Photo: ed br / Pexels
San Francisco startup Wafer claims its AI-driven optimization gets AMD's MI355X to 80% of Nvidia B200 throughput at less than half the cost
Emilio Andere has a simple thesis: Nvidia’s dominance isn’t really about hardware. It’s about software. And software, unlike silicon, can be rewritten.
The Wafer AI co-founder and CEO argues that AMD GPUs can match Nvidia’s performance when the code running on them is properly optimized.
The numbers behind the claim #
In July 2026, Wafer fine-tuned Z.AI’s GLM-5.2 model to run on AMD’s MI355X GPU. The result: approximately 80% of the throughput delivered by Nvidia’s B200, at less than half the cost.
Wafer’s broader pitch is that it can deliver 2 to 2.8x speedups over stock baselines across various open-source models. The startup builds AI agents that function as autonomous performance engineers, automatically optimizing GPU workloads for better throughput.
Average GPU utilization in production environments sits around 20%. That means organizations are effectively burning 80% of the compute they’re paying for.
From seed to Series A in five months #
On September 1, 2026, Wafer closed a $40 million Series A funding round co-led by Marathon and Chemistry. The deal valued the startup at over $200 million, a steep climb from its $4 million seed round completed just five months earlier in April 2026.
Backers include Google’s Jeff Dean and OpenAI’s Wojciech Zaremba. Andere and his co-founder Steven Arellano are both University of Chicago alumni. Their company is based in San Francisco.
Why Nvidia’s moat might be shallower than it looks #
Nvidia’s competitive advantage has never been purely about transistors. The real fortress is CUDA, the proprietary software ecosystem that makes developing for Nvidia GPUs dramatically easier than the alternatives. Virtually every major AI framework, every training pipeline, every inference engine has been built with CUDA in mind.
AMD’s answer is ROCm, its open-source GPU computing stack. AMD’s ROCm stack has undergone rapid iteration, with the company claiming 3.3x inference performance gains compared to previous versions.
Wafer’s approach adds another layer to this. Rather than waiting for AMD to close the software gap organically, the startup uses AI agents to automatically generate optimized code paths for AMD hardware.
What this means for the GPU market #
If Wafer’s optimizations hold up at scale, AMD has competitive hardware in the MI355X. But “competitive hardware” without competitive software has historically meant very little in the GPU compute market. The price-performance argument is particularly compelling for inference workloads. If AMD plus Wafer optimization delivers 80% of the performance at less than half the price, the total cost of ownership calculation shifts dramatically in AMD’s favor for many production use cases.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our