No HBM, No Problem: Positron Raises $875M to Scale Its Memory-First Inference Bet

The fiercest constraint in AI hardware right now is not compute — it is memory. And one Reno, Nevada-based startup just raised one of the year’s largest semiconductor rounds on exactly that thesis. Positron AI announced an $875 million Series C at a $5 billion post-money valuation, money it says will fully fund its next-generation inference silicon through production, without a single byte of scarce high-bandwidth memory in its next systems.

The round, announced Thursday, was co-led by NEA, Atreides Management, Valor Equity Partners, Andra Capital, SemiAnalysis Capital, and Jim Clark — the founder of Silicon Graphics and Netscape. Structurally, the financing came in two tranches: a $375 million Series C at a $3.5 billion pre-money valuation, followed by a Series C-1 of up to $500 million led by NEA and Jim Clark. The long investor list also includes DFJ Growth, Qatar Investment Authority, Hudson River Trading, Cisco Investments, and Naver Ventures. Four new board members arrive with the money: Forest Baskett of NEA, Gavin Baker of Atreides, Thomas Jermoluk from Jim Clark’s office, and — notably — Dylan Patel, the founder of SemiAnalysis, whose firm’s research has shaped how the industry thinks about inference economics.

The memory-first bet

Positron’s argument starts with a shift everyone in the industry can feel: AI infrastructure spending is pivoting from training frontier models to running them in production. Inference at scale is dominated by a different set of bottlenecks than training — memory capacity, memory bandwidth, and power — and the current answer to those bottlenecks, HBM stacked on advanced packaging, is supply-constrained to the point of distorting product roadmaps across the industry.

Positron’s counter-proposal is a memory-first architecture built on commodity LPDDR5X — the same class of memory found in laptops and phones, abundant and cheap. The company claims its design can realize more than 90% of available memory bandwidth, avoids dependence on HBM and CoWoS advanced packaging supply chains entirely, and supports both air-cooled and liquid-cooled data centers across varying rack densities.

That last claim carries weight because of Forest Baskett’s comment in the announcement. The NEA partner pointed out that even Nvidia’s Rubin Ultra roadmap has had to scale back — from a terabyte of HBM4E down toward 192GB per package — “simply because the supply isn’t there.” When the incumbent’s flagship roadmap bends under memory scarcity, a startup that engineered that dependency out of its bill of materials is making a timely argument.

From Atlas to Asimov to Titan

The money is not going toward a slide deck. Positron says more than 50 racks of its first-generation Atlas platform are already deployed at Oracle Cloud Infrastructure, with Parasail using that capacity for its inference service and Jump Trading and i3d.net among the production customers. Atlas is an inference appliance built around Positron’s Archer accelerators — earlier versions of which were reported to beat an Nvidia H200 on Llama 3.1 8B inference while drawing roughly a third of the power. Real racks, in a real hyperscaler cloud, serving paying inference traffic.

The Series C funds the next two steps on that roadmap. First is Asimov, Positron’s next-generation custom silicon, which is scheduled to tape out on TSMC’s N3P process at the end of 2026, with production expected in the second half of 2027. Each Asimov chip pairs Positron’s compute architecture with 288 GB to 2,304 GB of memory — a staggering ceiling when a leading HBM-equipped accelerator today carries 141–288 GB.

Second is Titan, the system that wraps four to eight Asimov chips into a single node. Positron is designing Titan to serve models with more than 16 trillion parameters and context windows exceeding 10 million tokens in a single node, scaling out to thousands of nodes. The round also funds a 2-megawatt-plus engineering data center and emulation platform to validate all of it.

Why the valuation jumped

Seven months ago, Positron closed a $230 million Series B at just over a $1 billion valuation. The Series C prices the company at $5 billion — a roughly 4x jump in half a year, and an even steeper climb from the $23.5 million seed round the company announced in February 2025. Reuters noted the two-tranche structure as investors piled in beyond the original target.

What changed between February and September is the proof. Atlas deployments at Oracle moved Positron from a promising architecture story to a company with production hardware running inside a top-tier cloud. Dylan Patel’s decision to invest personally and join the board is perhaps the sharpest signal: SemiAnalysis built its reputation on measuring what AI hardware actually delivers, and Patel said plainly that “most inference economics struggle under that scrutiny” — but that Positron’s architecture “addresses the real constraint, memory.”

The road ahead

The competitive bar is brutal. Nvidia dominates inference today, and a long line of well-funded challengers — Cerebras, Groq, d-Matrix, SambaNova, and the major cloud vendors’ own silicon among them — are all chasing the same workload shift. Positron’s differentiation is real but narrow: if LPDDR5X-based systems can genuinely deliver competitive tokens-per-second-per-dollar and per-watt at production quality, the supply-chain independence becomes a major advantage as HBM remains booked out. If the memory-first trade-off costs too much throughput, the cheap memory stops mattering.

The timeline is also unforgiving. Asimov must tape out on N3P by end of 2026 and hit production in 2027 — years in which Nvidia, AMD, and every custom-silicon program in the industry will also be shipping. But with $875 million banked, real deployments behind it, and a board stacked with semiconductor veterans, Positron has bought itself a genuine seat at the inference table. The AI industry has spent two years learning that memory is the wall. Positron just raised nearly a billion dollars on the bet that it found the door.