What happened
Etched said Tuesday it raised $800 million and locked in $1 billion in sales contracts for Sohu, its transformer-only inference chip, according to CryptoBriefing. The company didn't break out lead investors or the contract counterparties in the disclosure picked up by the outlet, but pairing the two numbers in a single announcement is the giveaway. Funding rounds at this size usually leak in tranches.
Bundling the order book with the raise tells you Etched wants the market to price the company on revenue visibility, not on slideware. Sohu is an ASIC built for one job. It runs transformer architectures and nothing else.
That's the bet. Drop the generality that makes a GPU useful for everything from gaming to scientific compute, and you free up die area for more transformer-specific math. The pitch to buyers is simple: lower cost per token on the workload that's eating the entire AI capex budget right now.
Why it matters
Nvidia's dominance in AI training is the headline. The quieter, more durable moat is inference, where models run in production every time a user hits a chatbot or an agent fires a tool call. That's where Etched is aiming.
Inference at scale is a margin business. A 30% cost-per-token advantage on a workload that runs billions of times a day is the kind of math that gets a hyperscaler CFO on the phone. The $1 billion in contracts, if it holds up under delivery, is the first credible signal that someone with real budget signed something binding.
The headline looks bullish. The execution risk doesn't go away. ASIC startups have a graveyard behind them.
