Positron raises $875M to get its next inference chip into silicon

Why it matters

The round funds Positron through tapeout and manufacturing, but the $5B valuation rests on Asimov reaching silicon on schedule and proving its economics on real workloads.

Positron AI announced that it raised $875 million at a $5 billion post-money valuation, giving co-founder Thomas Sohmers and CEO Mitesh Agrawal the capital to move Asimov, Positron's next-generation inference processor, from design into silicon.

Positron AI on X

The financing, detailed in a September 10th announcement, is meant to cover Asimov's tapeout on TSMC's N3P manufacturing process, a new engineering data center of more than 2 megawatts and the production ramp for Titan, a system built around four to eight Asimov chips. Positron is targeting tapeout by the end of 2026 and production in the second half of 2027.

That schedule makes the round a large, expensive bridge between the hardware Positron has already shipped and the custom silicon on which its $5 billion valuation depends. Asimov has not reached silicon. Its performance, manufacturing yield and economics will have to survive the transition from design targets to customer workloads.

The founder's second semiconductor swing

Sohmers started Positron in April 2023 with Edward Kmett, a functional-programming researcher and longtime Haskell contributor, and Barrett Woodside, whose earlier work included product roles at Nvidia, Google Cloud and Scale AI.

For Sohmers, Positron is another attempt to rethink the processors used in data centers. He left Massachusetts' Advanced Math and Science Academy at 17 to accept a 2013 Thiel Fellowship, then founded fabless semiconductor ventures REX Computing and Terapute Technologies. He later worked at Groq and Lambda, where he helped design an 800-GPU data center.

Positron's founding argument was blunt: inference, the process of running trained AI models, was becoming too expensive and power-hungry to remain dependent on general-purpose GPU infrastructure. The founders designed Atlas, Positron's first system, around that thesis and shipped the first customer server in August 2024, according to Positron's timeline.

Agrawal joined Positron as CEO in early 2025 after serving as COO and head of cloud at Lambda. His move brought an operator who had watched GPU cloud demand and infrastructure spending grow at close range into a hardware business trying to reduce those bills.

$875M, split into two checks

Positron says the financing includes NEA, Andra Capital, Atreides Management, Valor Equity Partners, SemiAnalysis Capital and Jim Clark. Forest Baskett of NEA, Gavin Baker of Atreides, Thomas Jermoluk from Jim Clark's family office and Dylan Patel are joining Positron's board.

That lineup puts semiconductor veterans, infrastructure investors and a prominent chip-industry analyst around the same table as Positron prepares for its most capital-intensive stage.

The valuation climbed rapidly. Positron announced a $230 million Series B in February 2026. Seven months later, the new financing values Positron at $5 billion, before Asimov has taped out and roughly a year before the planned production window opens.

The memory bet

Positron is building Asimov around a memory-first architecture. Positron says each chip will support between 288GB and 2,304GB of memory, while Titan will combine four to eight chips into a single system. Positron is designing Titan for models with more than 16 trillion parameters and context windows exceeding 10 million tokens.

Those specifications remain targets. The more immediate distinction is Positron's decision to use commodity LPDDR5X memory instead of relying on high-bandwidth memory and advanced CoWoS packaging, two heavily contested parts of the AI chip supply chain. Positron argues that the approach can deliver more usable memory capacity, lower power consumption and better inference economics without requiring customers to rebuild data centers around unusually dense liquid-cooled racks.

Atlas gives Sohmers and Agrawal some production experience to carry into that effort. Positron says more than 50 Atlas racks had been deployed at Oracle Cloud Infrastructure by August 2026. The supplied company material also names Cloudflare and Parasail as customers or evaluation partners. Positron has not disclosed the commercial terms of these relationships.

The deployments establish that Positron has moved first-generation hardware beyond demonstrations. They do not validate Asimov's promised memory bandwidth, power profile or performance per dollar. Custom silicon introduces a different set of risks, including tapeout timing, manufacturing yield, packaging, software support and the work required to run a broad mix of customer models reliably.

A valuation ahead of the silicon

Investors have shown a willingness to finance those risks across the inference market. In July, Etched raised $300 million at a $10.3 billion valuation, while SambaNova raised $1 billion at an $11 billion valuation. Groq, Cerebras and other specialized chipmakers are also pursuing workloads that might otherwise run on Nvidia GPUs.

Positron's pitch differs in its emphasis on memory capacity and ordinary LPDDR components. That choice could reduce exposure to scarce HBM and advanced-packaging capacity, while creating its own engineering and software demands. The architecture must keep enough of that memory bandwidth available to applications and compete across real models, batch sizes and latency requirements.

The $875 million round buys Sohmers and Agrawal time, manufacturing commitments and infrastructure. It also removes capital scarcity as an explanation if Asimov misses its targets. Positron has set late 2026 for tapeout and the second half of 2027 for production. The next proof point is silicon that performs outside Positron's roadmap.