Stanford's Ousterhout: AI workloads are outgrowing TCP
AI Engineer · youtube · 2026-09-17
In an AI Engineer talk, Stanford's John Ousterhout argues AI traffic has changed shape: agentic inference has shrunk GPU compute phases to milliseconds, making synchronization metadata exchange — not throughput — the bottleneck. TCP's sender-side congestion control learns about receiver-side queues via 1 bit, several round trips late. Homa, from Behnam Montazeri's Stanford dissertation, inverts this: receiver-driven grants, messages instead of byte streams, short messages overtaking long ones, benchmarked against TCP.
More from Infra
- Seagate IronWolf Pro HDDs doubled in a year, from $550 to $1,100 — HankYeomans · 2026-09-17
- Polymarket puts 31% odds on a US state enacting a data center moratorium by end of 2026 — Polymarket · 2026-09-17
- Data centers now supply roughly 45% of local tax revenue in Loudoun County, Virginia — Polymarket · 2026-09-17
- Google, Nvidia and Anthropic back Emerald AI coalition seeking 100 GW of grid capacity for data centers — TechCrunch AI · 2026-09-17
- Keeping vLLM's prefix cache warm between agent turns — bolts98 · 2026-09-17
- Running Qwen3.8-Flash-Next on 2x5080: should I move from llama.cpp to vLLM? — whatyathinkk · 2026-09-17