The US chip manufacturer Positron AI has raised 875 million dollars in a two-part funding round, reaching a valuation of five billion dollars. The company from Reno, Nevada, relies on inexpensive standard memory for its upcoming inference chip Asimov instead of scarce HBM memory. The valuation has more than quadrupled within seven months.
Investors quadruple the valuation within seven months
The funding round splits into two tranches: 375 million dollars Series C at a pre-money valuation of 3.5 billion dollars, plus a Series C-1 of up to 500 million dollars. The round is led by investment firms NEA, Atreides Management, Valor Equity Partners, and Andra Capital, along with SemiAnalysis Capital, run by industry analyst Dylan Patel; Netscape co-founder Jim Clark additionally leads the second tranche. Other backers include the Qatar Investment Authority, Cisco Investments, and trading firm Hudson River Trading. Positron had only closed a Series B of 230 million dollars at a valuation of 1.06 billion dollars in February 2026 – the new five-billion-dollar valuation marks a quadrupling within seven months. CEO Mitesh Agrawal, formerly COO of data center provider Lambda, will now sit on the board alongside investors Forest Baskett, Gavin Baker, Dylan Patel, and Thomas Jermoluk. Positron was founded back in 2023 by hardware engineers Thomas Sohmers and Barrett Woodside, who initially built the company on far smaller capital before today’s inference boom drew investor interest.
Asimov forgoes scarce high-performance memory
Unlike most AI accelerators, Positron’s upcoming Asimov chip skips High Bandwidth Memory in favor of standard LPDDR5X memory, the kind also found in smartphones and laptops. That pricier HBM memory has faced sharp increases since mid-2026: Samsung raised prices for its HBM4 chips by up to 20 percent. Depending on configuration, a single Asimov chip offers between 288 and 2,304 gigabytes of memory – far more than current accelerators reach per chip. The chip is set to tape out at TSMC on the 3-nanometer N3P process by the end of 2026, with mass production planned for the second half of 2027. Multiple Asimov chips can be combined into the Titan system, which the company says will serve models with more than 16 trillion parameters and context windows beyond ten million tokens. Whether the cheaper memory can actually match the bandwidth of specialized HBM chips in practice is independently unverified – the claims so far come solely from Positron itself, not from independent benchmark tests.
Customers already rely on predecessor system Atlas
Before Asimov launches, Positron is selling its current Atlas system, deployed in more than 50 racks – including at Oracle Cloud Infrastructure, trading firm Jump Trading, hosting provider i3d.net, and AI service Parasail. Experience from these ongoing Atlas deployments at Oracle fed directly into the design of Asimov and Titan, according to CEO Agrawal. Focusing on inference rather than training, Positron enters a market where Nvidia has already answered similar competition with its own specialized chips like Groq 3 LPX. The fresh capital will fund the completion of Asimov’s development, a test data center with more than two megawatts of power, and the ramp-up of Titan production. For the 70-person company, nine-tenths of whom are engineers by its own account, the round is also a bet on hitting the tight window to planned 2027 mass production without major delays while holding its own against far bigger rivals like Nvidia.
What will matter is whether Positron can keep its cost promise once Asimov reaches mass production in mid-2027 – so far the memory concept’s advantage rests solely on the company’s own claims. It also remains open whether customers will actually switch from the established Nvidia ecosystem to an as-yet-unproven chip for serving large language models, whose software stack still has to prove itself in the market.
