Gavin Uberti on Etched, Transformer ASICs, and Challenging Nvidia

Harvard dropout Gavin Uberti has bet an entire chip on a single idea — that inference will be the largest market on Earth, and that a processor hard-wired for the transformer can serve it an order of magnitude better than Nvidia’s general-purpose GPU.

Key ideas

1. Etched builds an application-specific chip — an ASIC, a processor etched for one job rather than programmable for any — that assumes every model worth running is a transformer. Fixing the architecture lets Uberti strip out the general-purpose ‘buffer’ the whole semiconductor stack carries, and reclaim performance no flexible chip can: ‘20% here, 50% there, 2x here’, compounding into a system radically better for inference.
2. Inference, not training, is the prize. ‘We know inference is going to be the biggest market in the world. Whoever produces the most tokens is going to be the most valuable company in the world.’ Every design decision reduces to one question — how to get the most token capacity online per watt.
3. The one-architecture wager is existential, and that is the edge. Google will not fail if its TPU fails, Meta will not fail if MTIA fails, OpenAI will not fail if its own chip fails — their revenue comes from elsewhere. For Etched the chip is the company, which is precisely why ‘the best chip in the world is built by a company that only builds that chip. It’s Nvidia.’
4. The culture is velocity, vertical integration, and parallelisation, run on an ‘assume it is possible’ first principle. The founders shipped a dozen engineers to Bangalore for six months, ran 24-hour handoff cycles, and pre-built the entire rack, software stack, and production line before silicon returned — reaching working inference in 40 days where a rival took ten months.
5. Sohu, the first product, wins on three technical bets: low-voltage inference (running at under half the voltage of any other AI chip, to beat thermal throttling and cram in more flops); prefill/decode disaggregation; and ‘cluster-scale memory’ — a custom interconnect that cuts chip-to-chip latency from Blackwell’s ~4,000 nanoseconds toward a handful, so a whole cluster’s memory acts as one pool.

Summary

The transformer-ASIC bet

Uberti’s founding observation is that every GPU, TPU, and AI accelerator in service was designed before ChatGPT and retrofitted to the workloads that followed. A general-purpose chip must hold ‘buffer’ at every layer — the EDA tools, power modules, boards, standard cells — to serve everything from data centres to edge IoT. Narrow the target to one use case and the constraints dissolve. His simple illustration: EDA tools sign off timing assuming a chip may run at freezing temperatures, but no data centre runs on ice, so a chip that will never drop below 80°C can discard that margin. Repeated across the stack, these reclaimed slivers compound. The deeper bet is architectural fixity: Etched assumes the transformer is the durable shape of intelligence and etches it in, trading the GPU’s generality for performance no flexible design can match.

Inference as the market

The thesis behind the whole company is that inference — serving tokens to users, not training models — becomes the largest market in the world, and eventually a majority of global GDP. Productivity is heading toward ‘agents per megawatt’. Today perhaps one-thousandth of humanity uses paid AI; serving billions concurrently, at the tokens-per-second coding and agentic work demand, requires orders of magnitude more infrastructure ‘from the wafer to the watt’. Uberti’s analogy is manufacturing: tokens today are hand-crafted like Renaissance screws by small general-purpose systems; he wants the economies of scale that made iPhones and cars identical for billionaire and average buyer alike, applied to token-making.

The one-architecture wager

Betting the entire chip on the transformer is the company’s central risk — if the dominant architecture shifts, a fixed-function chip cannot follow a GPU’s software. Uberti reframes the exposure as commitment. Because the chip is existential for Etched and merely optional for the hyperscalers building their own silicon, Etched attracts a calibre of talent, supplier belief, and intensity the side-projects cannot. He points to the raw numbers: the flop density of chips built by labs and hyperscalers trails Nvidia’s B300, because they need only build something ‘similar enough’ to dodge the Nvidia tax, not to win outright.

The startup culture

The operating philosophy is velocity, vertical integration, and relentless parallelisation, drawn by analogy from Uberti’s two-person high-school robotics team that ignored documentation and outreach to ‘just win’. ‘The best vendor is no vendor’: Etched builds chip, board, cold plates, interconnect, rack, and production line in-house — the only start-up building its own rack as well as its own chip. Talent is ‘bimodal’ — world-number-one ‘legends’ (they recruited the founder of Nvidia’s HGX/DGX rack team) paired with naïve first-principles young engineers, working together. The register is all-in: ‘You kind of have to be sick in the head to join our company’ — uprooting a comfortable career to move to San Jose for a pre-product chip company run by two 24-year-olds.

Competing with Nvidia

Uberti does not try to build a better GPU — futile against Nvidia’s TSMC pricing and model co-design. He competes on being different and existential. TSMC’s partnership, he argues, is the real moat: not the technology alone but a level of service — running yield experiments on their own dime — that no steelworks or ordinary supplier offers. Supply he treats as positive-sum, deliberately building Sohu on a different node (4nm) and a different HBM than Nvidia’s Rubin (3nm), so a customer’s choice is not ‘a gigawatt of GPU or a gigawatt of us’ but two gigawatts. His warning to himself: ‘if you have the most performant product and you can’t produce it… then you’re just a podcast’.

What Sohu does differently

The three technical bets are physical. First, low-voltage inference: because power rises with the square of voltage (Dennard scaling), and GPUs thermal-throttle before reaching full flop utilisation, Etched engineered a new power-delivery mechanism running at under half a GPU’s voltage — Bitcoin miners already run at a quarter, so it is physically possible. Second, prefill/decode (PD) disaggregation: separating the memory-loading ‘prefill’ pass from the token-generating ‘decode’ pass — ‘like loading the gun and then firing it’ — and shuttling the KV cache between them. Third, cluster-scale memory: a fully custom interconnect above Ethernet’s second layer that cuts point-to-point latency more than fivefold, so thousands of chips share memory bandwidth as a single pool rather than losing most of an 8x scale-up to hop latency. The nearest-death technical crisis — a clock-domain-crossing bug producing wrong attention results — was solved in two frightening weeks by aligning two clock signals to within 50 picoseconds, two billion times a second.

Speakers

• Gavin Uberti — co-founder and CEO of Etched; the primary voice in the conversation, telling the kernels-engineering and robotics origin story and walking through Sohu’s architecture.
• The co-founder appearing alongside him — addressed throughout as ‘Rob’ (Robert Wachen, Etched co-founder) — supplies the fundraising history and the personal cancer-survival story that opens his account. Chris Zhu is Etched’s third co-founder and does not feature. Because the source is an unlabelled caption transcript, some turns cannot be attributed with certainty between the two founders.

See also

• Gavin Baker on AI Infrastructure, Power Constraints, and Semiconductor Investing — the prior Invest Like the Best episode on the same wafer-to-watt stack, and Baker’s rule that a chip start-up must be ‘different and hard to do’ to survive Nvidia.
• Dylan Patel on the Token Economy, AI Supply-Demand, and the Permanent Underclass — the token economy and inference supply-demand that Uberti’s whole thesis rests on.
• Bitter Lesson — the compute-beats-ingenuity principle a fixed-architecture ASIC bets against: specialising hardware to one architecture wagers that the transformer, not a more general design, is the durable winner.
• Scaling Laws — why ever-larger models and ever-more inference underwrite the demand Etched is built to serve.
• Token Economics — tokens per watt, tokens per dollar, and the economies of scale in ‘token-making’ Uberti is chasing.
• Patrick O'Shaughnessy — host of Invest Like the Best, and an Etched investor who discloses his stake in the conversation.