Etched shipped its first rack. Not a demo. Not a paper benchmark. A working Sohu rack, running transformer inference in Jane Street’s live infrastructure. On Tuesday, the transformer-only chip startup announced $700 million in new funding at a $21 billion valuation — and confirmed that its chip is already earning its keep with one of the world’s most analytically demanding compute buyers. For developers who build inference-heavy AI applications, this changes the conversation from “what if” to “when.”
What Happened #
Etched’s Series D closed at $21 billion — a valuation that doubled in a single month after its $10.3 billion Series C in July. Jane Street, the quantitative trading firm, led the round and became the company’s first customer, actively running the Sohu chip in production. The round also includes Kleiner Perkins, Sequoia, Andreessen Horowitz, Peter Thiel, Bain Capital Ventures, and Blackstone.
The valuation velocity is worth noting: $5 billion in December, $10.3 billion in July, $21 billion in August. The inflection point was Jane Street’s validation. When a firm with Jane Street’s reputation for rigorous infrastructure evaluation puts its production workloads on your chip, other sophisticated buyers start returning calls. TechCrunch has the full funding breakdown.
The Numbers #
Etched’s performance claims are aggressive, but they follow from a clear architectural argument. The Sohu chip is built on TSMC’s N4P process (4nm) and carries 144 GB of HBM3E memory. What it does not have is programmability. Sohu hardcodes the transformer attention mechanism directly into silicon, achieving over 90% floating-point unit utilization on transformer workloads. Nvidia’s H100, designed for general-purpose AI, manages roughly 30% FPU utilization on the same workload.
The result, per Etched’s own benchmarks:
| Hardware (8-chip server) | Tokens/sec (Llama-70B) | vs. Sohu |
|---|---|---|
| Sohu | 500,000+ | — | | Nvidia Blackwell B200 | ~45,000 | 11x slower | | Nvidia H100 | 23,000 | 21x slower |
These are Etched’s numbers, not independent benchmarks. No neutral third party has published validated Sohu results at scale. That caveat matters — but so does the fact that Jane Street, not exactly known for taking vendor claims at face value, is running it in production.
The Bet — and the Real Risk #
Etched’s thesis is explicit: transformers are here to stay, and every watt spent on generality is a wasted. That bet looks strong in 2026. Every frontier model — GPT-5, Claude 4.6, Gemini 3.7, Llama 4 — is transformer-based. The architecture has held dominance since 2017.
But the risk is not that transformers disappear. The risk is that they evolve into something Sohu cannot serve as well. DeepSeek-V4, one of the most-downloaded open-source models in 2026, is a 671-billion-parameter Mixture-of-Experts architecture. MoE models are technically transformer-derived, but their dynamic routing structure undermines Sohu’s fixed-function advantage. Sohu cannot efficiently serve MoE architectures — and MoE is gaining ground precisely because it offers better efficiency at scale.
There is also the training gap. Sohu is inference-only. Organizations running their own training pipelines still need Nvidia or AMD hardware for that phase. Sohu enters the stack only at serving time. A detailed technical breakdown is available from Peak FLOPS on Substack.
What Developers Should Do Now #
The short answer: watch, don’t wait. There is no public API today. Sohu access is currently enterprise-only, sold at the rack level to organizations with the infrastructure to integrate it. Etched has opened a Sohu Developer Cloud registration for smaller teams and cloud-style access, with availability expected in late 2026 or early 2027.
If you are evaluating whether Sohu belongs in your stack when access opens: Good fit: High-throughput transformer inference — LLM APIs, real-time chat, streaming at scalePoor fit: MoE model serving, diffusion models, training workloads, any non-standard dense transformerNot yet: Production-critical workloads need independent benchmarks and enterprise support contracts that do not yet exist at Nvidia’s scale
The Bigger Picture #
The AI inference market is under pressure. The AI API price war unfolding in August — Anthropic, Google, and OpenAI all cutting inference costs — drives higher volume, which creates demand for exactly the throughput Sohu delivers. Lower prices drive higher usage. Higher usage demands more throughput. Specialized silicon is the logical endpoint of that compression.
Etched’s timing is not accidental. The company is betting that as inference becomes a commodity, the winners will be those who run it cheapest and fastest — not most flexibly. That bet is visible in every benchmark leaderboard, every new model release, and every AI product shipped in 2026. For more on the inference chip wars, Spheron’s Sohu vs. Nvidia breakdown is worth reading.
Whether Sohu disrupts Nvidia’s inference dominance or becomes a cautionary tale about architectural lock-in depends on factors no one can fully control. But the first rack is running. Jane Street paid for it. That’s a different conversation than six months ago.