AI has fundamentally changed the future of computing. We find ourselves in the midst of a massive infrastructure buildout to support a new class of compute-intensive software, with hyperscalers and AI labs investing hundreds of billions in capital per year to develop the next generation of datacenters. Today, these datacenters are filled with massive GPU farms as the computing core. Nvidia GPUs are the gold standard for general-purpose AI computing, well-suited for training large language models (LLMs) but less optimized for the full spectrum of AI inference. As we move into an age of ubiquitous AI, a new class of processors is emerging—purpose-built to power the next era of intelligent computing. This is where Positron comes in.
Today, DFJ Growth is delighted to announce our investment in Positron’s Series A financing and to partner with the company in building a leading platform for AI compute.
AI model training and inference have fundamentally different compute requirements. As the market matures, we believe the underlying hardware layer (a $235 billion market in 2025 according to Morgan Stanley) will diverge and specialize, with discrete training and inference segments emerging. Nvidia continues to be the benchmark for AI training, but its lead diminishes in inference, where performance and cost efficiency are paramount. Over time, as inference makes up the vast majority of compute demand, delivering fast, economical tokens will become increasingly critical. Most notably, Nvidia GPUs are limited in inference performance by their on-board memory bandwidth—the rate at which data can be fetched and output from LLMs. Positron has designed a novel chip architecture specifically to address this memory bottleneck and unlock faster, cheaper, and more power-efficient inference.
The semiconductor industry is typically known for prolonged development cycles. Positron has bucked this trend, delivering working chips in FPGA form to customers within two years of company formation. This is rare and immediately caught our team’s attention. Early deployments have not only allowed Positron to validate its product architecture but also provided a unique wedge into customers’ live data center environments. Their first product is delivering meaningful price/performance benefits relative to Nvidia GPUs, with a custom ASIC in development that is poised to deliver another leap in price/performance.
Building a generational company in any sector requires a world-class team who are driven by a common mission. Positron is led by Mitesh Agrawal (CEO) and Thomas Sohmers (founder and CTO). Together, they bring the leadership and direction needed to build a new leader in AI computing. Mitesh previously served as COO of Lambda Labs, growing the company to over $400M in annual revenue, and is intimately familiar with the AI customer segment and processor landscape. Thomas led the development of several chips before the age of 30 and has over a decade of experience in semiconductors, dropping out of high school to pursue a Thiel Fellowship and later serving as director of technology strategy at Groq. They are supported by a world-class, interdisciplinary team of experts steeped in chip design, machine learning, functional programming, and silicon manufacturing experience.
At DFJ Growth, we often talk about the core building blocks for advancing AI progress: models, data, and compute. Over the past few years, we’ve backed leaders in the first two layers and have been hunting for novel systems in the compute layer that more efficiently serve AI outputs – in other words, a new era processor purpose-built for AI inference. We immediately resonated with Mitesh and Thomas’s vision and approach to solving this pressing challenge.
Renowned sci-fi author Isaac Asimov imagined the positronic brain as the foundation for machine intelligence more than 80 years ago. Mitesh, Thomas, and team are bringing this vision to life by moving us one step closer to a world where AI is affordable and abundant.
