Positron AI, a U.S.-based semiconductor company specializing in inference-optimized hardware for generative AI, has secured $51.6 million in an oversubscribed Series A funding round, bringing its total capital raised this year to more than $75 million. The round was led by Valor Equity Partners, Atreides Management, and DFJ Growth, with participation from Flume Ventures, Resilience Reserve, 1517 Fund, and Unless.
The funding will accelerate deployment of Positron’s flagship product, Atlas, and support development of its second-generation inference system, Titan, expected in 2026. Atlas is currently being used in production environments and is designed to deliver higher energy efficiency and lower cost per inference compared to current market leaders.
With global AI infrastructure spending projected to exceed $320 billion in 2025, hardware bottlenecks, particularly in GPU availability, continue to challenge hyperscalers and enterprises. Positron offers a specialized alternative focused on inference workloads, a segment expected to see rising demand as generative AI applications scale. Atlas delivers 3.5 times the performance-per-dollar and up to 66% lower power consumption compared to NVIDIA’s H100, according to the company.
“Improving the cost and energy efficiency of AI inference is where the greatest opportunity lies,” said Randy Glein, co-founder of DFJ Growth. “Positron’s chip and memory architecture removes existing bottlenecks and democratizes access to AI.”
Founded in 2023, Positron was built around a capital-efficient model, bringing Atlas to market within 18 months using just $12.5 million in seed funding. The company is led by CEO Mitesh Agrawal, formerly COO at Lambda, with technical leadership from co-founders Thomas Sohmers and Edward Kmett.
Atlas uses a memory-optimized FPGA-based design that achieves 93% memory bandwidth utilization and supports large-scale transformer models within a standard 2-kilowatt server footprint. The system is compatible with Hugging Face models and integrates through an OpenAI-compatible API. Positron chips are fabricated in the U.S.
Dylan Patel, CEO of SemiAnalysis and an advisor to the company, noted that Positron’s approach to memory scaling allows for more high-speed capacity per chip than existing solutions. Titan, the upcoming product built on Positron’s custom Asimov silicon, will extend this capability further—supporting up to 16-trillion-parameter models and significantly expanding inference capacity.
The company’s early customer base includes Cloudflare and Parasail (via SnapServe), with additional enterprise deployments underway.
“We founded Positron to meet the demands of modern AI—running frontier models at the lowest cost per token,” said Agrawal. “Our architecture enables massive context lengths and model sizes to be served efficiently on a single system.”
Titan, set to launch next year, will feature up to two terabytes of directly attached high-speed memory per accelerator and support parallel hosting of multiple models. The system will use standard form factors and avoid exotic cooling requirements, easing integration for data centers.
Investors say Positron’s disciplined focus and functional software stack set it apart. “They’ve built a working solution on 2022-era FPGAs—before developing their ASIC,” said Gavin Baker, managing partner at Atreides Management. “That speaks volumes about their technical depth and execution.”
