Do we need to re-define ”Chutzpah” = ”Positron?” Mitesh Agrawal, CEO and co-founder of Positron AI, is leading one of the most audacious ventures in the tech world today: building silicon to compete directly with NVIDIA. Backed by over $70 million in funding from DFJ, Valor, Atreides, Oakseed (disclosure: I am an LP) and others, and already shipping product within 18 months of founding, Positron’s bet is that the market for inference—the part of AI that actually generates text, images, code, and video—will soon dwarf training of LLM models, which is NVIDIA's bread and butter. By focusing on efficiency (performance per watt, per dollar) rather than general-purpose GPUs, Positron aims to make AI more scalable, more affordable, and far less power-hungry BY A LARGE MARGIN. With early customers such as Cloudflare and ParaSail (and others), Positron has $10's of millions in bookings and is ramping up to reach $100s millions by 2026. Agrawal’s co-founders bring unusual strengths: Thomas Sohmers, a Thiel Fellow who taped out a silicon chip at age 19, and Edward Kmett, one of the world’s leading Haskell contributors. Together, they are challenging the world’s most valuable company. (BTW this may change my aversion to the whole idea of ”Thiel Fellow” which I used to call ”Thief Fellows.”)

Favorite Quotes

“We are living in an exponential growth of usage.”

“For inference, companies really care about efficiency—performance per watt and per dollar.”

“It took us just 18 months to go from founding to first product and first revenue.”

“The biggest challenge and the biggest fun is convincing customers you’ll be around in 3, 5, 10 years.” (We will be – with a raise of over $70 mio!)

“Ultimately, it comes down to a very unhinged belief that we will deliver faster.”

Key Themes

• Inference is the growth engine of AI: already $100 billion in compute spend this year and growing 5x year over year.
• NVIDIA’s GPUs dominate both training and inference, but they are general-purpose, power-hungry, and expensive—opening room for specialized chips.
• Positron’s edge is focus: custom silicon for inference, shipping fast, and driving down costs to unlock massive adoption.
• Scaling a silicon startup requires not only cutting-edge design but also building credibility with customers and managing capital-intensive production.
• Investors and early customers are betting that Positron can move faster than the giant, even if the ultimate outcome is acquisition.

Key Takeaways

• Inference is where the revenue is: companies pay per query or per generated output, and every cent saved per transaction scales massively.
• Speed to market matters: Positron went from founding to revenue in 18 months, compared to 4–5 years for most chip startups.
• Customer trust hinges on two things—performance today and confidence that Positron will still be shipping chips five years from now.
• Capital allocation is overwhelmingly R&D-driven (70%+), but profitability on sales helps fund growth.
• The organizational challenge ahead is maturing management and operations while retaining the boldness to outpace NVIDIA.