The $10.3 Billion Bet That Transformers Are Here Forever

Building a chip that can only do one thing is normally called a liability. Etched just raised $300 million at a $10.3 billion valuation to do exactly that — and the market called it the most valuable Series C Sequoia has ever led.

The company's Sohu chip hardcodes transformer attention mechanisms directly into silicon. It doesn't run general matrix multiplications. It doesn't support other neural network architectures. It runs transformers, and only transformers, at a throughput no GPU can match for those workloads. Eight Sohu chips replace roughly 160 Nvidia H100s on a Llama 70B inference task, pushing around 500,000 tokens per second. The efficiency advantage comes from eliminating the programmability layer that makes GPUs general-purpose — which also makes them wasteful for a task as structurally predictable as transformer attention.

The round is Sequoia's largest Series C on record. SK Hynix, Andreessen Horowitz, Jane Street, and Diffusion co-invested. Etched came out of stealth June 30 with first-pass silicon success on TSMC's N4P process and announced over $1 billion in signed customer contracts before the ink on the Series C was dry. That's not a company pitching a vision. That's a company closing a production financing round.

The investment thesis embedded in that valuation is worth reading carefully. A transformer-only chip is not a diversified bet on AI hardware. It is a single, high-conviction architectural claim: that the transformer — the attention mechanism at the core of GPT, Llama, Claude, Gemini, and virtually every major AI model in deployment today — becomes permanent infrastructure. Not a stepping stone. Not a dominant-but-replaceable paradigm. The TCP/IP of AI. The $10.3 billion valuation is the market pricing that claim as likely true.

The counterargument is real. State space models, mixture-of-experts architectures, and other approaches that diverge meaningfully from transformer attention are active research directions. If the next dominant AI architecture abandons the attention mechanism, Sohu becomes a very expensive silicon paperweight. Etched is making a bet that the companies who built fortunes on architectural permanence — Arm's RISC instruction set, x86's dominance in PCs — were right, and that AI is following the same pattern.

For founders building at the AI application layer, the Etched story matters less as a chip story and more as an inference cost story. If Sohu ships at scale, the per-token cost of running large language models falls substantially. Applications that were inference-cost-constrained — real-time financial analysis, high-frequency agentic workflows, continuous document processing — become economically viable at price points that weren't possible on GPU-based infrastructure. The chip company wins or loses based on architecture. The application companies win either way.