Three Harvard dropouts. Four years. A $10.3 billion AI chip company.

On July 23, AI chip startup Etched announced $300 million in Series C funding led by Sequoia Capital, with participation from a16z and SK Hynix, at a post-money valuation of $10.3 billion — more than doubling from its $5 billion valuation just seven months prior. Even more striking: according to sources, Etched is already advancing its next round, targeting a $20 billion valuation.

The company's core pitch is simple: we don't make GPUs. We make ASICs purpose-built for AI inference. No GPUs required — any AI model runs, and runs several times faster.

Why ASIC Inference Chips Are Hot Now

To understand why Etched is worth $10 billion-plus, you need to understand a structural shift underway in the AI compute market: training and inference are diverging.

For years everyone fixated on training compute — how many H100s GPT-5 trained on, how big Claude's training cluster was. But starting in H2 2025, a clearer trend emerged: inference compute demand is exploding, growing far faster than training. Why? Agents. An agent executing a task might call an LLM dozens of times; inference token consumption is dozens of times higher than traditional chatbots. When AI products move from "chatting" to "working," inference demand opens like a floodgate.

GPUs are general-purpose compute chips — they can do anything, but aren't optimal for anything. Training LLMs requires general-purpose compute because model architectures keep changing and chips need flexibility to adapt to tensor operations. But inference, especially large-scale deployment, is different: model architectures are relatively stable, latency and throughput requirements are extreme, and that's where purpose-built chips (ASICs) shine — strip out all unnecessary circuitry, throw every transistor at matrix multiplication and memory access, and you can beat GPU energy efficiency by multiple times.

This isn't a new concept. Google's TPU is an ASIC; AWS has Trainium and Inferentia; Microsoft has Maia. But those are all cloud-vendor in-house chips, not sold externally. Etched wants to be the "third-party AI inference ASIC" — selling to any company that needs large-scale inference, going head-to-head with Nvidia GPUs.

Three Harvard Dropouts and the Contrarian Bet

Etched's story fits the classic Silicon Valley narrative. The three founders are all Harvard dropouts — they started the company in 2022 before graduating. The industry consensus at the time: "The AI chip space is oversaturated, Nvidia is unbeatable, startups building AI chips are dead on arrival." The three didn't buy it. They believed GPUs were designed for graphics rendering, not AI, and that purpose-built chips for AI inference were an inevitability.

Sequoia led the $300 million round — reportedly one of the highest-valued Series C rounds Sequoia has ever led. a16z and SK Hynix's participation is also signaling: SK Hynix is the world's largest HBM memory supplier, meaning Etched will have supply security for high-end memory going forward.

Skeptics abound, of course. Nvidia's CUDA moat remains the wall every AI chip startup must scale. Even if your chip is 5x faster than a GPU, if mainstream AI frameworks don't support it, developers won't migrate, and the software stack is unstable, customers won't risk it. Etched claims its chips run any AI model without modification, but that claim needs large-scale customer validation.

One thing is certain: the AI inference market is exploding, and customers are hungry for alternatives to Nvidia. From AMD's Helios to Etched's ASICs, from cloud-vendor in-house chips to open-source inference engine optimizations, the entire industry is trying to loosen Nvidia's grip on inference. Whether Etched succeeds is an open question, but a $10.3 billion valuation shows: capital markets are willing to pay for the "Nvidia alternative" narrative.

After all, in a $1.4 trillion market, even capturing 5% is a $70 billion business.

See you tomorrow.