미명의 스타트업이 자체 개발한 실리콘이 엔비디아의 성능 지표 대비 80% 수준을 절반의 단가로 달성했다고 주장했습니다.

Nvidia Corp.’s (NASDAQ: NVDA) dominance in AI computing has long rested on more than powerful GPUs; its software ecosystem helps developers extract maximum performance from those chips. Wafer is betting that advantage can be challenged from another direction: using AI to optimize models for rival hardware.

According to The Information, the one-year-old startup recently raised $40 million in Series A funding at a valuation above $200 million. Its founders state that Wafer has also received multiple acquisition offers from larger inference and cloud providers.

Operating across both Nvidia and Advanced Micro Devices, Inc. (NASDAQ: AMD) architectures, Wafer runs inference workloads and is building AI agents that optimize models for specific tasks. In July, the company announced that it tuned Z.AI’s GLM-5.2 for AMD’s MI355X, reaching approximately 80% of Nvidia B200 throughput at less than half the cost in internal testing.

This forms the unusual core of Wafer’s pitch: the alternative to Nvidia may not require a superior chip if software can make existing alternatives perform competitively.

The evidence remains preliminary, and Wafer is still dwarfed by Nvidia. Yet its funding, acquisition interest, and expanding work with alternative accelerators suggest clear commercial value in an AI infrastructure layer designed to reduce industry dependence on a single chip ecosystem.

The central question is whether inference optimization can become a meaningful counterweight to Nvidia’s software advantage. Wafer remains a small-scale experiment, but it offers an early glimpse of a potentially significant shift: competition with Nvidia may increasingly occur in software, not just silicon.