Positron
Positron designs memory-optimized silicon for AI inference.
- Category
- Deep Tech
- Fund
- NZVC II
- Website
- www.positron.ai
What Positron does
Positron's Atlas appliance is shipping and serves Transformer models up to 500 billion parameters, and a model trained with the HuggingFace Transformers library runs on it without being rewritten. Thomas Sohmers, Barrett Woodside and Edward Kmett founded the company in 2023, and its Asimov chip, due in 2027, carries more than 2 terabytes of memory per chip.
The secret
The binding cost of running an AI model is electricity, not silicon. Optimizing the hardware around memory rather than raw compute cuts the power drawn per token, which is what turns an AI company into an energy company in everything but name.
Why NZVC invested in Positron
We backed Positron because inference is becoming the dominant AI workload and the power it draws is what decides whether serving a model is affordable. Designing the chip around memory rather than raw compute attacks that number directly, and the Atlas appliance is already in production serving models up to 500 billion parameters. Thomas Sohmers, Barrett Woodside and Edward Kmett have been building this hardware since 2023, and every model in the HuggingFace Transformers library runs on it unchanged, which removes the switching cost that usually keeps new inference hardware out.
Positron in the press
- Positron jumps up to the big-league investment circleJon Peddie Research · 6 February 2026
- Positron raised $230 million at a valuation above $1 billionTechCrunch, FinSMEs, Yahoo Finance (Business Wire) · 4 February 2026
- Positron raised an oversubscribed $51.6 million Series AYahoo Finance (Business Wire) · 28 July 2025
- The Silver Lining profiled Positron and its Reno chip operationThe Silver Lining