xAI Leases Colossus 1 to Anthropic: The Masterstroke That Turns a Problematic Cluster into a Golden Goose

Por Juan Carlos Carvallo – Analista Financiero Blog: https://www.juancarloscarvallo.com/blog 9 de mayo de 2026

On May 6, 2026, SpaceXAI (the entity that now integrates xAI following the merger) announced a landmark deal: Anthropic will lease the entire capacity of Colossus 1, one of the world’s largest AI supercomputers, featuring over 220,000 NVIDIA GPUs (a mix of H100, H200, and GB200) and more than 300 MW of power.

At first glance, many interpreted this as Elon Musk surrendering to a competitor. The reality is quite different: this is one of the smartest and most pragmatic moves we have seen in the AI race. It’s no longer just about who has the “smartest” model — it’s about who survives the brutal economics of compute.

What is Colossus 1 and Why Lease It?

Colossus 1 was built at record speed in Memphis to train Grok, xAI’s AI model. However, its heterogeneous architecture — combining three different generations of GPUs (H100, H200, and GB200) — has created significant challenges for large-scale distributed training:

• Straggler effect: Slower GPUs hold back the entire cluster, reducing effective FLOPs utilization to just 11%, far below the 40%+ achieved by Meta or Google.

• Network topology issues and “power smoothing” problems with the Blackwell GPUs further complicate the stack.

• Result: Inefficient for training frontier models.

For this reason, xAI has already migrated its training workloads to Colossus 2, a fully homogeneous Blackwell-based cluster that delivers superior efficiency. Colossus 1, on the other hand, is ideal for inference workloads — exactly what Anthropic needs for Claude Pro, Max, Code, and enterprise services. Inference requires simple parallelization, benefits from a single tenant (eliminating jitter), and does not demand perfect synchronization.

The Economic Importance of the Lease

The leasing agreement between xAI and Anthropic is estimated to generate $5 to $6 billion in annual recurring revenue. This figure nearly offsets the operational and construction costs of Colossus 1, transforming a problematic cluster into a solid, high-margin cash-flow generator.

On the other side, Anthropic is reporting a run-rate revenue exceeding $30 billion (up from $9B at the end of 2025), driven by Claude Code and more than 1,000 enterprise clients paying over $1 million per year each. The deal has already enabled increased usage limits on Claude Code and the API, accelerating growth even further.

This move proves that in the AI race, having the most advanced model is not enough — how you manage compute infrastructure is equally critical. The brutal economics of compute are forcing companies to seek pragmatic solutions to maximize return on investment.

How Colossus 1 Becomes the Golden Goose

Although Colossus 1 is not optimal for training new models, its inference capabilities make it extremely valuable. Inference requires less synchronization and can take full advantage of heterogeneous hardware without efficiency losses. Anthropic can now run multiple instances of its Claude models without the bottlenecks that plagued training.

Additionally, having a single tenant in full control of the cluster eliminates jitter — a common issue in multi-tenant environments that affects latency and stability. This turns Colossus 1 into a highly reliable resource for enterprise services demanding high availability and consistent performance.

Looking ahead to SpaceXAI’s imminent IPO (target valuation between $1.75 and $2 trillion, with roadshow expected in June), this deal is pure gold: it shifts the narrative from “AGI cash-burning lab” to “neo-cloud” player in the style of AWS that efficiently monetizes infrastructure.

Impact on Competition and the AI Market

This transaction is reshaping the AI market in several ways:

• For xAI/SpaceXAI: Predictable recurring revenue, full focus on Colossus 2 for future Grok models, and a “picks and shovels” business model that funds the mission to understand the universe. Elon continues suing OpenAI while collecting rent from Anthropic — pure pragmatism.

• For Anthropic: Immediate access to hundreds of thousands of GPUs without massive capex. It solves compute bottlenecks, boosts capacity for subscribers and enterprise clients, and opens the door to joint gigawatt-scale orbital compute projects.

• Pressure on other competitors: Companies like Meta, Google, and OpenAI must rethink their infrastructure strategies to compete with business models that combine ownership with smart leasing.

Such agreements could mark a new trend in the industry, where collaboration and shared resources become as important as model innovation.

Key Lessons for the AI Industry

The Colossus 1 story offers several important takeaways:

• Not all clusters are created equal: Architecture and hardware homogeneity directly impact efficiency and the types of workloads they can handle.

• Compute economics are crucial: Beyond raw power, cost and effective utilization determine project viability.

• Flexibility and pragmatism win: Adapting infrastructure to real needs (e.g., using a cluster for inference instead of training) can turn a problematic asset into a revenue generator.

• Collaboration between companies: Sharing resources can accelerate technology development and deployment for everyone’s benefit.

  
  

The Future of “AI for Rent”

The xAI-Anthropic agreement may be just the beginning of a new era. As models grow in size and complexity, infrastructure demand will continue rising, but so will the importance of cost and efficiency.

We are likely to see more specialized cluster leasing contracts, allowing companies with excess capacity to monetize their assets while others gain access to cutting-edge technology without massive upfront investments. This democratizes high-performance AI and fosters a more dynamic, collaborative ecosystem.

Conclusion: Business Maturity in the AI Era

This operation demonstrates first-principles thinking: don’t fall in love with the hardware — use it in the most efficient way possible. The AI race will no longer be won solely by the one who builds the biggest cluster, but by the one who doesn’t go bankrupt paying the electricity bill.

xAI/SpaceXAI is not retreating. It is positioning itself as critical infrastructure that funds its own ambition. And in doing so, it is teaching the industry a valuable lesson: in AI, economics matter as much as intelligence.

What do you think? Do you believe we will see more compute leasing agreements, or will the major labs continue burning capital on their own clusters? Would you like a deeper pre-IPO analysis of SpaceXAI or Anthropic as an investment? Leave your comments below.

Disclaimer: This article is based on publicly available information as of May 9, 2026. It is not investment advice. Always conduct your own due diligence.