Positron's AI chips challenge NVIDIA's dominance in the data center
Positron, a Reno-based startup, just raised $23.5 million to scale up its US-made AI chips.
Their new Atlas server is built for heavy-duty AI tasks and is already shipping.
The big deal? It's faster and more energy-efficient than NVIDIA's DGX H100 and H200 systems.
Atlas offers better performance per watt than NVIDIA's DGX H200
Atlas delivers 4.54x better performance per watt and 3.08x better value for your money compared to NVIDIA's DGX H200.
In real numbers: Atlas hits 280 tokens/second per user at 2000W, while the H200 manages only 182 tokens at a much higher 5900W.
It is compatible with all Hugging Face transformer models
Thanks to its memory-smart design, Atlas uses over 93% of its memory bandwidth (way above the typical GPU's 10-30%), making it about 70% faster and using two-thirds less power than NVIDIA's top chips.
Plus, it works with all Hugging Face transformer models and offers an OpenAI-compatible API—so plugging it into existing workflows is easy.
Atlas cuts data center costs by half
Atlas doesn't just save on energy—it also slashes data center CapEx by 50% compared to NVIDIA's H100/H200 systems.
For businesses running lots of AI workloads, that means smarter spending without sacrificing performance or sustainability.