Etched Raises $300M at a $10.3B Valuation, Betting Everything on Transformers Staying Dominant
Source: TechCrunch — 2026-07-23
Summary
Etched, the Cupertino chip startup building Sohu — an ASIC (application-specific chip) that hardwires the transformer architecture directly into silicon rather than running it as software on a general-purpose GPU — closed a $300 million Series C at a $10.3 billion valuation, led by Sequoia with Andreessen Horowitz, SK Hynix, Jane Street, and Diffusion Capital participating. The bet is architectural: by burning transformer math into fixed-function circuits instead of programmable compute units, Etched claims dramatically higher inference throughput than Nvidia's H100 — at the cost of being unable to run anything that isn't a standard transformer.
Key Takeaways
- Sohu is an ASIC, not a GPU: it implements multi-head attention (QKV projection, softmax, output projection), the feed-forward network, and layer normalization as dedicated hardware blocks, so data flows through a fixed physical pipeline instead of being interpreted by programmable compute cores.
- Etched claims roughly 20x the inference throughput of an Nvidia H100 for transformer workloads, attributing it to fixed-function attention circuits and a low-voltage design that sustains around 80% of peak FLOP utilization — GPUs typically run well below that on real workloads because of instruction/scheduling overhead the ASIC simply doesn't have.
- The concrete number: an 8-chip Sohu server is claimed to deliver 500,000 tokens/second running Llama 70B — about 62,500 tokens/second per chip.
- The tradeoff is real and by design: because the transformer computation is physically baked into the chip, Sohu cannot run convolutional networks, diffusion models, mixture-of-experts routing, or state-space-model architectures — it is a bet that standard transformer attention keeps being the thing worth accelerating.
- The $300M round (participants include Sequoia, a16z, SK Hynix, and Jane Street) is a signal that investors are willing to fund architecture-specific silicon now, rather than only general-purpose accelerators, even though it means betting against future architecture shifts away from the transformer.
Reel Script
Hook A chip startup just raised $300 million on a bet that sounds almost reckless: build a chip that can only run one type of AI model — and nothing else, ever.
Core Concept Every GPU you've heard of, including Nvidia's, is a general-purpose machine — it's built to run all kinds of math fast, and then software tells it, step by step, "now do this operation, now do that one." That flexibility costs speed, because the chip spends real energy and time just figuring out and scheduling what to do next. Etched's chip, called Sohu, throws that flexibility away on purpose. It's an ASIC — application-specific integrated circuit — which just means the chip's physical circuitry is the program. Instead of software telling a general chip how to do transformer math — the attention and feed-forward layers that basically every modern LLM is built from — Etched etched that exact math directly into the silicon. Think of it like the difference between a Swiss Army knife and a single fixed blade shaped exactly like the one cut you need to make: the fixed blade is way faster at that one cut, and useless for anything else.
Hands-On Here's what that tradeoff buys them, in real numbers. Etched claims about 20x the inference throughput of an Nvidia H100 on transformer workloads, and says a server built from 8 of these chips can push 500,000 tokens per second running Llama 70B — that works out to roughly 62,500 tokens per second per chip. The reason they can hit that is the fixed-function attention circuits skip the scheduling overhead a GPU pays on every operation, letting the chip reportedly sustain around 80% of its theoretical peak FLOP utilization, where real-world GPU workloads typically fall well short of that. The cost of that speed is architectural lock-in: because the transformer math is physically wired in, Sohu literally cannot execute a convolutional network, a diffusion model, a mixture-of-experts router, or a state-space model — if the AI field's next big architecture isn't a transformer, this chip can't run it, full stop.
Takeaway This is a pure bet on one architecture staying dominant long enough to justify chips that can't do anything else — and Sequoia, a16z, and SK Hynix just put $300 million behind that bet at a $10.3 billion valuation. If you're building inference infrastructure today, the real takeaway is that specialization-for-speed is now a viable business model in AI hardware, not just a research curiosity — but it comes with real architectural risk you should weigh before committing a stack to it.