Etched

NVIDIA has signaled intent to build transformer-optimized inference chips. Their Blackwell architecture already includes transformer-specific optimizations. With 80%+ market share in AI accelerators, NVIDIA can bundle inference silicon with existing training GPU contracts, making it nearly impossible for Etched to compete on distribution.

Etched has ~45 employees attempting to tape out a custom ASIC. Chip design typically requires 200+ engineers and costs $30M+ per tapeout. A single mask error can cost 6-12 months. The team is talented but thin for the complexity of the problem.

Etched is betting entirely on transformer architecture dominance. If state-space models (Mamba), mixture-of-experts, or other architectures displace transformers, Etched's ASIC becomes a stranded asset. The AI architecture landscape is shifting rapidly.

$120M raised but chip fabrication is extremely capital-intensive. Each tapeout attempt costs tens of millions. Without revenue, the company may need additional capital at a time when semiconductor startup valuations are under pressure.