Positron AI Raises $875 Million to Challenge Nvidia in Inference Chips

Positron AI, a three-year-old startup building computer chips for running trained artificial intelligence models, has raised $875 million in new funding at a $5 billion valuation, one of the largest sums yet committed to a company positioning itself as an alternative to Nvidia in the AI inference market. The financing was disclosed in a press announcement from the Reno, Nevada-based company on September 10.

The round values Positron at roughly five times what it was worth seven months ago, when it raised a $230 million Series B in February. It marks a rapid escalation for a company that raised a $23.5 million seed round in February 2025 and, by its own account, has now taken in more than $1.1 billion in total funding in under two years.

A two-tranche round with an unusual roster

According to the company's announcement, the $875 million was raised in two tranches: a $375 million Series C at a $3.5 billion pre-money valuation, co-led by NEA, Andra Capital, Atreides Management, Valor Equity Partners and Dylan Patel's SemiAnalysis Capital, and a Series C-1 of up to $500 million led by NEA and Netscape co-founder Jim Clark. Additional participants named in the release include DFJ Growth, the Qatar Investment Authority, Hudson River Trading, Cisco Investments and Naver Ventures.

A separate announcement from Liberty Global Tech Ventures, the venture arm of the publicly traded telecommunications company, confirmed it also put money into the round, without disclosing the size of its stake. Liberty Global said the deal adds Positron to a portfolio of AI investments that already includes Higgsfield, XBOW, Legora and ElevenLabs, and pointed to inference computing — the process of running, rather than training, AI models — as a market it expects to reach $1.3 trillion by 2032.

As part of the financing, Positron is adding four new board members: NEA partner Forest Baskett, Atreides' Gavin Baker, Thomas Jermoluk of Jim Clark's office, and Patel, whose SemiAnalysis newsletter and consultancy has become an influential voice in semiconductor and AI infrastructure analysis.

Betting on memory instead of raw compute

Positron was founded in 2023 by Thomas Sohmers, a former Thiel Fellow who previously built processors at his own startup REX Computing and worked as a hardware architect at Lambda, along with Barrett Woodside and mathematician Edward Kmett. The company's pitch is that most AI inference workloads are limited not by raw computing power but by how much data a chip can move in and out of memory — and that Nvidia's GPUs, built around high-bandwidth memory optimized for training, are inefficient for the job of serving already-trained models to users.

Positron's current-generation Atlas systems are already running commercially, including a deployment of more than 50 racks inside Oracle Cloud Infrastructure, according to the company. Its next-generation chip, code-named Asimov, is expected to tape out on Taiwan Semiconductor Manufacturing Co.'s N3P process by the end of 2026, with production targeted for the second half of 2027. Positron says Asimov will offer between 288 gigabytes and 2,304 gigabytes of memory per chip, compared with the 384 gigabytes Nvidia has specified for its upcoming Rubin architecture. Four to eight Asimov chips combine into a system called Titan, which Positron says is designed to run models with more than 16 trillion parameters and context windows beyond 10 million tokens.

Sohmers, Positron's chief technology officer, has said the company's current hardware already delivers roughly two to five times the performance per watt and per dollar of comparable Nvidia systems on inference workloads — a claim independent benchmarking has not yet verified at scale, and one that will be tested directly once Asimov ships against Nvidia's own next generation of chips.

Investors point to memory as the real bottleneck

The new investors framed their bets around that memory-first argument rather than around Positron unseating Nvidia outright. Dylan Patel, SemiAnalysis Capital's founder, said in the company's announcement that "Positron's architecture addresses the real constraint, memory, without depending on HBM or advanced packaging" — a reference to the costly high-bandwidth memory and packaging techniques that have become a bottleneck across the AI chip industry. Andra Capital's Paul Tuan offered a similar assessment, saying the company's approach addresses "the performance, power and deployment constraints of modern AI systems."

"Positron is making the boldest memory-first bet in AI hardware. Atlas is running at scale inside Oracle's cloud today," said Forest Baskett, a partner at NEA, in the funding announcement.

Liberty Global's Bobbie Maltiel tied the investment to the broader economics of AI deployment rather than to Positron specifically, saying that "the ability to deliver token usage at the lowest possible cost is key to the full benefits being delivered for society." Positron chief executive Mitesh Agrawal, who joined the company last year after roles at Lambda, said the new capital was aimed at compressing Positron's product cycle: "Speed matters in this market, both in how quickly we ship new generations of silicon and in how quickly they reach customers."

A crowded field betting against Nvidia's grip

Positron is one of a growing list of well-funded challengers — alongside companies such as Groq, Cerebras and d-Matrix — trying to carve out a share of the inference market as it grows faster than the training market that first built Nvidia's dominance. What distinguishes Positron's pitch, and much of the investor commentary around this round, is a narrower focus: rather than competing across all AI workloads, the company is wagering that memory capacity and bandwidth, not raw processing throughput, will decide who wins the economics of running large models in production. The round, first reported by Verdict and other outlets on September 10 and 11, pushes Positron's valuation to $5 billion even though its flagship Asimov chip has not yet left the design stage, underscoring how much investor capital is chasing potential alternatives to Nvidia in AI infrastructure.

That remains the central uncertainty hanging over the round. Positron's revenue and customer base beyond Oracle Cloud Infrastructure, Cloudflare and the AI platform Parasail have not been disclosed, and the company's boldest performance claims apply to Asimov, a chip that will not reach production until the second half of 2027 at the earliest — well after Nvidia's own next-generation Rubin systems are expected to ship. Whether Positron's memory-first architecture holds its advantage against Nvidia's response, and whether the company can convert its now-doubled valuation into contracts at the scale its investors are betting on, will not be clear until Asimov and Titan actually reach customers.