Etched has turned a narrow chip bet into a $10.3 billion company, but the real story is not Harvard-dropout theater. It is whether AI inference is finally valuable enough for customers to buy hardware built around one dominant workload.
On July 23, Etched said it had raised a $300 million Series C led by Sequoia, with Andreessen Horowitz, Jane Street, Diffusion, SK Hynix, Peter Thiel, and Jump Trading among the backers. Reuters reported that the round valued the San Jose company at $10.3 billion and represented the highest valuation ever for a Sequoia-led Series C. Seven months earlier, Etched was worth about $5 billion. That is the story.
Look at the timing. Etched came out of stealth on June 30, said it had working A0 silicon on TSMC's N4P process, and disclosed more than $1 billion in signed customer contracts. It also said it had opened an 80,000-square-foot, 10-megawatt facility near its San Jose headquarters to expand production and prototyping. For a chip startup, those details matter more than the romance of three founders leaving Harvard.
Etched's original pitch was brutally narrow: Sohu, an application-specific integrated circuit built for transformer inference. The company has said an eight-chip Sohu server can run Llama 70B at more than 500,000 tokens per second, compared with about 23,000 tokens per second for an eight-H100 Nvidia system. Those are Etched's own figures. Independent benchmarks still aren't public, so you shouldn't treat the 20x gap as settled fact.
But you also shouldn't ignore it.
The chip bet is narrower than the valuation
The logic is straightforward. If most high-value AI workloads keep running on transformer-heavy models, a chip that cuts away general-purpose GPU flexibility can be faster and cheaper on the work customers actually need. Nvidia's H100 and H200 GPUs dominate because they can do many things. That strength has a cost. Etched is betting customers won't keep paying for all that flexibility when the bill they feel every day is inference.
This is still a hard bet. In its June 30 announcement, Etched said its inference systems are running models including DeepSeek, Qwen, Mamba, and Llama, and are designed to support models of many shapes and very large parameter counts. That is a more flexible message than the old transformer-only line. The risk hasn't disappeared, though, because Etched's performance argument still depends on the market staying concentrated around workloads its silicon handles unusually well.
That is where Nvidia remains so difficult to attack. The company doesn't only sell chips. It sells CUDA, supply relationships, networking, systems, developer trust, and a procurement default that every AI infrastructure buyer already understands. If you're building an AI product and your model changes next quarter, Nvidia gives you room to move. Etched has to prove that its speed and cost gains are large enough to make that freedom feel overpriced.
Frankly, that is the only version of the story worth taking seriously. Not the dorm-room version. Not the rejection-and-redemption version. The question is whether inference has become a big enough market for specialized hardware to stop looking like a dangerous constraint and start looking like the obvious economic choice.
Sequoia's reversal is useful, but not magical
Gavin Uberti, Chris Zhu, and Robert Wachen founded Etched in 2022 after leaving Harvard. Sequoia's own July 23 note said that when Gavin, Chris, and Rob were still in their dorm rooms, building an inference-specific system looked deeply contrarian. Tech Funding News reported that Sequoia had passed on the founders before leading this round. That reversal is neat. Don't make it folklore.
The better signal is who joined and why. SK Hynix is not a tourist in this market. Jane Street and Jump Trading know what custom systems and low-latency infrastructure cost when they fail. Andreessen Horowitz published its own argument that inference is becoming the largest market in AI, and Sequoia wrote that Etched has production-ready custom silicon and is ready to ship in 2026. Investors can be wrong, but this is not a soft-consensus SaaS round where everyone is buying the same spreadsheet.
MarketScale reported before the Series C closed that Etched had been in talks around back-to-back financing that could point toward a $20 billion valuation. Treat that carefully. A target number is not a signed financing, and venture markets love future valuation talk when the current round is still fresh. Still, it tells you how quickly backers are trying to frame Etched as more than another Nvidia challenger.
The company now has the problem every serious hardware startup eventually earns. It has to manufacture, deploy, support customers, and prove its claimed throughput outside its own materials. A $10.3 billion valuation doesn't make that easier. It raises the bar. The first racks and the first public customer evidence will matter more than the next investor quote.
Also read: Why your startup sales deck keeps losing enterprise deals • Encord is collecting brain-wave data to solve physical AI's training data crisis • Hermes Agent crosses 214,000 GitHub stars as developers abandon commercial AI agent frameworks
