Etched Raises $300M for a Chip That Only Runs Transformers

What you need to know

• The round closed on 23 July 2026. $300 million Series C at a $10.3 billion valuation, led by Sequoia Capital, with a16z, Jane Street and SK Hynix participating. Reporting described it as Sequoia's highest-valued Series C.
• Etched exited stealth in late June 2026. Roughly $800 million raised in total, and over $1 billion in signed customer contracts already booked.
• The product is Sohu, a transformer-only ASIC. Fabricated on TSMC's N4P process, with 144 GB of HBM3E per chip and — on Etched's own published figures, which nobody independent has verified — roughly 1.8 times the memory bandwidth of an H100 SXM5.
• The performance figures are vendor claims. No third-party benchmark organisation has published independent throughput measurements from physical Sohu hardware under production conditions, and the chip has not shipped in volume.
• The consequence for most builders is pricing, not procurement. Credible alternative silicon is what keeps GPU rental rates falling. You will feel this in your token bill long before you see the chip.

The round, and the investor signal underneath it

The headline number is $300 million, but the interesting number is $10.3 billion for a company whose product is not yet shipping in volume. Sequoia led, and the round has been reported as the firm's highest-valued Series C — a distinction that says as much about how the inference market is being priced as it does about Etched.

One detail in the history is worth pausing on: by widely reported accounts, Sequoia is backing founders it had previously passed on. That is not gossip, it is a proxy for how fast the thesis moved. A transformer-only ASIC is an easy pitch to decline while the architecture might change next quarter, and a hard one to decline once the constraint has shifted from "can we train it" to "can we afford to serve it".

The participant list reinforces the reading. SK Hynix is a memory supplier, and an inference ASIC lives or dies on high-bandwidth memory availability. Jane Street is a trading firm with a genuine, unglamorous appetite for low-latency inference of its own. These are not tourist cheques.

The $1 billion in signed contracts, booked before the round, is the number that actually underwrites the valuation. It is also the number that tells you what the bet is, which we will come to.

What a transformer-only ASIC actually buys

A GPU is a general-purpose machine. It spends a real fraction of its die and its power budget on being able to run anything — scheduling, control logic, flexible datapaths, a memory hierarchy designed for workloads nobody has written yet. That generality is why the same silicon serves a recommender and a diffusion model, and it is also why utilisation on real inference workloads is often disappointing.

Sohu deletes the generality. The transformer attention mechanism is hard-coded into the chip as fixed-function logic rather than expressed as instructions running on general-purpose compute. Etched says this lets it use roughly 90% of the chip's FLOPS. If you have ever profiled a serving stack and watched a very expensive accelerator sit at a fraction of its theoretical throughput, you understand immediately why that figure is the whole pitch.

Pair that with 144 GB of HBM3E per chip and, again on Etched's own unverified figures, roughly 1.8 times the memory bandwidth of an H100 SXM5 — which carries 80 GB and about 3.35 TB/s — and the shape of the argument is clear. Inference at scale is a memory-bandwidth problem wearing a compute-problem costume. More bandwidth, less idle silicon.

And what it trades away

Everything the fixed-function pipeline was not designed for.

This is where an even-handed reading matters more than enthusiasm. The single most valuable property of a GPU over the last decade has not been its peak FLOPS. It has been the ability to run an architecture nobody had heard of eighteen months earlier, on hardware bought before that architecture existed. Every serious inference operator has cashed in that option at least once.

The things that break a hard-coded pipeline are not exotic. Mixture-of-experts routing changes the shape of the computation between tokens in a way that a fixed dataflow handles badly. State-space and hybrid layers interleave a fundamentally different operator alongside attention. Novel attention variants — different masking, different positional treatment, different sparsity — are precisely the class of change that a programmable machine absorbs in a kernel rewrite and a fixed-function unit cannot absorb at all.

So read the $1 billion in signed contracts for what it is: a bet, placed by buyers with real workloads, that the transformer block is stable enough through this hardware generation to be worth burning into silicon. That is a defensible bet. Attention has been remarkably durable, and a chip bought in 2026 does not need to survive until 2036 — it needs to pay back over a deployment cycle. But it is a bet, and anyone telling you it is a certainty is selling something.

The numbers Etched is claiming, and why nobody has checked them

Here is the comparison Etched puts forward. Every Sohu figure below, and both of the GPU throughput figures it is measured against, come from Etched's own materials.

