AI weekly digest: Capacity crunch, mega launches, and an industry rewiring itself
A week dominated by infrastructure: Anthropic leased compute from xAI to fix a capacity crunch, OpenAI weighed suing Apple, Cerebras popped 107% on IPO, Nvidia hit $40B in AI bets, Thinking Machines dropped a 276B realtime model, and Opus 4.7 Fast landed in Claude Code.
The week's center of gravity was infrastructure, not models. Anthropic struck a compute deal with xAI to claw back capacity, Cerebras IPO'd into a 107% first-day pop, and Nvidia's $40B in investments kept the AI capex flywheel spinning. Models still showed up — Opus 4.7 Fast, Thinking Machines' 276B-A12B realtime model, and another wave of Anthropic interpretability work — but the more interesting signal is that the buildout is now visibly straining the labs that started it.
1. Anthropic leases compute from xAI to relieve developer capacity crunch
After weeks of throttled rate limits, banned accounts and angry developers, Anthropic confirmed it has reached an agreement with xAI to lease data center space. The deal lets Anthropic restore Claude Pro access and raise API rate limits while it catches up to demand that ran ahead of its capacity plans.
Source: The Pulse: Did capacity shortages turn Anthropic hostile to devs?
2. Apple–OpenAI partnership frays; OpenAI exploring legal action
OpenAI has retained outside counsel and is preparing possible legal action against Apple, claiming Apple never made an honest effort to integrate ChatGPT into its app ecosystem after their two-year-old deal. Apple is set to open its platform to rival providers later this year, raising the stakes for any unwind.
Source: TLDR — Apple vs OpenAI
3. Claude Opus 4.7 Fast enters research preview
Anthropic shipped a Fast variant of Opus 4.7 across the API, Claude Code, Cursor, Emergent, Factory, v0 and Warp. It targets the same quality bar with materially lower latency — the practical effect for coding agents is shorter inner-loop times, especially on long edit-build-test cycles.
Source: TLDR — Opus 4.7 Fast
4. Thinking Machines drops a 276B-A12B realtime interaction model
Thinking Machines Lab released a research preview of its Interaction Models — a 276B parameter MoE with 12B active per token, designed for native real-time human–AI collaboration across audio, video and text. Reported as a SOTA jump on realtime voice/video benchmarks; weights are limited preview.
Source: AINews — TML-Interaction-Small 276B-A12B
5. Anthropic's Mythos used to find the first public macOS M5 kernel exploit
Researchers used an early version of Anthropic's Mythos model to develop the first public memory-corruption exploit on Apple's M5 silicon, bypassing Memory Integrity Enforcement — the multi-billion-dollar mitigation Apple shipped to make exactly these attacks obsolete. The chain still required human expertise; Mythos compressed the discovery loop to about five days.
Source: Calif.io — First public M5 kernel exploit
6. Cerebras IPO pops 107% on first trading day
Cerebras Systems opened 75% above its IPO price and at one point traded up 157% before settling at +107%. The reaction is being read less as another dotcom-style frenzy and more as Wall Street finally pricing in inference demand as a durable, capital-hungry workload distinct from training.
Source: Exponential View — Cerebras and the IPO pop
7. Nvidia commits $40B in AI investments year-to-date
Nvidia disclosed it has now made over $40B in AI-related commitments in 2026 alone — funding model labs, infrastructure plays and downstream startups whose growth keeps GPU demand intact. The pattern is a self-reinforcing capital loop that policy and antitrust watchers are starting to flag.
Source: TLDR — Nvidia invests $40B
8. Figure AI's Helix-02 humanoids run autonomous 8-hour factory shifts
Figure says its Helix-02 system now lets humanoid robots sustain full eight-hour factory shifts without human intervention or resets. The unified network combines vision, touch, proprioception and whole-body control in a single model and supports multi-robot coordination on long-horizon tasks.
Source: Interesting Engineering — Figure Helix-02
9. Anthropic publishes interpretability papers on Natural Language Autoencoders
Anthropic released new interpretability work on Natural Language Autoencoders — methods for projecting model internals back into readable English. The Sequence flagged it as a meaningful step beyond circuit-tracing: instead of mapping features to neurons, you get sentence-level descriptions of what a model is "thinking" at a given step.
Source: The Sequence — Reading Claude's Mind in English
10. /goal becomes an agent primitive — now in Claude Code and Codex CLI
Both Claude Code and OpenAI's Codex CLI shipped a /goal command this week. The pattern is the same: you write down what "done" looks like and the agent works toward it, checking its own state, until that goal is met. Expect this to become a standard primitive in coding agents alongside /plan and /review.