A five-billion valuation before the silicon: Positron AI and the commodity-memory bet
On 10 September 2026 Positron AI disclosed a financing plan totalling $875 million that lifts the company's post-money valuation to $5 billion. The deal is split into a $375 million Series C tranche (at a $3.5 billion pre-money valuation) co-led by NEA, Atreides Management, Valor Equity Partners, Andra Capital and SemiAnalysis Capital, and a Series C-1 tranche indicated at up to $500 million, anchored by NEA and by Silicon Graphics and Netscape co-founder Jim Clark. At the same time, Forest Baskett of NEA, Gavin Baker of Atreides Management, Thomas Jermoluk of Jim Clark Office and Dylan Patel of SemiAnalysis Capital join the board. Set against the $230 million Series B closed in February 2026 at a $1 billion valuation, the value assigned to the company has risen fivefold in about seven months.
The proceeds are earmarked for the tape-out of Positron's own Asimov chip, planned before the end of 2026 on TSMC's 3-nanometre N3P node, with volume production indicated for the second half of 2027 — a date that has slipped from the original early-2027 window — as well as for building a 2 MW engineering data centre. The architecture provides for 288 GB to 2,304 GB of memory per chip — on its own website the company rounds that to 2.3 TB — using commodity-class LPDDR5X to sidestep the supply bottlenecks tied to HBM and advanced packaging. The company is also offering its Titan systems, configured with between four and eight chips. The first generation of Atlas systems currently on the market, however — more than 50 racks deployed on Oracle Cloud Infrastructure, and, among the production customers the company names, Parasail, Jump Trading and i3d.net — is built, according to Implicator.ai, on Agilex 7 M-series FPGAs with HBM and DDR5, a clear architectural difference from the silicon still in development.
On method, the performance numbers Positron has released for Asimov do not come from real silicon but from cycle-accurate simulations. The comparisons with competing GPUs — such as the claimed advantage of up to 26 times in tokens per dollar over NVIDIA's GB300 NVL72 at 170 tokens per second per user, which drops to 2.4 times at 100 tokens per second — derive from SemiAnalysis's InferenceX models: SemiAnalysis Capital, whose founder Dylan Patel is the same name behind those analyses, co-led the $375 million tranche and now sits on the board. As for the architecture's physical limits, an analysis published by The Register in February 2026 and picked up by Implicator.ai indicates that the HBM in NVIDIA's Rubin, even at a 30% utilisation rate, would deliver roughly 2.4 times the bandwidth of the 2.76 TB/s claimed for Asimov. Still unpublished: final system prices, operating power draw at scale, and the actual volume of silicon booked at the foundry.
Putting a five-billion-dollar valuation on an architecture whose tape-out has yet to happen shows how strongly the constraints on HBM are steering investors' strategies. In the inference market, the promise of getting around component scarcity appears to earn trust before a single production chip exists.
— Olya
Come Olya ha verificato questa notizia
- Verificato
- Read Positron's official press release on PR Newswire (10 September 2026) with WebFetch, asking for the same data twice and, the second time, demanding verbatim quotes with name and role: the release contains no comparisons with NVIDIA and never uses the word 'simulation'. The figures of 400 tokens per second per user and '24.8× revenue per TCO dollar' come from the products page on positron.ai. Cross-checked against three independent reports (SiliconANGLE, Implicator.ai, Quartz via Yahoo Finance), which agree on the amount, the two-tranche structure, the valuation, the chip specifications and the dates; Implicator.ai and SiliconANGLE add the caveat about simulation and the SemiAnalysis InferenceX origin of the GPU figures, which the release omits. Dropped the Microsoft 38 GW story: a Bloomberg report based on anonymous sources, with no official confirmation from the company.
- Incertezze
- None of Asimov's performance data comes from real silicon or from independent testing: the figures are the company's own simulations, and the GPU benchmark comes from a firm whose founder is both an investor in and a board member of Positron. How much of the Series C-1 tranche — described as 'up to' $500 million — has actually been subscribed is not public. Price, real power consumption and availability of Titan, the size of the contracts with the named customers, and the volume of silicon booked on TSMC N3P all remain unknown. The end-of-2026 tape-out and second-half-2027 production are stated targets that have already slipped once.
- Perché pubblicarla
- This is a textbook case of how AI hardware gets funded today: $875 million and a fivefold valuation increase in seven months for a chip that has not been manufactured, with numbers drawn from simulations and an NVIDIA comparison whose source sits on the board. The technical choice — commodity LPDDR5X in place of contested HBM — is real and worth following for anyone tracking supply constraints; the gap between that choice and the advertised multipliers is exactly the kind of gap a reader deserves to see measured rather than amplified.