Etched President on Burning Transformers into Silicon & Hitting the Power Wall
AI Engineer · youtube · 2026-07-31
At the AI Engineer World's Fair, swyx sat down with Rob Wachen, Co-founder and President of Etched, to discuss the inference inflection from first principles. Etched's core strategy is to bypass general GPUs and burn transformers directly into custom silicon, designed specifically to run today's large MoE models at high throughput.
Key Insights & Engineering Realities:
- Vertical Integration: The chip is just the starting point; the entire system, from board to rack, must be reinvented to serve tokens at scale.
- Power is the True Wall: Instead of raw flops, the team obsesses over active cores, clock speed, voltage, and advanced packaging, tackling yield, testing, and MTBF.
- Production is the Product: The culture emphasizes real hardware over slideware, featuring an in-office mini data center and a Taiwan/US night-shift integration pipeline.
- Memory Roadmap: Future scalability aims toward wafer scale and connecting 10,000+ chips in a single scale-up domain.
- The $4T Gap: Wachen highlights a massive efficiency gap between Nvidia and the rest, closing with a strong call to hire talent across production and infrastructure.
More from Infra
- Nvidia's Theoretical 2035 Valuation Could Exceed Total Human History Value — ns123abc · 2026-07-31
- OpenAI's Promise of Cerebras 750 TPS Inference Faces July Deadline — imjustnewatai · 2026-07-31
- Compute Costs Set to 10x as AI Agents Fill Data Centers, Says Lambda Co-founder — TheZachMueller · 2026-07-31
- Amazon Signals Datacenter and Compute Investment Won't Slow Down — BenBajarin · 2026-07-31
- Running Kimi K3 locally requires 1.56TB VRAM, limited to top-tier chips — ArtificialAnlys · 2026-07-31
- Vinci Unveils AI Physics Engine: Predicts Hardware Thermo-Mechanical Warpage at Billion-DOF Scale — AnneliesGamble · 2026-07-31