Positron's $875m round prices an inference chip at $5bn before it has taped out
The Reno chip start-up has raised a two-tranche Series C to bring its Asimov processor to TSMC, betting that commodity memory, not scarce HBM, is the way to cheaper AI inference.
Positron AI, a Reno, Nevada maker of hardware for running artificial intelligence models, announced on September 10 an $875m Series C financing at a $5bn post-money valuation. The round was co-led by NEA, Atreides Management, Valor Equity Partners, Andra Capital, Dylan Patel's SemiAnalysis Capital and Jim Clark, the founder of Silicon Graphics and Netscape. It comes seven months after a $230m Series B in February that valued the company at more than $1bn.
The headline numbers place Positron among the best funded challengers to Nvidia in the inference market, and Dealroom ranks the deal among the largest Series C rounds ever raised by a US semiconductor company. But the more interesting feature is what the money is buying. Positron's next-generation chip, Asimov, has not yet been manufactured. Investors are paying a growth-stage price for a design, a production plan and a thesis about where the bottleneck in AI computing really sits.
A round in two parts
The financing came in two tranches. The first is a $375m Series C at a $3.5bn pre-money valuation. The second, a Series C-1 of up to $500m, is led by NEA and Jim Clark. Other investors include DFJ Growth, the Qatar Investment Authority and Resilience Reserve, with VentureTech Alliance, Hudson River Trading, Cisco Investments and Naver Ventures joining as strategic investors. Forest Baskett of NEA, Gavin Baker of Atreides, Thomas Jermoluk of Jim Clark's office and Dylan Patel join the board.
The structure deserves careful reading. On our arithmetic, the first tranche alone implies a post-money value of $3.875bn. For the full $875m to produce a $5bn post-money figure, the C-1 money has to come in at a higher implied price, roughly $4.5bn before that tranche. The words "up to" also matter: the headline amount and valuation assume the second tranche is fully drawn. Staged rounds of this kind are common in hardware, where capital needs rise sharply around manufacturing milestones, and they let the most committed investors set a higher mark while protecting themselves if the schedule slips.
The memory bet
Positron's argument is that inference, the work of running trained models for users, is limited less by raw arithmetic than by memory: how much data a chip can hold close to its processors and how quickly it can move it. Nvidia's answer to that problem is high-bandwidth memory, or HBM, stacked next to the processor using advanced packaging such as TSMC's CoWoS. Both are expensive and in short supply.
Positron is designing around them. Its "memory-first" systems use commodity LPDDR5X memory, the kind used in mobile devices, and the company says its design reaches more than 90% of available memory bandwidth. Asimov is due to tape out on TSMC's N3P process at the end of 2026, with production expected in the second half of 2027, and each chip will offer 288GB to 2,304GB of memory. Titan, the system built around it, will combine four to eight Asimov chips and is designed for models above 16 trillion parameters with context windows of more than 10 million tokens per node.
Dylan Patel, whose research firm is closely followed by chip investors, summed up the case: "Positron's architecture addresses the real constraint, memory". NEA's Forest Baskett was more direct about the risk appetite: "We led this round because we believe Positron is making the boldest memory-first bet in AI hardware."
The bandwidth figures are company claims for a chip that does not yet exist in silicon, and they should be read that way. Commodity memory offers capacity and supply security, but its raw bandwidth is lower than HBM's. Whether a better utilisation rate makes up the difference on real workloads is the central technical question, and it will be answered only after tapeout.
What already works
Positron is not starting from zero. Its first-generation Atlas system is being deployed in more than 50 racks at Oracle Cloud Infrastructure, and customers include Parasail, Jump Trading and i3d.net. That matters for two reasons. It shows the company can sell and support systems in production data centres, which many chip start-ups never achieve, and it gives the team operating data on the inference workloads Asimov is meant to serve.
The customer and investor lists also overlap in a revealing way. Jump Trading co-led the Series B and is a production customer. Hudson River Trading, another quantitative trading firm, joins this round as a strategic investor. Trading firms care intensely about latency and cost per query, and their participation suggests that the buyers who scrutinise hardware most closely see enough in Positron to put money behind it. The Qatar Investment Authority, which also took part in the Series B, returns as well; Positron opened an office at the Dubai International Financial Centre earlier this year, its first outside the United States.
The schedule is the risk
The new money will fund the Asimov tapeout, a 2MW-plus engineering data centre and emulation platform, and the production ramp for Titan. Chief executive Mitesh Agrawal set out the priorities: "Our focus now is to tape out Asimov, bring Titan to production, and scale manufacturing to meet the demand in front of us."
When the company raised its Series B in February, Asimov was described as shipping in early 2027. The plan announced this week puts production in the second half of 2027. A slip of a few months is ordinary in semiconductors, but it is a reminder of how much execution risk sits between a funding announcement and revenue. A first tapeout on a leading-edge node can require a respin; supply of memory, packaging and wafer capacity has to be secured in advance; and Nvidia, the incumbent, will have shipped further generations of its own hardware by the time Asimov reaches customers.
For the venture market, the round is another data point in the repricing of AI infrastructure. A roughly fivefold rise in valuation in seven months, before the flagship product exists, shows how much capital is willing to fund alternatives to Nvidia in inference, the part of the market expected to grow as AI models move from training labs into everyday use.
Our view
Positron's raise is a rational bet with an unusually clear test. The company has identified a real constraint, the cost and scarcity of high-bandwidth memory, and has built a credible first product with production customers. Investors are paying for the option that Asimov delivers on its claims, and they have structured the round to stage that exposure.
For founders in AI hardware, the lesson is that capital is available at striking prices for teams that can show working systems and a distinct architectural thesis, not just a faster chip. For investors, the valuation leaves little room for delay or a disappointing first silicon. What to watch: whether Asimov tapes out on schedule at the end of 2026, whether the full C-1 tranche is drawn, how many racks Atlas reaches at Oracle and elsewhere, and whether independent benchmarks confirm the memory utilisation claims. Until then, $5bn is the price of a promise, backed by some of the most technically demanding investors in the market.
Sources
- Latham & Watkins: Latham & Watkins Advises Positron AI on US$875 Million Series C
- Dealroom: Positron raises $875M Series C at $5B valuation for inference chips
- Pulse 2.0: Positron AI Raises $875 Million Series C At $5 Billion Valuation To Scale Memory-First AI Inference Chips
- Middle East AI News: Positron AI raises $875 million at $5 billion valuation
- SiliconANGLE: Positron AI raises $230M at over $1B valuation to build energy-efficient AI accelerator hardware
Filed under Startups. Written by InvestTimes Editorial under our editorial standards. Commentary and analysis for information only, not investment advice.
