The Signal — May 7, 2026
The AI infrastructure arms race intensified yesterday as major labs locked down massive compute deals and published new networking standards while a scrappy startup proved you can train frontier-class models without NVIDIA.
Anthropic Strikes SpaceX Colossus-1 Compute Deal, Doubles Claude Code Limits
Anthropic announced yesterday at its developer conference that it has secured access to SpaceX's entire Colossus One data center, a facility with over 300 megawatts of power and 220,000-plus NVIDIA GPUs. The deal is Anthropic's most aggressive move yet to address surging demand for Claude, which has strained capacity for months.
As a direct result of the SpaceX agreement and other recently closed compute partnerships, the company revealed it is doubling rate limits for Claude Code users. The capacity expansion targets the developer tier specifically, where usage has consistently bumped against existing ceilings.
The strange-bedfellows optics are hard to miss: Anthropic, founded by ex-OpenAI safety researchers, is now a major customer of Elon Musk's space company turned AI infrastructure provider. SpaceX's Colossus One facility represents one of the largest single-site GPU concentrations in existence, and Anthropic apparently needed every watt of it. Even well-funded labs with existing cloud partnerships (Anthropic has deep ties to Amazon and Google) still apparently face compute constraints severe enough to send them shopping at unconventional vendors.
Sources: Anthropic · Ars Technica · Axios · WIRED
OpenAI and Partners Launch MRC Protocol for AI Supercomputer Networking
OpenAI, AMD, Broadcom, Intel, Microsoft, and NVIDIA jointly published MRC (Multipath Reliable Connection), a new open networking protocol designed to solve one of the hardest bottlenecks in large-scale AI training: keeping 100,000-plus GPUs communicating efficiently without exotic proprietary hardware.
MRC enables multi-plane high-speed networks using only two tiers of standard Ethernet switches, a sharp simplification over current approaches that often require expensive InfiniBand fabric or custom silicon. The protocol is already running in production at OpenAI's Oracle Cloud Infrastructure site in Abilene, Texas, and at Microsoft's Fairwater supercomputers.
The consortium released MRC through the Open Compute Project, making it freely available for anyone building large training clusters. These companies compete fiercely on chips, models, cloud services, and training infrastructure, but evidently agreed that networking was better treated as shared plumbing than a competitive moat. For smaller labs and sovereign AI efforts that lack access to proprietary interconnects, MRC could lower the barrier to building GPU supercomputers at scale.
Sources: OpenAI · The Decoder · AMD
Zyphra Releases ZAYA1-8B: First Large-Scale Reasoning MoE Trained on AMD Hardware
Zyphra released ZAYA1-8B, a reasoning-focused Mixture-of-Experts model trained entirely on AMD MI300X GPUs running on IBM Cloud infrastructure. According to Zyphra's benchmarks, the model delivers competitive reasoning and math performance, along with strong coding benchmarks at its 8-billion parameter scale. The company emphasizes what it calls "intelligence density per parameter."
More interesting than the model itself is what trained it. Zyphra claims this is the first large-scale MoE trained end-to-end on an integrated AMD hardware, software, and networking stack. If the results hold up to independent evaluation, it's a meaningful crack in NVIDIA's near-total monopoly on frontier model training. AMD has shipped competitive inference hardware for some time, but training workloads have remained stubbornly CUDA-locked.
The model weights are available on Hugging Face under an open license. Zyphra is positioning the release as proof that AMD's training stack has matured enough for serious work, not just benchmarking exercises. The broader research community will stress-test those claims soon enough, but a credible non-NVIDIA training run at this scale is news regardless.
Sources: Zyphra · PR Newswire · MarkTechPost · Hugging Face
On the Editor's Desk
Held today: Anthropic published interesting midtraining research on model specs, but covering two Anthropic stories would skew the edition. The EU AI Act Brussels enforcement talks that reportedly collapsed were stale (the underlying meetings were April 28-29). Reuters ran a piece on Chinese firms scrambling for Huawei AI chips post-DeepSeek V4, but it relied entirely on unnamed sources. Similarly, reports of US-China AI security talks and DeepMind unionization efforts both lacked primary confirmation. All remain on the watchlist if corroborated.
