Recommended Articles
Fundraising

Positron AI Raises $875M at $5B Valuation to Prove Commodity Memory Can Beat HBM in AI Inference

By Angeled · October 19, 2026 · Fundraising
Positron AI Raises $875M at $5B Valuation to Prove Commodity Memory Can Beat HBM in AI Inference
Positron AI Raises $875M at $5B Valuation to Prove Commodity Memory Can Beat HBM in AI Inference

Positron AI, the Reno-based AI inference hardware startup, has raised $875 million in Series C financing at a $5 billion post-money valuation — roughly five times the ~$1 billion valuation it held after its $230 million round in February 2026. The financing was announced on September 10, 2026 in two tranches: a $375 million Series C at a $3.5 billion pre-money valuation, and an up to $500 million Series C-1.

The syndicate

The round was co-led by NEA, Atreides Management, Valor Equity Partners, Andra Capital, and SemiAnalysis Capital — an unusual co-lead that pairs a top-tier venture firm with a specialist AI analyst firm (SemiAnalysis, led by Dylan Patel) — with Silicon Graphics and Netscape founder Jim Clark co-leading the C-1 alongside NEA. Also participating: DFJ Growth, the Qatar Investment Authority, Hudson River Trading, Cisco Investments, and Naver Ventures. Four investors joined the board, including Gavin Baker of Atreides and Dylan Patel of SemiAnalysis.

The architectural bet: inference is a memory problem

Positron's core claim is that large-scale AI inference is increasingly constrained by memory capacity, usable memory bandwidth, and power — not peak compute. Its next chip, Asimov, is designed around commodity LPDDR5X memory — the DRAM class found in phones and laptops — instead of High Bandwidth Memory (HBM), the expensive 3D-stacked specialty memory beside Nvidia's H100, H200, and Rubin accelerators.

The bet has two economic legs:

  1. Supply. HBM supply is chronically constrained and expensive, with advanced CoWoS packaging adding cost and lead times. LPDDR5X is abundant and cheap.
  2. Utilization. Positron argues the problem isn't how much bandwidth an accelerator peaks at, but how much of it actually gets used. It claims its architecture can realize more than 90% of available memory bandwidth, versus typical GPU decode utilization below 30%.

Asimov is designed with 288 GB to 2,304 GB of memory per chip, versus 80 GB of HBM3 on an H100, and supports air as well as liquid cooling. It is scheduled to tape out on TSMC's N3P process in October 2026 — just 16 months after the design process began — with production targeted for the second half of 2027. Its systems arm, Titan, will combine four to eight Asimov processors, targeting models beyond 16 trillion parameters and context windows exceeding 10 million tokens in a single node, scaling to thousands of nodes. Positron says simulations show Titan delivering up to 26x more tokens per dollar than Nvidia's Blackwell GB300 NVL72 — a claim not yet tested against live silicon.

Real customers, real racks

Unlike many AI chip stories, Positron has production experience. Its first-generation Atlas platform is shipping today, with more than 50 Atlas racks deployed at Oracle Cloud Infrastructure, where the inference company Parasail uses that capacity; Jump Trading and i3d.net are identified as production customers. The new capital also funds a 2MW+ engineering data center and emulation platform and the Titan production ramp.

The company was founded in 2023 by Thomas Sohmers (co-founder and chairman) and CEO Mitesh Agrawal.

The timing is not accidental

The raise landed two days after Qualcomm and Amazon disclosed a multi-generational LPDDR-based inference partnership worth up to $60 billion in purchases — a deal that independently validated Positron's architectural thesis. Inference, not training, is where the industry's compute economics are now being decided, and everyone is hunting alternatives to Nvidia's ecosystem.

Why it matters for founders and investors

For founders: Positron illustrates a winning pattern in hard-tech fundraising — ship a real product before asking for silicon money. Atlas's revenue-generating deployment at OCI gave the syndicate something to underwrite beyond slides. If you're building hardware, the Series C that pays for tapeout is won or lost on traction from generation one.

For investors: the $5B mark-up in seven months prices in successful tapeout and a 2027 production ramp — both hard milestones with physical risk. But the underlying trend (inference economics, HBM constraints) is one of the strongest secular bets in compute. The diligence question is utilization: does the >90% bandwidth claim hold in real customer workloads?

Angeled.com connects founders building the next layer of the AI stack with investors.


Sources: Converge Digest; Data Center Dynamics; Tech Times; The Investor Society. Performance and bandwidth figures are company-reported and not independently confirmed.

Read Next