Etched eyes $20B: specialized silicon sets your token price

Etched is reportedly negotiating two funding rounds at the same time — one at roughly $10 billion, another at roughly $20 billion — for a chip that does exactly one thing: run transformer models. You will never buy one. You will absolutely pay a price shaped by whether it works. Inference silicon is the floor under your token bill, and that floor is being renegotiated right now.

What actually happened

The Wall Street Journal reported on July 17 that Etched is pursuing back-to-back rounds, as covered by PYMNTS: one led by Sequoia Capital around $10 billion, and a second led by existing investor Jane Street that would put the company near $20 billion. Its prior mark was $5 billion post-money, set by a $500 million round that closed in December 2025.

Three weeks earlier, Etched came out of stealth. In its own announcement, the company said it had raised $800 million across previously unannounced rounds and signed over $1 billion in customer contracts. Its chip, Sohu, is an ASIC built only for transformer inference — no general-purpose GPU flexibility, which is the entire bet. Etched says its A0 silicon came back from TSMC's N4P process earlier this year, that it's validating a rack-scale product with customers, and that first racks ship this summer. Bloomberg reported Jane Street and a TSMC-linked venture arm among the investors.

Two rounds at a 2x valuation gap, negotiated simultaneously, is not a detail to skip. That spread is the market saying it does not know what this is worth yet.

Why inference chips matter for your business

Here's the chain nobody draws for you. TSMC makes the wafer. Etched, Groq, Cerebras, and Nvidia turn it into inference capacity. Your model vendor rents that capacity. Your per-token price is the last link. When somebody removes the general-purpose overhead from that stack, the savings eventually reach your invoice — or they don't, and they become somebody's margin.

The specialization bet cuts both ways. A transformer-only ASIC is dramatically cheaper per token while transformers are the architecture. If the frontier moves to something structurally different, that silicon is a very expensive paperweight. That's the risk investors are pricing at $10B versus $20B, and it's why Etched's $1B in contracts is the number that matters more than the valuation.

You don't need a position on any of this. You need your stack to be able to take advantage of it. When inference gets cheap on specialized hardware, it shows up as a new provider with better prices — and you capture that only if switching models is a config change, not a rewrite. We've said this about open weights and about model aliases; it's the same discipline. Route through one abstraction layer. Log cost per completed task, not cost per token. Then when the floor drops, you actually notice.

Sources: PYMNTS — AI Chip Startup Etched Eyes $20 Billion Valuation, Etched — Emerges From Stealth With Working Chip.