The Intake
The Intake — Thursday, May 7, 2026
Editor’s note
Anthropic named its compute supplier and the operational consequence in a single release yesterday: 300MW from SpaceX's Colossus 1, and Claude Code's five-hour limit doubled with peak-hour caps removed. The causal link is explicit in the announcement — the deal enables the headroom.
The Code with Claude 2026 conference ran alongside, producing a second layer of product news: multi-agent orchestration moved to public beta, and Dreaming — a cross-session self-inspection feature — was introduced. Routines, the async-agent-automation feature shipped April 14, ran as a conference keynote topic.
A brief model note: Zyphra released ZAYA1-8B yesterday — the first publicly documented large-scale mixture-of-experts model trained exclusively on AMD MI300X hardware. Worth a read for anyone tracking the non-NVIDIA training substrate.
On the substrate
Anthropic secures 300MW from SpaceX Colossus; Claude Code limits double and peak-hour caps removed
Anthropic Simon Willison liveblog Anthropic Routines
Anthropic secured more than 300MW of new capacity from a SpaceX partnership at the Colossus 1 data center — over 220,000 NVIDIA GPUs — and announced the operational consequence in the same release: Claude Code's five-hour rate limit doubles for Pro, Max, Team, and seat-based Enterprise plans effective May 6, and peak-hour limit reductions are removed for Pro and Max accounts.
Most compute partnerships of this size are announced separately from their downstream product effects, if they are announced at all. Naming both in the same release is the notable structural choice — readers can trace the causal link explicitly: this much new capacity, this much new headroom for these tiers.
The Code with Claude 2026 conference ran May 6 alongside the deal disclosure and produced two announcements that change what agents can do on the next turn. Multi-agent orchestration moved to public beta: agents can spawn, coordinate, and receive results from subagents through a structured API without manual prompt-chaining. A cross-session self-inspection loop, Dreaming, has the agent review its own prior outputs across a configured window and surface drift, gaps, or contradictions. Routines, the async-automation feature already shipping since April 14, was a third keynote topic — the conference covered it rather than announced it.
For operators running or evaluating async agent workflows, the Routines documentation is the operator-grade reference. The trigger-configuration spec, the permission model for spawned subagents, and the session-boundary behavior under Dreaming are where implementation decisions will diverge from the conference-demo framing. Simon Willison's liveblog captured the afternoon session in real time.
ZAYA1-8B: Zyphra's MoE model trained on AMD MI300X
Zyphra released ZAYA1-8B on May 6 — a mixture-of-experts model trained exclusively on AMD MI300X hardware, and the first publicly documented large-scale MoE to run that training pipeline on AMD rather than NVIDIA.
Weights are available on Hugging Face; Zyphra's post carries the training methodology. ZAYA1-8B demonstrates that AMD MI300X clusters are now producing MoE architectures at a scale previously confined to NVIDIA H100 and H200 deployments. Zyphra's post does not include head-to-head benchmarks against NVIDIA-trained MoE models at this scale — that is the benchmark question to watch once independent eval coverage arrives. Single-source vendor item; treat performance framing accordingly.
For operators
Claude Code rate-limit doubling: re-check fallback logic before assuming it's quiet
Anthropic announcement Anthropic limits page
Claude Code's five-hour rate limit doubled for Pro, Max, Team, and seat-based Enterprise plans on May 6, and peak-hour limit reductions were removed for Pro and Max. The decision this prompts for personal-agent operators: review whether existing workflow configurations were shaped around the prior constraints and no longer need to be.
If your Claude Code configuration includes explicit rate-limit fallback logic — retry schedules, model downgrade paths triggered by capacity errors — test whether that logic fires correctly under the new limits before assuming it no longer triggers. Silent fallback logic that was load-bearing under the old caps may now be running rarely enough that drift goes unnoticed until the next time it does fire. The limits page carries the current tier breakdown; the prior limits are not archived there, so document your own baseline now if you intend to compare against pre-May 6 behavior.
Multi-agent orchestration beta access does not come with the limit increase automatically — it is a separate enrollment.