported by the Financial Times. Google informed Meta of this around March 2026, and the restrictions were still in place at the end of June 2026.

The measure has delayed a range of internal AI projects at Meta. Meta uses Gemini for a variety of tasks, from detecting fraud and removing harmful content to customer service, software development and advertising tools. Other Google customers are reportedly experiencing similar restrictions, but the impact is greatest at Meta due to its exceptionally high level of demand.

The incident exposes a structural tension in the AI sector: demand for computing power is growing faster than technology companies can build new capacity, even among those with the deepest pockets.

How Meta is responding to the shortage

To reduce its dependence on external models, Meta has encouraged employees to use AI resources more efficiently. Internally, the company measures that consumption in so-called AI tokens, units that track how much computing power a user or application consumes.

At the same time, Meta is shifting some of its workloads to Muse Spark, the proprietary AI model it introduced in April 2026. By running more tasks on its own infrastructure, Meta aims to reduce its exposure to capacity constraints at external providers. This shift reflects broader trends in the industry: large technology companies are increasingly building their own models and computing clusters to manage supply risks.

Google's infrastructure problem and the deal with SpaceX

Google's inability to meet Meta's demand is notable given the company's scale. In the first quarter of 2026, Google Cloud revenue exceeded $20 billion per quarter for the first time, while the backlog of signed but not yet fulfilled cloud contracts now stands at more than $460 billion. Demand is therefore growing faster than Google can bring new capacity online.

To bridge that gap, Google reached an agreement with SpaceX in June 2026 to lease computing capacity for approximately $920 million per month. This arrangement, in which a cloud provider itself leases capacity from a third party, indicates how acute the situation has become. Through its Starlink network and associated ground infrastructure, SpaceX has computing resources that Google can deploy in the short term.

CEO Sundar Pichai acknowledged in April 2026, when presenting the quarterly results, that the company is struggling to keep pace with capacity expansion. Google is investing tens of billions of dollars in chips, data centres and energy supply, but those investments do not translate into available computing power immediately.

What this says about the state of the AI industry

The conflict between Google and Meta over Gemini capacity is not an isolated incident but a symptom. Demand for AI computing power is rising at a pace that structurally outstrips the expansion of physical infrastructure, data centres, grid connections and the production of specialised chips. This applies even to companies investing billions per quarter in infrastructure.

For the Dutch and European AI scene, this is a relevant observation. Startups and scale-ups building their products on API access to large foundation models face the same risk as Meta: an external provider can limit available capacity, whether due to scarcity or by prioritising larger customers. Investors and founders currently developing AI applications increasingly view infrastructure security as an independent strategic variable, alongside model selection and data quality. The fact that European policymakers are simultaneously working on sovereign cloud and AI infrastructure, through initiatives such as IPCEI-CIS and the European AI Office, takes on an added layer of urgency in light of incidents like this.