When AI Rivals Become Each Other's Landlords

This week xAI started selling Anthropic GPU time and OpenAI shipped a networking protocol with NVIDIA, AMD, Intel, Broadcom, and Microsoft. The vendor walls are visibly cracking.

This week's vendor moves broke the tidy 'AI camp' narrative the industry has been telling itself since 2024. xAI agreed to give Anthropic access to its Colossus 1 supercomputer, while OpenAI shipped a networking protocol co-developed with five chip and cloud competitors. Whatever else they are, frontier model providers are now functionally landlords and tenants of one another's infrastructure.

The xAI/Anthropic deal is the more surprising of the two. xAI announced on May 6, 2026 that it will provide Anthropic with capacity on Colossus 1 — which xAI calls one of the world's largest AI supercomputers — with both companies signaling interest in 'multiple gigawatts of orbital AI compute capacity' down the line. Two of the highest-profile rivals in foundation models are now openly in a paid infrastructure relationship.

The OpenAI announcement is structurally similar but framed as engineering. OpenAI released Multipath Reliable Connection (MRC) on May 6, 2026, an open networking protocol developed with AMD, Broadcom, Intel, Microsoft, and NVIDIA, designed to keep large training clusters in sync. OpenAI says MRC is already deployed across its largest sites including its Abilene, Texas Oracle Cloud Infrastructure facility and Microsoft's Fairwater supercomputers. That a single protocol now spans Microsoft (OpenAI's biggest backer) and NVIDIA (which sells GPUs to every rival) tells you the inputs to frontier training are getting expensive enough that nobody can afford private versions of them anymore.

Meanwhile the product layer keeps moving. OpenAI began rolling out GPT-5.5 Instant on May 5, 2026 as the new ChatGPT default, with claimed factuality gains in medicine, law, and finance — claims worth verifying before betting a workflow on them. Google DeepMind shipped a year-in-review for its Gemini-powered coding agent AlphaEvolve. Anthropic ran its Code with Claude developer event in San Francisco and unveiled the Anthropic Institute on May 7, 2026, a research arm focused on economic diffusion, threats and resilience, AI in the wild, and AI-driven R&D. Each release nominally targets a different audience, but all four vendors are now drawing on a shared substrate of chips, fabric, and (now) leased capacity.

For builders, the practical implication is that compute geography matters less than it did six months ago — your model choice no longer dictates which clouds you can deploy on. For buyers negotiating procurement, the lock-in story is weaker: if Anthropic is renting from xAI and OpenAI is shipping protocols with NVIDIA and Intel, the captive-ecosystem pitch from any single vendor needs harder questions.

The bottom line: Stop drawing the 2026 vendor map by which lab made which model — draw it by who controls compute, fabric, and energy, because that's where the real partnerships are forming.

Sources