AI digest: Infrastructure week
OpenAI drops a new networking protocol whilst inference engines and benchmarking tools get serious upgrades.
This week brought some proper infrastructure announcements alongside the usual product launches.
OpenAI builds its own networking protocol
OpenAI released MRC (Multipath Reliable Connection), a networking protocol designed for massive GPU clusters. The protocol spreads packets across hundreds of paths and recovers from network failures in microseconds, enabling clusters with over 100,000 GPUs using just two tiers of Ethernet switches. When you’re building at OpenAI’s scale, apparently you need to reinvent the network stack too.
TokenSpeed targets inference bottlenecks
The LightSeek Foundation launched TokenSpeed, an open-source inference engine aimed at matching TensorRT-LLM performance for agentic workloads. As coding assistants like Claude Code and Cursor move from developer tools to core infrastructure, inference efficiency is becoming the real bottleneck. This feels like the right problem to solve, though we’ll see if it actually delivers the performance claims.
Meta releases massive EEG benchmark
Meta AI dropped NeuralBench, a benchmarking framework for NeuroAI models covering 36 EEG tasks across 94 datasets and 13,603 hours of brain recordings. The standardised interface across 14 different architectures should make comparing brain-computer interface models much easier. Proper benchmarking infrastructure like this often signals a field is getting serious about reproducible progress.
OpenAI adds voice features to API
OpenAI launched new voice intelligence features in their API, targeting customer service and education applications. The timing feels deliberate, coming right after their networking protocol announcement. They’re clearly positioning for enterprise scale across both infrastructure and user-facing features.