# Etched Ships First Transformer ASIC Rack at $21B Valuation > Source: > Published: 2026-08-25 10:11:18+00:00 On August 18, Etched did two things simultaneously that most chip startups never manage even once: it shipped its first physical rack to a paying customer and raised $700 million at a **$21 billion valuation**. The customer is Jane Street. The valuation tripled since December. For developers, the headline worth paying attention to is not the funding — it is that a chip engineered to run *only* transformer models is now in production at one of the most operationally demanding computing environments in finance. ## One Chip, One Job [Etched’s Sohu chip](https://www.globenewswire.com/news-release/2026/08/18/3347095/0/en/etched-raises-700m-at-a-21b-valuation-and-completes-first-customer-delivery-to-jane-street.html) makes a stark architectural choice: transformer attention is burned directly into fixed-function silicon rather than being orchestrated in software on general-purpose tensor cores. The chip cannot run diffusion models, CNNs, state space models like Mamba, or mixture-of-experts architectures with dynamic routing. It does transformer inference, and nothing else. Built on TSMC’s 4-nanometer process with 144GB of HBM3 memory per chip, an eight-chip Sohu rack is sized to replace approximately 160 H100 GPUs for transformer inference workloads. Etched claims the eight-chip server delivers **500,000 tokens per second on Llama-3 70B**. For reference, an equivalent H100 configuration produces roughly 23,000 tokens per second, and Blackwell lands around 43,000 to 45,000. That is a claimed 20x advantage over H100 and 10x over Blackwell. These are Etched’s own numbers — no independent benchmark lab has published results from physical Sohu hardware under production conditions. The benchmark configuration also favors Sohu’s architecture: high-batch throughput with long input sequences, not low-latency single-request workloads where GPUs remain more competitive. ## Why Jane Street Matters More Than $700M Chip startups that arrive with impressive benchmark decks and no shipping silicon are common. What is uncommon is a paying customer with a live deployment. Jane Street operates trading infrastructure where latency is measured in microseconds and compute efficiency translates directly into profit. The firm tested Etched’s hardware against real-world workloads before cutting a check — and then led the funding round. That sequence matters. Jane Street did not invest on faith; it invested after testing. The delivery makes Etched’s performance claims significantly less abstract. It does not validate every number, and no developer outside Jane Street’s datacenter has seen independent results. But it does separate Etched from the long list of AI chip companies that raised money on simulations and never delivered working silicon at production scale. [TechCrunch noted](https://techcrunch.com/2026/08/18/etcheds-valuation-doubles-to-21b-in-a-month/) that Etched’s valuation has now tripled in under eight months — $5B in December, $10.3B in July, $21B today. ## The Bet and the Risk Etched’s transformer-only design is the most extreme bet in the current inference hardware market. Groq and Cerebras, the other inference-focused challengers, use SRAM-centric architectures that can run multiple model types. Etched cannot. If the industry meaningfully shifts toward hybrid SSM-transformer designs — architectures that interleave attention blocks with Mamba-style state space layers — Sohu cannot run them. CEO Gavin Uberti acknowledges the risk directly: if transformers fall out of favor, Etched designs a new chip. That is a multi-year design cycle. It is a genuine risk, and most coverage undersells it. The counter-argument has some force. Every frontier model — Claude, GPT, Gemini, Llama — is still a transformer as of August 2026. Mamba-2, the most mature SSM alternative, had to [rewrite its core to use matrix multiplications](https://www.spheron.network/blog/etched-ai-sohu-vs-nvidia-transformer-asic-inference/) because scans cannot saturate tensor cores — which pushes it toward transformer-compatible computational patterns. The $1 billion in customer contracts Etched has secured suggests buyers have made a judgment that transformers are not going anywhere on a timeline that makes Sohu a bad investment. ## When Does This Reach Developers? Right now, Sohu is rack-level hardware. Developers do not interact with it directly. The path to impact runs through cloud providers and AI companies buying racks, deploying them for inference, and offering faster or cheaper API endpoints as a result. Etched says its customer base already includes AI companies and cloud providers, which means the downstream effects on inference APIs are a matter of when, not whether. Inference costs have [already collapsed](https://wafer.substack.com/p/breaking-down-etcheds-sohu) from roughly $20 per million tokens in 2022 to under $0.40 in 2026. Purpose-built silicon optimized exclusively for transformer workloads should accelerate that decline further. Whether Sohu survives long enough to drive it — or gets outcompeted by hyperscaler custom ASICs that are not available on the open market — remains to be seen. What is no longer speculative is whether the hardware exists. It is sitting in Jane Street’s datacenter.