Anthropic Borrowed xAI's Compute. That's the Real Story This Week.

Code with Claude grabbed the headlines, but the bigger vendor news was Anthropic accepting compute from xAI's Colossus 1 — a deal that quietly redraws the AI supply chain.

The week of May 5-7, 2026 looked like Anthropic's coronation: a sold-out developer conference, a teased Claude Mythos Preview, a public HackerOne bug bounty, and the launch of The Anthropic Institute. Then xAI announced it was lending Anthropic the Colossus 1 supercomputer — a deal that quietly tells us more about the next 18 months than any model release.

The Anthropic week

At Code with Claude in San Francisco, Anthropic stacked announcements. It opened its security bug bounty publicly on HackerOne, unveiled the research agenda for The Anthropic Institute covering economic diffusion, threats and resilience, AI systems in the wild, and AI-driven R&D, and published research on Natural Language Autoencoders — interpretability tools that translate model activations into readable text. The company also donated its Petri alignment tool to Meridian Labs for independent stewardship. The throughline: Anthropic is positioning itself as the safety-and-research vendor, not just the coding vendor.

The compute twist

The surprise came from xAI. SpaceXAI announced it would provide Anthropic with access to Colossus 1, with both companies expressing interest in multiple gigawatts of orbital AI compute capacity. Simon Willison flagged the under-reported details: xAI keeps the larger Colossus 2, Colossus 1 has a poor environmental record, and Elon Musk reserved the right to "reclaim the compute if their AI engages in actions that harm humanity." That last clause is novel — a compute lender with a moral kill switch written into the contract.

Everyone else moved too

OpenAI shipped GPT-Realtime-2 in the API, pitched as GPT-5-class reasoning for voice agents (as of May 2026), and a Codex Chrome extension that runs in parallel across browser tabs. Google promoted Gemini 3.1 Flash-Lite from preview to general availability at the same pricing. And, per a third-party benchmark, Alibaba's Qwen 3.6 Max Preview reportedly outperforms Claude 4.5 Opus on agentic coding — a claim worth verifying, but not dismissing.

For builders, the agentic-coding race is now a four-way contest, and your IDE will likely route to whichever model is cheapest per task. For buyers, the lesson is harder: model quality is converging, but compute supply chains are diverging. Where your model trains may matter as much as how it benchmarks.

The bottom line: Vendor due diligence in 2026 means asking not just "how good is the model" but "whose data center is it running on, and who can pull the plug."

Sources