A startup founded by three Harvard dropouts has just become one of the most valuable young companies in AI hardware — on the strength of a chip that can run only one thing. Etched has closed a $300 million Series C at a $10.3 billion valuation, led by Sequoia, in a bet that a transformer-specific processor can displace Nvidia at the inference layer even as the incumbent's dominance looks unassailable.
The Deal
The round was led by Sequoia — which says the valuation is the highest it has ever backed at the Series C stage — with participation from Andreessen Horowitz, SK Hynix, Jane Street, and Diffusion Capital, alongside earlier investors. There is a notable irony in Sequoia's leadership: the firm once rejected Etched's founders before ultimately backing them at a five-eleven-figure valuation.
The company's ascent has been rapid. Etched was valued at $5 billion in December when it raised $500 million, meaning it has roughly doubled its valuation in about seven months. It only emerged from stealth on June 30, 2026, disclosing $800 million raised to date, a working chip, and more than $1 billion in signed customer contracts. Backers beyond the lead investors reportedly include Peter Thiel, Andrej Karpathy, Dylan Field, and Amjad Masad.
The Sohu Bet
Etched's flagship is the Sohu chip, a transformer-specific ASIC built on TSMC's 4nm process. Where Nvidia's GPUs are general-purpose accelerators capable of running any neural network architecture, Sohu does exactly one thing — run transformer models — and, according to Etched's own benchmarks, does it dramatically faster.
The company's published comparisons are striking. An eight-GPU cluster of Nvidia H100s serves Llama-3 70B at roughly 25,000 tokens per second, and an eight-chip cluster of Nvidia's newer B200 "Blackwell" parts reaches about 43,000 tokens per second. An eight-chip Sohu cluster, Etched claims, hits 500,000 tokens per second — an order-of-magnitude leap that, if it holds up in production, would reset the economics of large-scale AI serving.
Etched now employs around 400 people and is scaling production across a new 80,000-square-foot California facility and a factory in Taiwan, with first rack shipments slated for the summer.
The Economics Driving It
The thesis behind Etched rests on a shift in where AI money is spent. Training a frontier model is a one-time, capital-intensive event; serving it is a perpetual cost that scales with every query. As enterprises move from pilots into production and agentic systems multiply the number of model calls per task, inference is becoming the dominant line item in AI budgets.
That is why a wave of capital is flooding into inference infrastructure — and why a specialized chip that trades flexibility for raw throughput and efficiency could find a lucrative niche. If nearly every high-value model in production is a transformer, the logic goes, why pay for hardware that can run architectures nobody is deploying at scale?
The presence of SK Hynix on the cap table adds another dimension. Memory bandwidth, not raw compute, is the binding constraint for many inference workloads, and a partnership with one of the world's leading memory makers could give Etched an edge in feeding data to Sohu fast enough to sustain its throughput claims. It also signals that established semiconductor players see the specialized-inference thesis as credible enough to back.
Why It Matters
Etched's raise is a concentrated wager on one of the biggest open questions in AI economics: whether specialized silicon can chip away at Nvidia's grip on the accelerator market. Nvidia's moat has always rested partly on generality — its GPUs run anything, backed by the mature CUDA software ecosystem. A transformer-only chip attacks that moat precisely where AI spending is growing fastest.
The core risk is equally clear, and it is architectural. Sohu is hard-coded for transformers. If the research community produces a successor architecture that supplants transformers at scale — a state-space model, an attention-free design, or something not yet named — Sohu's advantage could evaporate overnight. It is a high-conviction, high-fragility bet, and its payoff hinges on the transformer remaining the dominant AI architecture for years to come.
For the broader market, Etched's valuation is a barometer of investor appetite. Even as some analysts warn that AI funding is turning more selective, capital is still flowing aggressively toward companies with real revenue, defensible technology, and a credible path to serving inference at lower cost. Etched has convinced some of the sharpest names in venture that it is one of them. Whether Sohu's benchmarks translate into shipped racks and satisfied customers will determine if this is a landmark bet — or a cautionary tale.
