Positron AI raises $875 million at a $5 billion valuation -- five times its February mark -- betting on commodity memory over Nvidia's HBM

Positron AI has raised $875 million in a two-tranche round that values the inference-chip startup at $5 billion -- five times what it was worth seven months ago, when a $230 million Series B valued the company at just over $1 billion -- to fund a bet that commodity, smartphone-grade memory can out-compete Nvidia's specialized high-bandwidth memory on the one AI workload growing faster than chipmakers can supply it for: running already-trained models at scale, rather than training new ones.

The round closed September 10 and was co-led by NEA, Atreides Management, Valor Equity Partners, Andra Capital and SemiAnalysis Capital -- the research firm run by chip analyst Dylan Patel, who joins Positron's board alongside NEA's Forest Baskett, Atreides' Gavin Baker and Netscape co-founder Jim Clark's representative Thomas Jermoluk. Qatar Investment Authority, Cisco Investments and Hudson River Trading also participated. Positron says the money funds the tape-out of its Asimov chip, a 2-plus-megawatt engineering data center to test it in, and the production ramp of Titan, the multi-chip system built around it. SemiAnalysis's participation is itself a signal worth naming: Patel's firm has spent 2026 publishing some of the most-cited independent GPU benchmarking in the industry, and a benchmarking house taking a board seat at a company it will presumably keep covering is a conflict Positron's own materials don't address.

Positron is not the only well-funded challenger chasing Nvidia on inference economics rather than raw model training. This newsroom's own reporting on the inference-chip pivot in late August found Nvidia, OpenAI/Broadcom and three separate startups all converging on performance-per-watt as the metric that matters, with three inference-chip startups closing more than $1.5 billion in combined funding the same week -- alongside established rivals Groq, Cerebras and SambaNova, all building custom silicon around the same bet that inference, not training, is where the compute shortage actually bites. Positron's angle within that crowd is specifically memory capacity rather than raw throughput or power efficiency, which is a narrower and more falsifiable claim than most of the category's marketing -- once Asimov exists to test.

The technical bet is a direct challenge to how every major AI chipmaker builds for inference. Nvidia's GPUs pair compute with high-bandwidth memory (HBM) -- fast, but capacity-constrained and, per The Register's February reporting on Positron's approach, in tight enough supply that it has become the industry's real bottleneck. Positron's Asimov chip instead uses LPDDR5X, the same class of memory found in phones and laptops: cheaper, more abundant, and -- per specifications in Positron's own announcement -- available in far larger quantities per chip, from 288 gigabytes up to 2.3 terabytes when expanded over CXL.

That capacity gap is real and documented: Nvidia's H100 ships with 80GB of HBM3 and the H200 with 141GB of HBM3e, both well short of Asimov's stated range. What isn't independently verified yet is whether the gap translates into the performance Positron claims -- as much as 26 times the tokens per dollar of Nvidia's Blackwell GB300 NVL72, on the strength of hitting roughly 90% memory-bandwidth utilization against what the company says GPUs typically achieve in the real world (about 30%). Every one of those figures comes from Positron's own simulations; nobody outside the company has run Asimov, because Asimov does not exist as shipping silicon yet. The Register's earlier reporting flagged the same gap in February: Nvidia's Rubin generation still moves data roughly 2.4 times faster than Positron's on-package bandwidth, and a memory-capacity advantage does not by itself settle whether a workload was ever memory-bound in the first place.

"Our focus now is to tape out Asimov, bring Titan to production, and scale manufacturing to meet the demand in front of us." -- Mitesh Agrawal, Positron AI CEO

Positron isn't starting from zero. Atlas, an earlier, GPU-adjacent rack product, is already running in production: more than 50 racks are deployed at Oracle Cloud Infrastructure, with Parasail, Jump Trading and i3d.net named as customers. Asimov and Titan are a generational bet layered on top of that base -- Asimov's tape-out is targeted for the end of 2026, with production pushed to the second half of 2027, a schedule that has already slipped once from the February round's stated early-2027 target.

The valuation, in other words, is a bet on a roadmap more than a product. Positron's own numbers -- the memory-capacity comparison, the deployed Atlas base -- are sourced and real. The numbers that would actually settle whether Asimov beats Nvidia on inference economics do not exist yet, because the chip they'd be measured on has not been built.