IMPLICATOR
Positron AI raised $875 million at a $5 billion post-money valuation on Sept. 10, 2026, to take its Asimov inference chip to tapeout and production. The valuation is nearly five times the $1 billion it carried after a $230 million Series B in February 2026. Asimov uses commodity LPDDR5X memory instead of the HBM that supplies Nvidia’s accelerators and remains short of demand, but the chip has not taped out.
What Changed
- Positron AI raised $875 million at a $5 billion post-money valuation on Sept. 10, nearly five times the $1 billion it carried after its February 2026 Series B.
- The round came in two tranches: a $375 million Series C at a $3.5 billion pre-money valuation, and a Series C-1 of up to $500 million led by NEA and Jim Clark.
- Asimov, built on LPDDR5X instead of HBM, tapes out at the end of 2026. Production, targeted for early 2027 in February, now sits in the second half of 2027.
- Asimov's performance figures come from simulations. The GPU comparison data comes from SemiAnalysis, whose founder's fund co-led a tranche and who will join the board.
AI-generated summary, reviewed by an editor. More on our AI guidelines.
The round
The financing announced Sept. 10 came in two tranches. A $375 million Series C valued the Reno, Nevada, company at $3.5 billion before the investment and was co-led by NEA, Andra Capital, Atreides Management, Valor Equity Partners and Dylan Patel’s SemiAnalysis Capital. A Series C-1 of up to $500 million was led by NEA and Jim Clark, the Silicon Graphics founder and Netscape co-founder.
Qatar Investment Authority returned after investing in the February 2026 Series B. Cisco Investments, Hudson River Trading and Naver Ventures joined as strategic investors. Forest Baskett of NEA, Gavin Baker of Atreides, Thomas Jermoluk of Jim Clark Office and Patel will join Positron’s board.
The money will fund Asimov’s tapeout, a data center and emulation platform rated above 2 megawatts, and the Titan production ramp, including commitments for LPDDR5X supply.
FREE · ABOUT FIVE MINUTES
Follow the inference chips built to rival Nvidia.
Implicator cuts the flood to the developments worth your attention, with concise reporting on what changed and why it matters. Delivered every weekday at 4:45 a.m. Pacific, 7:45 a.m. Eastern.
From San Francisco. No spam. Unsubscribe anytime.
The memory bet
Asimov is planned for TSMC’s N3P process, with tapeout at the end of 2026. In February 2026, production was targeted for early 2027; the Sept. 10 release places it in the second half of 2027. The chip is designed with 288 gigabytes to 2,304 gigabytes of memory and power use of about 400 watts. Positron says it can use more than 90 percent of available memory bandwidth, compared with less than 30 percent for GPUs running the same models. Titan is designed to combine four or eight Asimov chips.
What runs at Oracle
More than 50 racks of Positron’s first-generation Atlas system were being deployed at Oracle Cloud Infrastructure as of Sept. 10, 2026. Parasail uses the capacity for its inference service, while Jump Trading and i3d.net are production customers.
Atlas is built on Agilex 7 M-series FPGAs with HBM and DDR5. The LPDDR5X design planned for Asimov has not shipped. Patel said Positron’s architecture works “without depending on HBM or advanced packaging for its next-generation systems.”
The numbers are simulated
Positron’s Titan comparison claims 26 times more tokens per dollar than Nvidia’s GB300 NVL72 at 170 tokens per second per user, the fastest Blackwell rate shown. At 100 tokens per second per user, the page shows 2.4 times more tokens per dollar.
Know someone who'd find this useful? ✉️ Email it to a friend in one click, or they can subscribe free here.
The Asimov performance figures are based on cycle-accurate simulations, and Asimov has not taped out, so none of its figures come from production silicon. The GPU data comes from SemiAnalysis InferenceX. Dylan Patel’s SemiAnalysis Capital co-led the $375 million Series C tranche, and Patel, the firm’s founder and chief executive, will join Positron’s board.
The Register calculated in February 2026 that Nvidia’s Rubin offered 22 terabytes per second of peak HBM bandwidth, against roughly 3 terabytes per second for Asimov. Even if GPUs reached only 30 percent of peak bandwidth, Rubin’s memory would still be about 2.4 times faster. Positron’s product page lists 2.76 terabytes per second of realizable bandwidth and no teraFLOPS figure.
Tom’s Hardware wrote in July 2025 that Positron’s Atlas comparison “requires verification by a third party.”
“Our focus now is to tape out Asimov, bring Titan to production, and scale manufacturing to meet the demand in front of us,” said Mitesh Agrawal, Positron’s chief executive.
Frequently Asked Questions
How much did Positron raise, and at what valuation?
Positron AI announced $875 million at a $5 billion post-money valuation on Sept. 10, 2026. It came in two tranches: a $375 million Series C at a $3.5 billion pre-money valuation, and a Series C-1 of up to $500 million led by NEA and Jim Clark.
What is Asimov?
Asimov is Positron's next-generation inference chip. It is planned for TSMC's N3P process with 288 gigabytes to 2,304 gigabytes of memory and power use of about 400 watts. Tapeout is set for the end of 2026, with production in the second half of 2027.
Why does Positron use LPDDR5X instead of HBM?
HBM remains short of demand. Positron says its design can use more than 90 percent of available memory bandwidth, compared with less than 30 percent for GPUs. The Register calculated in February 2026 that Nvidia's Rubin memory would still be about 2.4 times faster even at 30 percent of peak.
What is Positron running at Oracle today?
More than 50 racks of Atlas, its first-generation system, were being deployed at Oracle Cloud Infrastructure as of Sept. 10. Atlas is built on Agilex 7 M-series FPGAs with HBM and DDR5, so the LPDDR5X design planned for Asimov has not shipped.
Are Positron's performance claims independently tested?
Asimov's figures are based on cycle-accurate simulations, and the chip has not taped out. The Titan page claims 26 times more tokens per dollar than Nvidia's GB300 NVL72 at 170 tokens per second per user, and 2.4 times at 100. The GPU data comes from SemiAnalysis InferenceX.
AI-generated summary, reviewed by an editor. More on our AI guidelines.
Marcus Schuler
Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai
