Gavin Uberti
Co-founder and CEO of Etched, the semiconductor start-up building Sohu — a chip etched for one job, running the transformer models behind modern AI faster and cheaper than a general-purpose GPU. Uberti dropped out of Harvard with his co-founders to build it, and became one of the youngest chief executives in semiconductors, raising the company’s first $100m round in his early twenties.
Before Etched, Uberti wrote low-level GPU code — kernels, the hand-tuned routines that squeeze performance from raw silicon — from the age of 17, at companies later bought by Apple and Nvidia. That work taught him the lesson the company is built on: the maths of AI is easy, but moving data around and between chips is the true bottleneck. He pairs it with a competitor’s instinct forged in high-school robotics, where a two-person team he ran ignored documentation and outreach to build only the highest-scoring machine — and held a world record.
Core positions
- Inference — serving tokens to users, not training models — becomes the largest market on Earth, and whoever produces the most tokens becomes the most valuable company. Every design choice reduces to token capacity online per watt.
- The winning AI chip is specialised, not general. Fixing the architecture to the transformer strips out the ‘buffer’ a general-purpose chip carries at every layer and reclaims performance no flexible design can match.
- The chip being existential to the company is a feature. Nvidia builds the best chip because it builds only that chip; hyperscalers’ in-house silicon is a side-project that will never fail their business, and so never commands the same intensity.
- Physics, not software, is where Etched competes: low-voltage inference to beat thermal throttling, prefill/decode disaggregation, and ‘cluster-scale memory’ — a custom interconnect that lets thousands of chips share bandwidth as one pool.
- Velocity, extreme vertical integration, and ‘assume it is possible’ as an operating stance — pre-build everything that can be built before silicon returns, and treat production itself as the product.
In the wiki
| Episode | Source | Date |
|---|---|---|
| Gavin Uberti on Etched, Transformer ASICs, and Challenging Nvidia | Invest Like the Best | 30 June 2026 |
See also
- Bitter Lesson — the compute-versus-ingenuity principle that a fixed-architecture ASIC bets against.
- Token Economics — tokens per watt, tokens per dollar, and the economies of scale in token-making.
- Gavin Baker on AI Infrastructure, Power Constraints, and Semiconductor Investing — a fellow Invest Like the Best guest on the same wafer-to-watt stack.
- Patrick O'Shaughnessy — host of Invest Like the Best.