Etched Doubles Valuation in 7 Months: The $10.3 Billion Bet on a Transformer-Only Chip

Create a chip that does only one thing — run transformer-based AI models — and ignore everything else. That is the radical bet of Etched, a startup founded in 2022 by three Harvard dropouts. It just closed a $300 million round led by Sequoia, giving it a $10.3 billion valuation — double what it was worth in December.

The number is striking, but what drives it is not a new versatile GPU. Instead, it is an ASIC — an application-specific integrated circuit — designed exclusively for transformer inference. If the industry moves to another AI architecture, the chip loses its reason to exist. For now, heavyweight investors like Andreessen Horowitz (a16z), SK Hynix, Jane Street, and Diffusion Capital are willing to take that risk.

What Etched Is Selling

The startup has roughly 400 employees, chips manufactured by TSMC, and its own 2 MW data center. Rather than competing directly with Nvidia's GPUs on training, Etched focuses on inference — the moment a trained model generates responses for users, as with ChatGPT.

According to the company, its processor delivers far superior performance compared to traditional GPUs on this specific task, with lower power consumption. The promise is to drastically cut the cost of running large language models, one of the sector's main financial bottlenecks. Etched also claims to have $1 billion in reserved orders and calls the round the largest Series C ever led by Sequoia — assertions that come from the startup itself, with no external validation.

Why Investors Are Betting So High

Money came in fast. In December, Etched had raised $500 million at a $5 billion valuation. Seven months later, that number doubled. For venture capital firms, the logic is clear: if the transformer architecture continues to dominate generative AI — and all signs point that way — a custom-designed chip for it could capture a massive share of the inference market.

The thesis gains strength from demand for models like GPT, Gemini, and Claude, all based on transformers, and from pressure to find alternatives to Nvidia's near-monopoly. The entry of strategic investors such as memory giant SK Hynix also signals that Etched is building a serious supply chain. But the bet has a structural vulnerability: it works only as long as the transformer remains the standard.

A Chip Held Hostage by One Architecture

Etched's ASIC is built to execute specific operations from the architecture introduced by the 2017 paper "Attention Is All You Need." This makes it extremely efficient for that family of models, but also renders it completely useless if a new paradigm emerges — whether based on state-space models, mixture-of-experts with unmappable operations, or any architecture that abandons attention as its central mechanism.

That dependence contrasts with the flexibility of GPUs, which handle multiple workloads, and even with other ASICs targeting broader AI segments. Etched has no Plan B: the product lives and dies with the transformer. So far, the market has accepted this concentrated risk. But tech history is full of standards that got replaced faster than expected.

What Remains Unproven

The performance metrics backing the valuation come from the company itself. There are no independent tests, third-party benchmarks, or publicly disclosed customer validation. The gap between what a prototype shows under controlled conditions and what a chip delivers at data-center scale is wide — and so far, Etched has only shown the former.

Manufacturing at TSMC puts the project on a real industrial track, but the startup still needs to prove it can deliver functioning chips in production with consistent yield, and that reserved orders will turn into firm contracts. The $10.3 billion valuation, therefore, is a measure of expectation — not results.

With the Series C cash, Etched plans to expand its team and infrastructure. If it can prove that a chip dedicated solely to transformers is the key to cheaper inference, the bet may pay off. If not, the company will have been a brilliant experiment that the market priced too early — and too expensively.