The AI Compute Crisis: Google Pays SpaceX £720M a Month, Anthropic £975M

The week of 16 June 2026 produced a series of extraordinary revelations about the state of AI infrastructure. Google is paying SpaceX $920 million per month for compute capacity. Anthropic is paying SpaceX $1.25 billion per month to access the Colossus 1 facility. Microsoft, whose Azure cloud is one of the largest in the world, has been forced to route GitHub’s AI features through Amazon Web Services due to capacity constraints. The AI industry is facing a structural infrastructure crisis — and the consequences are being felt by every business and consumer that uses AI tools.

Why Is There an AI Compute Crisis?

Training and running large language models requires specialised hardware, primarily graphics processing units (GPUs) and custom AI accelerators. The demand for this hardware has grown exponentially since the release of ChatGPT in late 2022 and has continued to accelerate with each new generation of more capable models.

Nvidia, which manufactures the H100 and H200 GPUs that power most frontier AI systems, has been running at full production capacity for over two years. Lead times for large GPU orders remain measured in months. While competitors like AMD, Intel, and Google (with its TPU chips) are increasing supply, they have not been able to close the gap.

The result is a situation where even the world’s largest technology companies — Google, Microsoft, Meta, Anthropic — cannot get as much compute as they need from conventional sources. They are turning to unconventional providers, and SpaceX has emerged as one of the most significant.

SpaceX’s Role: From Rockets to AI Infrastructure

SpaceX’s involvement in AI compute is not widely understood outside the technology industry. The company operates the Colossus 1 facility — a large-scale data centre originally built to train xAI’s Grok models — which has spare capacity that SpaceX is now licensing to other AI companies. The facility reportedly houses one of the largest clusters of Nvidia H100 GPUs in existence.

Anthropic’s $1.25 billion monthly payment for Colossus 1 access is staggering in its scale. To put it in context, Anthropic’s total annual revenue in 2025 was estimated at around $1.5 billion. The company is now spending almost as much per month on compute as it earns per year — a situation that is only sustainable because of its substantial venture capital backing and the expectation of exponential revenue growth.

Google’s $920 million monthly payment to SpaceX is equally striking. Google operates one of the largest cloud computing networks in the world, including data centres across Europe, North America, and Asia. The fact that it is paying a competitor for additional capacity illustrates just how severe the supply constraint has become.

Microsoft Routes GitHub Through AWS

Perhaps the most counterintuitive revelation from the week was that Microsoft — owner of Azure, the second-largest cloud platform in the world — has been routing some of GitHub’s AI-powered features through Amazon Web Services. GitHub Copilot, the AI code completion tool used by over 10 million developers worldwide, requires enormous amounts of inference compute. When Azure capacity was insufficient, Microsoft turned to its most direct cloud competitor.

This is the computing equivalent of a major supermarket buying shelf space from a rival. It demonstrates that the compute crisis has transcended competitive dynamics. The demand for AI infrastructure is so great that companies are setting aside normal commercial considerations to access the hardware they need.

For UK businesses that rely on Microsoft Azure or GitHub Copilot, this routing through AWS is largely invisible — the service continues to function normally. However, it raises questions about data residency and compliance. UK organisations that have specific data sovereignty requirements under UK GDPR should verify with Microsoft that their data is not being processed on AWS infrastructure without their knowledge.

What the Crisis Means for AI Pricing and Access

When supply is constrained and costs are rising dramatically, prices tend to follow. Several AI API providers have raised prices in 2026, and the trend is likely to continue if compute costs remain elevated. Anthropic’s API pricing, OpenAI’s GPT-4 costs, and Google’s Gemini rates have all risen over the past 12 months.

For UK businesses that have built products or workflows on AI APIs, this cost inflation is a growing concern. A tool that cost £500 per month to run in 2025 may cost £750 or more in 2026. Businesses should model their AI costs under scenarios of 50% to 100% price increases to ensure their unit economics remain viable.

There is a countervailing force: smaller, more efficient models are becoming increasingly capable. Anthropic’s Claude Haiku, OpenAI’s GPT-4o Mini, and Google’s Gemini Flash are all capable of handling many common tasks at a fraction of the cost of frontier models. UK businesses that route appropriate tasks to cheaper models can partially offset the cost inflation at the frontier.

Gemini 3.5 Pro Still Not Publicly Available

The compute crisis provides context for another development from the week of 16 June: Google’s Gemini 3.5 Pro, widely anticipated as Google’s most capable model, remains in limited Vertex preview for select enterprise customers as of 19 June 2026. Polymarket traders are pricing approximately 50 to 55% odds of a public release before 30 June.

The delay almost certainly reflects compute constraints. Running a frontier model at scale for millions of users requires an enormous amount of GPU capacity. If Google is already paying SpaceX $920 million per month for additional compute, adding another frontier model to its public offering would require further capacity — capacity that may not yet be available.

UK developers and businesses that have been waiting for Gemini 3.5 Pro access should plan for the possibility that public availability slips into July or beyond. Alternative frontier models, including Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-4o, remain available through their respective APIs.

The UK’s AI Infrastructure Position

The compute crisis has significant implications for the UK’s AI strategy. The UK government has committed to building a domestic AI infrastructure, including the announcement of the AI Opportunities Action Plan in early 2026, which includes investment in sovereign AI compute. However, the scale of investment required to compete with the infrastructure being built by American companies and their SpaceX compute provider is enormous.

The UK’s AI Safety Institute conducts safety evaluations of frontier models on behalf of the government. The compute infrastructure on which those evaluations are run is largely provided by the AI labs themselves — which means the UK is reliant on American companies for its own AI safety testing capabilities. The compute crisis may accelerate pressure on the government to develop independent sovereign compute infrastructure.

What This Means for You

If you use AI tools for work — whether that is Claude, ChatGPT, Gemini, or Copilot — you are relying on infrastructure that is under significant strain. In practical terms, this may manifest as slower response times during peak hours, occasional capacity errors, or gradual price increases on paid plans.

For UK businesses building AI-dependent products, the infrastructure crisis is an argument for architectural flexibility: avoid building in hard dependencies on a single AI provider. Design your systems so that you can switch between Claude, GPT, and Gemini with relatively low effort. The provider that offers the best price-performance ratio in 2026 may not be the leader in 2027.

This article is for educational purposes only.