The model was only one component in each of today's strongest stories. OpenAI split voice routing from session state and opened its GPU transport protocol; MathNet separated mathematical retrieval from solving; and Cerebras offered public investors a direct, capital-intensive alternative to conventional GPU infrastructure.
1. OpenAI separates voice packet routing from session state
OpenAI described a WebRTC design that separates a lightweight UDP relay from stateful transceivers. The relay derives routing from ICE credentials and forwards opaque packets, while each transceiver retains ICE, DTLS, SRTP, and session state for ChatGPT voice and Realtime API connections.
The split replaces broad public UDP port ranges with a small relay surface and lets Kubernetes scale stateful workers independently. OpenAI provides a production account rather than a comparative latency study, but the architecture cleanly isolates packet steering from the expensive cryptographic and conversational state of a one-to-one voice session.
Sources: OpenAI on low-latency voice infrastructure
2. MathNet exposes retrieval as a separate mathematical bottleneck
MathNet released 30,676 expert-authored competition problems and solutions from 47 countries in 17 languages. Its benchmark covers direct solving, retrieval of mathematically equivalent problems, and retrieval-augmented solving; every tested embedding model scored below 5% Recall@1.
Human verification and multilingual breadth reduce two common benchmark shortcuts, while prompts, graders, and the held-out split still define the rankings. The retrieval collapse is the notable result: a model can solve a supplied Olympiad problem yet fail to locate a structurally equivalent one, a gap hidden by aggregate reasoning scores.
Sources: MIT's MathNet benchmark
3. Cerebras seeks $3.5 billion in a public-market test for AI chips
Cerebras proposed selling 28 million shares at $115 to $125, raising $3.5 billion and implying a $26.6 billion market capitalization at the top of the range. TechCrunch reported that indicated orders substantially exceeded the shares offered before pricing.
The offering was open, so demand could still alter pricing or execution. A completed listing would create a public valuation and disclosure regime for specialized inference hardware, giving investors and customers a comparison point for capital requirements, revenue concentration, and margins outside Nvidia's GPU ecosystem.
Sources: TechCrunch on Cerebras's order book · Cerebras's offering announcement
4. OpenAI opens the transport protocol inside its largest GPU clusters
OpenAI published Multipath Reliable Connection through the Open Compute Project after developing it with AMD, Broadcom, Intel, Microsoft, and Nvidia. MRC spreads transfers across many paths, uses SRv6 source routing, and is deployed in OpenAI's largest Nvidia GB200 supercomputers.
OpenAI reports two-tier fabrics above 100,000 GPUs and recovery from path failures on microsecond timescales. Those production claims remain the vendor's, but an open specification changes the evidence available: switch and NIC suppliers can implement the protocol, and cluster operators can test congestion, failover, and collective performance without reproducing OpenAI's entire stack.