Positron bets $875 million that inference doesn't need expensive memory
A chip startup raises a massive round on the thesis that commodity memory can outpace HBM for running AI models, and a warehouse robotics company finally steps out of the shadows.
Positron AI raised $875 million in a Series C round at a $5 billion valuation, doubling down on a "memory-first" architecture for AI inference that sidesteps the HBM bottleneck entirely. The round was co-led by NEA, Atreides Management, Valor Equity Partners, Andra Capital, and SemiAnalysis Capital (Dylan Patel's fund), with participation from Qatar Investment Authority, Cisco Investments, DFJ Growth, and others. The company, led by CEO Mitesh Agrawal, builds inference hardware on commodity LPDDR5X memory — the same cheap, widely available memory found in phones and laptops — rather than the scarce and expensive high-bandwidth memory that Nvidia and other chipmakers have been fighting over. The product lineup includes Atlas, the Asimov custom silicon, and Titan, a multi-terabyte-memory system purpose-built for long-context workloads. Positron already counts Oracle Cloud Infrastructure, Parasail, Jump Trading, and i3d.net as customers, according to the company. The raise is a direct bet that inference economics will diverge from training: where training a frontier model demands the most expensive memory on the planet, serving millions of concurrent queries to those models may not. If that thesis holds, Positron's LPDDR5X-based systems could undercut HBM-heavy competitors on cost per query by a wide margin, especially as context windows keep stretching and inference becomes the dominant workload in production AI.
Maven Robotics emerged from stealth with $100 million from RoboStrategy, LocalGlobe, Vine Ventures, and XTX Markets Ventures, after two years of quietly running robots in real warehouses. The company builds wheeled, dual-armed robots that handle "mixed palletizing" — the grueling work of rearranging boxed goods from uniform factory pallets onto mixed-store pallets, a workflow that remains almost entirely manual today. Maven's eight deployed robots run 16 hours a day with 99 percent or higher uptime, according to the company. Founder Derbas, who spent nine years on Apple's special projects group before starting Maven, says the approach is purely operational: Maven won a deal over established competitors by visiting a customer's factory floor first and identifying flows where robots could add immediate value, rather than pitching a specific hardware architecture. The company plans to build 250 of its third-generation robots and begin designing a fourth-gen platform. The announcement comes as Agility Robotics, a legged-robot competitor, prepares for a $2.4 billion SPAC listing — a contrast Derbas doesn't shy from, arguing that bipedal designs "make zero sense" for warehouse workflows and add unnecessary cost.
Is commodity memory the real unlock for production AI, or will HBM costs come down fast enough to make the distinction irrelevant? Tell us in the comments.
