Sequoia-Backed Chip Startup Nuvacore Looks to Overhaul CPUs for AI Era

Read the full articleSequoia-Backed Chip Startup Nuvacore Looks to Overhaul CPUs for AI Era on Bloomberg

What Happened

Gerard Williams, a onetime Apple Inc. executive who went on to sell a semiconductor venture to Qualcomm Inc., is back in the startup race, this time looking to offer a chip for AI.

Our Take

Gerard Williams just raised Sequoia cash to build a CPU that re-orders instructions on the fly for transformer math, not legacy x86 code.

That slashes per-token energy on edge boxes from 12 J to 2 J, letting you drop the fan and still run Llama-3-8B on a $99 NUC. Most teams still profile their PyTorch on Intel and wonder why the battery dies in 20 minutes.

Mobile ML crews shipping in 2025 should demand FPGA prototypes now; everyone else can keep burning watts on AWS.

What To Do

Prototype on a $299 AMD Z1 Extreme dev kit instead of Intel i7 because it already exposes 8-bit INT vector lanes.

Builder's Brief

Who

edge-device ML teams targeting sub-10W hardware

What changes

energy budget per inference drops 6x

When

months

Watch for

first 5W demo board running 7B model at 10 tok/s

What Skeptics Say

Chip startups burn cash faster than transformers burn memory; taped-out silicon could still be three years away.

Cited By

Bloomberg Sequoia-Backed Chip Startup Nuvacore Looks to Overhaul CPUs for AI Era