Positron AI has raised $875M in Series C funding, a bet that the bottleneck in AI is shifting from compute to memory.
The Reno company builds inference systems that prioritize memory capacity, bandwidth and power efficiency over raw speed, using commodity LPDDR5X memory in air-cooled or liquid-cooled racks. Its products include Atlas, custom silicon called Asimov, and Titan, a multi-terabyte-memory machine aimed at long-context workloads.
The round was co-led by NEA, Atreides Management, Valor Equity Partners, Andra Capital, SemiAnalysis Capital and Jim Clark, with DFJ Growth, the Qatar Investment Authority, Cisco Investments and Naver Ventures among the participants. The breadth of the syndicate, mixing chip veterans, sovereign wealth and strategics, reflects how contested inference hardware has become.
Chief executive Mitesh Agrawal’s customers already include Oracle Cloud Infrastructure, Parasail, Jump Trading and i3d.net. Positron says it will spend the money on operations and product development.
Running models is now the cost center that matters most as agents multiply the number of queries. Startups promising cheaper tokens per watt are drawing record checks, and Positron’s raise puts it among the best funded of them.
