Positron AI

Founded in spring 2023 by Thomas Sohmers (Thiel Fellow, CTO), Edward Kmett (Chief Scientist), and Barrett Woodside (VP of Product) in Reno, Nevada, Positron AI is building a vertically integrated inference hardware stack to challenge Nvidia on the specific workload GPUs were never designed for: autoregressive LLM token generation. The company's flagship Atlas system is manufactured in the United States (Arizona) and claims to use less than a third of the power of an H100 while matching its throughput — positioning it for cost-sensitive, power-constrained inference deployments across cloud providers, networking, gaming, and content moderation. In February 2025 former Lambda COO Mitesh Agrawal joined as CEO to lead commercial scaling. The company raised a $23.5M seed in February 2025, a $51.6M oversubscribed Series A in July 2025 (led by Valor Equity Partners, DFJ Growth, and Atreides Management), and a $230M Series B in February 2026 at a post-money valuation above $1B — achieving unicorn status 34 months after founding. The Series B was co-led by ARENA Private Wealth, Jump Trading, and Unless, with strategic participation from Arm and the Qatar Investment Authority. Proceeds fund tape-out of Asimov, Positron's next-generation custom silicon targeting five times the energy efficiency of Nvidia's Rubin architecture, with production slated for early 2027. The follow-on Titan system will offer up to 2 TB of directly attached high-speed memory per accelerator, enabling single-system inference of models up to 16 trillion parameters.

Positron is one of the few post-seed AI chip startups that has shipped silicon to production data centers at scale, not just taped out a demo chip. Its $230M unicorn round — co-led by quantitative trading firms with deep inference cost sensitivity — suggests real commercial traction, and its Asimov roadmap directly targets Nvidia's next Rubin generation on efficiency metrics that matter most at inference scale: tokens per watt and tokens per dollar.

positron.ai AI AcceleratorsUSAprivate

At a glance

Total raised

$305.1M

last round series b

Headcount

—

no public estimate on record

Headquarters

Reno

USA

Founded

2023

3 years active

Status

Private

AI Accelerators

Latest valuation

$1B

Series B · 2026

For subscribers

Subscribe to see the full profile

Including Funding rounds, Bull / Bear thesis, Stock + earnings, Roster changes, Patents, News, and Open roles.

Founding

50% off

/month

91 of 100 spots left

Full

$10

/month

Available once all 100 Founding Member spots are claimed.

Get full access

Already subscribed? Sign in →

Funding history

Total raised

$305.1M

Rounds tracked

Latest stage

Series B

Latest valuation

$1B

Date	Stage	Amount	Valuation	Lead investors
Feb 4, 2026	Series B	$230M	$1B	ARENA Private Wealth, Jump Trading, Unless
Jul 28, 2025	Series A	$51.6M	—	Valor Equity Partners, Atreides Management, DFJ Growth
Feb 11, 2025	Seed	$23.5M	—	Valor Equity Partners, Atreides Management

How they make money

Hardware + recurring servicesPattern

Positron's primary revenue source is the direct sale of its Atlas rack-scale inference appliance to cloud providers, data centers, and enterprises running large language model inference at scale. Hardware is manufactured in Arizona and sold as a full system rather than individual chips, positioning it as a capital purchase decision for operators seeking to reduce inference cost per token and power consumption versus Nvidia H100-class GPUs. The company offers a "Test Flight" program — free dedicated access to Atlas servers hosted in Positron's own engineering facility — as a low-friction on-ramp to conversion, consistent with a high-touch enterprise hardware sales motion. Recurring revenue layers sit on top of hardware sales in the form of software support, model-optimization services, and field application engineering. Positron employs engineers who work directly with customers to tune model deployment and optimize performance, a service model typical of specialized silicon vendors. While Positron has not publicly detailed formal subscription tiers, the combination of proprietary inference software stack (required to get peak performance from Atlas), ongoing model-compatibility updates as new frontier models release, and the upcoming Asimov/Titan hardware refresh creates a natural upgrade cycle and ongoing services relationship with each customer. The Series B proceeds are earmarked primarily for Asimov tape-out and manufacturing scale, indicating hardware margin — not SaaS — remains the dominant economic engine through at least 2027.

Generated by editorial-sourced · June 4, 2026Analytical summary · not investment advice

Founder backgrounds

Thomas Sohmers

Co-founder & CTO

Thomas Sohmers began his computing career at 13 as a researcher at MIT's Institute for Soldier Nanotechnologies, where he focused on high-performance distributed computing and augmented reality systems. Awarded a Thiel Fellowship in 2013 at age 17 — which he used to found REX Computing, a startup designing low-power processors for mobile base stations and HPC applications — he subsequently built out hardware for cryptocurrency ASICs before joining Lambda as Principal Hardware Architect, where he architected much of the Lambda GPU Cloud, now one of the largest pure-play GPU clouds in the US. He then joined Groq as Director of Technology Strategy before co-founding Positron AI in spring 2023. Sohmers was named to the Forbes 30 Under 30 Energy list in 2015 and holds no traditional university degree, having pursued his career entirely through hands-on engineering and entrepreneurship.

