Etched's $21B Bet: Inference Silicon Is Now a One-Trick Wager

A chip startup that builds silicon for exactly one kind of mathematics became a $21 billion company in August 2026. Thirteen months earlier it was worth $5 billion. That pace is not normal, even for AI.

Etched is the name.

The bet is narrow.

The bet: hard-wired transformers

Etched was founded in 2022 by three Harvard dropouts. Its first product, Sohu, is an application-specific integrated circuit (ASIC) built to run one thing: the transformer computation at the heart of large language models. A GPU is a general programmable machine. Sohu bakes the transformer math directly into silicon, with no operating system, no instruction set, and no room for anything else.

The pitch is throughput per dollar. Etched says one eight-chip Sohu server delivers more than 500,000 tokens per second on Llama 70B and replaces roughly 160 Nvidia H100 GPUs. The claim is unverified. No independent benchmark has been published, and the company only returned first-pass silicon from TSMC in 2026.

That distinction matters for any buyer.

Why a quant fund wrote the cheque

Jane Street led the round. It also took delivery of Etched's first rack-scale system in July and runs it inside its own datacenter. The same institution is the largest new investor and the first production customer.

We tested the chip and are pleased with the early results. Etched's unique approach to inference delivers the precision we will need to support our most demanding workloads. We're excited to now have our own rack running in our datacenter.— Jane Street, statement in Etched's 18 August 2026 funding note

When the buyer is also the investor, signal quality changes. Jane Street's returns depend on getting answers fast and accurately on its own hardware. If the chip had failed the shakedown, the round would not have closed at this size.

Inference is where the money moves

Training a model gets the headlines. Running it, over and over, is where the permanent compute bill lives. Deloitte projected inference would account for roughly two-thirds of all AI compute in 2026. Custom ASIC shipments are growing about 45 percent this year, against 16 percent for GPUs.

Nvidia still holds between 80 and 85 percent of the data-center AI accelerator market, down from 92 percent in 2023. The gap is not closing because challengers beat Nvidia on share. It is closing because hyperscalers and specialist buyers want options, and a one percent slice of a market this large is a real business.

What happens if transformers fade

Etched's cost advantage exists because the chip refuses to do anything but transformers. That is a bet that the dominant architecture of the last decade stays dominant. It is a strong bet. Dense transformers still power most commercial models, including GPT-4 and Claude.

It is not a safe bet. Mixture-of-experts (MoE) models activate only part of their parameters per token, which needs memory access patterns a fixed-function circuit handles poorly. Most major labs have already shipped MoE systems. Etched has noticed. Its August messaging quietly widened from "transformer-only ASIC" to "frontier inference clusters" that, per the company, now run large MoE models and some non-transformer designs.

That reframing is the interesting part. A pure transformer ASIC creates lock-in a serious CIO must model over a three-to-five-year hardware life. A cluster that co-designs chips, memory, and interconnect around large MoE and long-context workloads, while still accommodating other designs, is a more durable wager.

We made this point before. As we wrote in August on where the AI infrastructure value sits, the stack is splitting into layers investors can own separately. And in a piece on ultra-low-power AI silicon, the same pattern appeared: specialized hardware wins on efficiency only when the workload is stable enough to justify giving up generality.

How long before Nvidia's next generation closes the gap?

Development scenarios