NVIDIA exec on AI factory economics: why performance per watt comes first and when to retire old GPUs
kimmonismus · x · 2026-11-03
kimmonismus sits down with NVIDIA's Shruti Koparkar to discuss the economics and software behind AI infrastructure. Key topics: why she tests performance per watt first when power is the binding constraint; what CUDA and CUDA-X actually do; how Pinterest has reused the same GPU fleet across recommendation systems, LLMs and vision-language models; and the decision logic for keeping, repurposing or replacing older GPUs when new models launch. Full video on YouTube.
Related event: NVIDIA Exec on AI Factory Economics: Performance per Watt First(2 posts)→
More from Companies & People
- Agent builder and Leverage Loops author joins Anthropic as Member of Technical Staff — henloitsjoyce · 2026-11-04
- Sentry CEO: engineers who can't do systems design will see salaries cut in half — zeeg · 2026-11-04
- Google's planning AI pilot cuts officer assessment time 58%, 70% of reports accepted unedited — nukebuddy · 2026-11-04
- Factory launches Deployment team, cites Nvidia and Blackstone as users — maddiehfaulkner · 2026-11-04
- Zuck reportedly values Nat Friedman equal to his boss Alexandr Wang — ShakeelHashim · 2026-11-04
- Dean Ball's OpenAI job: "AGI-pilling" the global affairs team — ShakeelHashim · 2026-11-04