Edward Kmett

Co-founder & Chief Scientist

Edward Kmett holds an M.S. in Computer Science from Eastern Michigan University (2005–2006) and is one of the most prolific contributors in the Haskell open-source ecosystem, having authored or maintained over 300 libraries on GitHub and served as founding chair of the Haskell Core Libraries Committee. He serves on the boards of the Topos Institute and the Haskell Foundation, and his research interests span category theory, type theory, and programming-language safety. Before co-founding Positron AI in April 2023, Kmett held senior roles at Groq as Head of Technology and Architecture and later Senior Technical Advisor, and previously worked at S&P Global Market Intelligence (where he designed the Ermine programming language), Digital Asset (smart-contract language engineering), and the Machine Intelligence Research Institute. His deep background in compiler design, hardware-software co-design, and formal methods informs Positron's inference-optimized ISA and software stack.

Barrett Woodside

Co-founder & VP of Product

Barrett Woodside co-founded Positron AI in 2023, contributing early product strategy and go-to-market direction for the Atlas inference appliance. Prior to Positron, he worked at Scale AI where he helped launch the Physical AI team and drove go-to-market for Scale Spellbook, Scale's first LLM inference and fine-tuning product, and earlier held roles at Intel and SigOpt. He studied economics and international studies and holds additional data-science credentials. Woodside subsequently departed Positron and, as of 2026, holds a growth and strategy role at Opal Security.

Mitesh Agrawal

CEO (joined February 2025)

Mitesh Agrawal joined Positron AI as CEO in February 2025, having spent eight-and-a-half years as COO at Lambda, the AI cloud company, where he helped scale revenues from $500K to nearly $500M in annualized run-rate and participated in raising over $1B in capital. He holds a B.S. in Mechanical Engineering from Georgia Institute of Technology, an M.B.A. from UC Berkeley's Haas School of Business, and a master's-level credential from Stanford University. His decision to leave Lambda — a company deeply integrated into the Nvidia GPU ecosystem — for Positron AI was widely interpreted as a strong signal of commercial opportunity in the purpose-built inference silicon market. At Positron, Agrawal leads commercial scaling and has overseen the company's Series A ($51.6M, July 2025) and Series B ($230M, February 2026) rounds.

AI bull / bear

Bull case

Positron is one of the very few post-seed AI chip startups to have shipped silicon to production data centers at scale rather than a demo tape-out — Atlas is already deployed by cloud providers serving gaming, content moderation, and token-as-a-service workloads. That real-revenue proof point, combined with the composition of its Series B (co-led by quantitative trading firms Jump Trading and ARENA Private Wealth, who have direct, acute sensitivity to inference cost per token), strongly suggests genuine commercial traction rather than speculative hype. The Asimov roadmap targets five times the energy efficiency of Nvidia's Rubin architecture at a time when inference electricity costs are becoming a first-order constraint for large-scale LLM operators, giving Positron a durable wedge as token volumes compound. The company's US-manufactured angle (Arizona fabs) opens a meaningful government and regulated-enterprise procurement channel that offshore silicon cannot easily access. Achieving unicorn valuation 34 months after founding on $305M raised — with Arm and the Qatar Investment Authority as strategic participants — provides both validation and the balance sheet to fund Asimov tape-out through 2027 without another dilutive round. If Asimov delivers on its efficiency roadmap and Titan's 2 TB per-accelerator memory enables single-system inference of 16T-parameter models, Positron could become the default inference appliance for frontier-model operators.

Bear case

The AI chip graveyard is littered with startups that shipped promising first-generation silicon but could not build the software ecosystem — drivers, quantization libraries, model-serving frameworks, CUDA-equivalent toolchains — that makes hardware sticky. Nvidia's moat is primarily software (CUDA, cuDNN, TensorRT), not transistors, and Positron has not publicly demonstrated a comparable developer ecosystem at scale. Customer concentration risk is high: losing one or two large cloud customers before Asimov ships could strain the runway, and the window between Atlas and Asimov's 2027 production date is a prolonged period of competitive exposure. Broader macro risk is also material: if the LLM inference market bifurcates toward hosted hyperscaler APIs (AWS Trainium, Google TPU, custom Microsoft silicon), demand for third-party inference appliances could remain a niche. Positron's $1B+ valuation implies significant revenue multiples that require both Asimov to arrive on schedule and the inference-appliance TAM to materialize at the scale the market currently projects — neither of which is guaranteed.

Generated by editorial-sourced · June 4, 2026Refreshes quarterly · see disclosures

In the news

Positron AI · last 30 days

No articles ingested yet for Positron AI. Once the hourly news pipeline is live, every article the classifier tags as mentioning this company appears here with its one-line AI summary and sentiment.

Watch

Leadership

Thomas Sohmers

Co-founder & CTO & CTO & Co-founder

Edward Kmett

Co-founder & Chief Scientist & Chief Scientist & Co-founder

—

Barrett Woodside

Co-founder & VP of Product

—

Mitesh Agrawal

CEO (joined February 2025; former COO of Lambda)

—

Products

Atlas
Positron's first-generation LLM-inference accelerator — billed as the world's first purpose-built inference appliance rather than a repurposed graphics chip. Manufactured in the United States (Arizona), Atlas uses less than one-third the power of Nvidia's H100 while delivering comparable throughput for autoregressive token generation. It ships as a full rack-scale system and is already deployed by production data centers and cloud providers serving gaming, content moderation, and token-as-a-service workloads.
Titan (announced)
The next-generation inference system powered by Positron's forthcoming Asimov custom silicon. Titan is designed to offer up to 2 TB of directly attached high-speed memory per accelerator, enabling single-system inference of models up to 16 trillion parameters. Tape-out of Asimov is targeted for late 2026 with production expected in early 2027.