Leadership1 publisher3 min readPublished
Investors quintupled Positron's valuation months before Asimov's tapeout
The $875 million round prices an inference chip whose performance figures come from cycle-accurate simulations and whose production date has moved from early 2027 into the second half. The system Positron ships today runs on HBM.
The Board Room · Leadership desk
What happened
- Positron AI raised $875 million at a $5 billion post-money valuation on Sept. 10, 2026, to take its Asimov inference chip through tapeout and into production.
- Production of Asimov, which Positron targeted for early 2027 when it raised in February, now sits in the second half of 2027.
- Asimov's performance figures come from cycle-accurate simulations, and because the chip has not taped out, none of them come from production silicon.
Compiled by The Board RoomSomething wrong?How this is made
Why it matters
- constraint Anyone whose 2027 capacity plan leans on non-HBM parts now has to absorb a two-quarter slip that was booked before the riskiest milestone, tapeout, has happened at all.
- decision How big the economic case looks depends on the latency target a buyer picks: the advantage Positron shows at 170 tokens per second per user is 26 times, and at 100 it is 2.4 times.
- exposure Procurement committees looking for independent GPU comparison data have nowhere clean to get it, because the firm that supplied Positron's comparison is now a shareholder with a board seat.
- contradiction Positron has paying customers on shipping hardware, but that hardware uses HBM and DDR5, so the deployments say nothing about the LPDDR5X design the $5 billion price is built on.
The two tranches in this round were not priced alike. The $375 million Series C valued Positron at $3.5 billion pre-money, which puts the company at $3.875 billion once that tranche closes [3][1]. The round is described as $5 billion post-money, so if the Series C-1 of up to $500 million is drawn in full, it is priced off roughly $4.5 billion pre-money [1][3][2]. That is about $625 million of appreciation between two tranches announced on the same day [3].
The engineering question the money does not settle is bandwidth. The Register calculated in February 2026 that Nvidia's Rubin offers 22 terabytes per second of peak HBM bandwidth against roughly 3 terabytes per second for Asimov, and that Rubin would still be about 2.4 times faster even if GPUs realized only 30 percent of peak [18]. Thirty percent of 22 is 6.6 terabytes per second. Positron's product page lists 2.76 terabytes per second of realizable bandwidth, and 6.6 divided by 2.76 is 2.4 [19][4]. Positron's answer is utilization: it says Asimov can use more than 90 percent of available memory bandwidth, against less than 30 percent for GPUs running the same models [10]. That figure comes from cycle-accurate simulation [16].
The cost claim moves with latency. Positron's Titan comparison shows 26 times more tokens per dollar than Nvidia's GB300 NVL72 at 170 tokens per second per user, the fastest Blackwell rate shown, and 2.4 times at 100 tokens per second per user, so the advantage falls by a factor of about 11 between two points on the same page [15][5].
The hardware Positron has in the field runs on HBM. More than 50 racks of its first-generation Atlas system were being deployed at Oracle Cloud Infrastructure as of Sept. 10, with Parasail using the capacity for its inference service and Jump Trading and i3d.net as production customers [12]. Atlas runs on Agilex 7 M-series FPGAs with HBM and DDR5, and the LPDDR5X design planned for Asimov has not shipped [13]. Dylan Patel said Positron's architecture works "without depending on HBM or advanced packaging for its next-generation systems" [14].
Pre-tapeout pricing is ordinary in fabless silicon, and Cisco Investments, Hudson River Trading and Naver Ventures did diligence a reader cannot see [4]. What changed since February is the calendar: production was targeted for early 2027 and the Sept. 10 release places it in the second half, a slip of at least two quarters booked before the chip has taped out [8][6]. Tapeout is scheduled for the end of 2026 on TSMC's N3P process [7].
The board-deck version is that a $5 billion valuation, a returning sovereign investor and three strategic corporates de-risk the part [1][4]. What the money buys is itemized: the tapeout, a data center and emulation platform rated above 2 megawatts, and the Titan production ramp including commitments for LPDDR5X supply [6]. None of that produces a measured number before 2027.
Anyone sizing inference capacity who reserves 2027 volume against Asimov is accepting a simulated bandwidth utilization figure and a GPU comparison set supplied by SemiAnalysis, whose founder co-led a tranche and is joining the board [16][17]. Positron's own product page carries no teraFLOPS figure at all [19].
What to watch
- Whether Asimov actually tapes out on TSMC N3P by the end of 2026, the milestone the current schedule already slipped ahead of.
- A measured bandwidth figure from production silicon, published by a party that does not hold Positron equity or a board seat.
- Whether the Series C-1 is drawn in full, since the $5 billion post-money figure depends on the whole $500 million arriving.