Usable AI capacity emerged as a delivery chain extending beyond GPU totals. Capital, cluster networking, and customer quotas moved together; voice split across transport and specialist models; high-authority agents made identity part of execution control. New benchmarks showed why task stages and evaluator disagreement have to remain visible.
1. Compute supply becomes usable through capital, fabric, and allocation
Cerebras opened an IPO roadshow for 28 million shares at $115 to $125, implying up to $3.5 billion in gross proceeds. OpenAI published MRC after deploying it in its largest Nvidia GB200 supercomputers; the protocol supports two-tier fabrics above 100,000 GPUs and routes around failures in microseconds. Anthropic then tied 220,000 GPUs at SpaceX's Colossus 1 directly to doubled Claude Code limits and higher API rate limits.
The three developments price separate conversions in the compute chain. The IPO prices capital for a specialized hardware supplier, network design turns installed accelerators into synchronized training capacity, and service quotas expose how much capacity reaches customers. GPU totals omit both utilization and allocation; Anthropic's limit increase is unusually concrete evidence that newly contracted hardware changed the product surface.
Sources: Cerebras's proposed initial public offering · OpenAI's MRC networking design · Anthropic's SpaceX compute and usage-limit announcement
2. Voice AI splits into routing, session state, and specialist models
OpenAI's WebRTC architecture sends opaque UDP packets through a lightweight relay while stateful transceivers retain ICE, DTLS, SRTP, and session ownership. Three days later, it released GPT-Realtime-2 for reasoning and tools, GPT-Realtime-Translate for speech from more than 70 input languages into 13 outputs, and GPT-Realtime-Whisper for streaming transcription.
The two releases decompose one voice interaction twice. Infrastructure can scale packet ingress independently from cryptographic session state, while applications can purchase reasoning, translation, or transcription under separate token and per-minute meters. Failure attribution becomes sharper: transport jitter occurs before inference, translation quality belongs to a specialist model, and a tool error can arise after both layers performed correctly.
Sources: OpenAI's low-latency voice infrastructure · OpenAI's new Realtime API models
3. High-authority agents stack identity and action-review boundaries
Perplexity's Personal Computer can operate across local files, native Mac applications, the web, and more than 400 connectors, with tasks initiated and approvals issued from an iPhone. OpenAI's GPT-5.5-Cyber preview grants more permissive behavior only to verified defenders. Codex Auto-review separately sends sandbox-crossing requests to a GPT-5.4 Thinking reviewer, which approved 99.1% of escalations in internal traffic.
The controls sit at three points: actor verification before model access, human approval during a remote task, and model review at the sandbox boundary. Identity establishes who is accountable; environmental policy defines which actions require review; the reviewer decides whether one crossing proceeds. OpenAI's warning that Auto-review can be misled prevents its approval rate from serving as a containment score.
Sources: Perplexity's Personal Computer for Mac · OpenAI's Trusted Access for GPT-5.5-Cyber · OpenAI's Codex Auto-review research
4. Stage-specific benchmarks expose where aggregate rankings fail
MathNet separates Olympiad problem solving, equivalent-problem retrieval, and retrieval-augmented solving across 30,676 expert-authored problems in 17 languages; every tested embedding model remained below 5% Recall@1. Contra Labs split 93 creative prompts across ideation, mockup, and refinement, and found that no model led all three phases in any of five domains.
The comparison distinguishes two reasons a leaderboard winner can mislead. MathNet locates a technical handoff where strong solving coexists with weak retrieval. Contra locates an evaluation handoff where shared standards give way to legitimate differences in taste. A useful rank therefore belongs to a named stage and scoring population; collapsing either dimension discards the mechanism behind the score.
Sources: MIT's MathNet benchmark · Contra Labs' Human Creativity Benchmark