Inference Hardware That Keeps Up With the Models: Serving Frontier LLMs on Reconfigurable Silicon

Frontier LLMs are getting bigger and changing shape faster than the chips built to run them. The result is inference that costs too much, runs too slow, or both. Saman Naderiparizi CTO of Elastix.ai will walk through why legacy inference accelerators hit a wall with today’s models, and how reconfigurable silicon changes the math. This is a 30-minute technical talk for people who build, deploy or pay for AI infrastructure. What you’ll take away: Where the bottlenecks are: why fixed-function accelerators struggle as model architectures evolve How reconfigurable silicon works: dynamic hardware flexibility that adapts execution to the model at runtime What this means in production: practical approaches to cutting latency and raising throughput at scale Who this is for: ML engineers, infrastructure leads, CTOs and technical founders running or evaluating LLM inference. Speaker: Saman Naderiparizi, Elastix.aiFormat: 30-minute talk with time for questionsTrack: AI Systems & Infrastructure AI Systems & Infrastructure day is made possible with the support of Elastix.ai. Elastix.ai is a Seattle startup building inference infrastructure for frontier LLMs on reconfigurable silicon. Founded by former Apple and Meta machine learning engineers, Elastix runs large models on off-the-shelf FPGA servers that fit a standard air-cooled rack and plug in through a vLLM back end with an OpenAI-compatible API.