Why this matters if you will never buy one

Almost nobody reading this will procure a Sohu. That does not make the round irrelevant to your infrastructure bill, and the mechanism is worth being explicit about.

The inference layer is where the incumbent is actually contestable. Training is defended by an enormous, decades-deep software ecosystem and by workloads that change constantly. Inference is the opposite: a narrow, repetitive, latency-sensitive operation that runs the same handful of architectures billions of times a day. Narrow and repetitive is what ASICs eat. The whole Etched thesis is architectural lock-in in reverse — accepting a hard constraint on what you can run in exchange for not paying the generality tax on every token.

The second-order effect is pricing. Market analysts estimate that H100 on-demand rates have compressed to roughly $1.80–$2.50 per hour across the major providers, with spot dipping to about $0.60–$0.90 per hour off-peak. Those are estimates, not published rate cards, and they move. But the direction is consistent, and credible alternative silicon is one of the main reasons it stays consistent. That is the same dynamic we traced when Crusoe dropped MI300X pricing to $1.71 an hour, and it feeds directly into the arithmetic in our inference cost economics playbook.

It is also why the falling per-token price has not made anyone's bill smaller. Per-token costs drop while total organisational spend rises, because agentic workloads multiply call volume faster than the unit price falls. A single user request that used to be one model call is now a planning call, six tool calls and a summarisation. Cheaper tokens and a bigger invoice are not a contradiction; they are the same trend.

The de-Nvidia-fication pattern this belongs to

Etched is one instrument in a broader piece of music. Qualcomm's $3.9 billion Modular acquisition was an explicit bet against CUDA at the software layer, attacking the portability problem that keeps buyers locked to one vendor. The large labs went a different route and moved workloads onto alternative accelerators directly, as covered in the great TPU migration. And the financial plumbing has caught up: inference chips now function as collateral in their own right, which we looked at in the first AI loan backed by inference chips rather than GPUs.

Four different attacks: software portability, in-house silicon, purpose-built ASICs, and financing structures that lower the cost of buying any of them. None displaces the incumbent alone. Collectively they change what the incumbent can charge, which is the outcome that reaches your budget.

What it means in Bengaluru and in Manchester

The dual-market picture is genuinely different on each side, and neither side gets the chip.

Indian builders drawing on subsidised GPU capacity through the IndiaAI programme are buying against an administered price rather than a market one, set by policy cycles rather than competitive pressure. So the gap between subsidised and commercial rates narrows from the commercial side, not the subsidised one — and any team planning to graduate onto a commercial provider should model that transition on today's falling rates rather than yesterday's.

UK builders on commercial cloud in the London region feel price competition sooner and more directly, but they also carry the full cost of it. For them the relevant question is contract length: signing a long reserved-capacity commitment in a market where the floor is being pushed down by four independent forces is how teams end up paying above spot for two years.

Neither group gets Sohu-class hardware early. Early ASIC supply goes to the large contract holders — the buyers behind that $1 billion of signed commitments. That is how every constrained hardware ramp has worked, and there is no reason this one differs. If you want the benefit, it arrives as a cheaper GPU-hour in AWS Mumbai or London, not as a new server.

That is the right evaluation frame, and it applies to any alternative silicon, not just this one. Running those checks needs one thing you may not have yet: a repeatable load profile drawn from your own traffic, so that when a vendor hands you a throughput figure you have something honest to hold it against. Building that profile is the first half of our vLLM production playbook, and it is worth doing whether or not you ever look at a Sohu.

The bet, stated plainly

Etched has raised roughly $800 million in total and is valued at $10.3 billion on the proposition that the transformer block is stable enough to be worth carving into silicon, and that inference volume is large enough that the resulting efficiency is worth more than flexibility. Over $1 billion in signed contracts says that some serious buyers agree.

The counter-case is not that they are wrong about transformers. It is that the industry has repeatedly found value in architectures that did not exist when the hardware was bought, and a chip that cannot follow the models has a shorter useful life than its depreciation schedule assumes. That risk is real, priceable, and unresolved by any number currently public — because none of those numbers have been independently measured.

Watch for the first third-party benchmark on physical hardware. That single publication will tell you more about whether $10.3 billion was sensible than every press release between now and then. More coverage of how AI infrastructure money is moving sits in our funding news section.