Etched just raised $300 million at a $10.3 billion valuation. The round was led by Sequoia and included Andreessen Horowitz, SK Hynix, and Jane Street. Seven months ago, the company was worth $5 billion. Before that, its seed round valued it at roughly $34 million. That trajectory is unusually steep, and it tells you something about where serious money thinks the AI hardware market is going.
The company's product is called Sohu, a chip built to do one thing: run AI models as fast as possible. When you use ChatGPT, Claude, or any similar tool, the response you receive is generated by a process called inference. It is computationally expensive, it happens billions of times a day, and the cost of doing it efficiently is one of the biggest operational concerns for any company relying on AI. Etched's entire pitch is that a chip designed specifically for that task will always beat a general-purpose chip on both speed and cost.
The performance claims are aggressive. Etched says a single server with eight Sohu chips can process 500,000 tokens per second running one of Meta's widely used models, and that this outperforms 160 of Nvidia's best chips while using less power and space. For context, those 160 Nvidia chips would cost somewhere between $4 million and $5 million at current market prices. If Etched's numbers hold up in real-world use, the economics are compelling. The problem: those numbers remain self-reported. No independent benchmarks exist yet, and the first full systems are only now being tested by early customers.
This is not a reason to dismiss the company. The chip was manufactured by TSMC, the world's leading chipmaker, and it worked correctly on its first manufacturing run, which is a notable engineering achievement. Andrej Karpathy, Geoffrey Hinton, and engineers from OpenAI have reportedly tried the hardware directly and invested. These are not people who invest based on slides.
The broader market context helps explain the valuation pace. The inference market is projected to exceed $50 billion in 2026, growing faster than the training market for the first time. Barclays estimates Nvidia will capture only around half of inference computing over the long term, leaving the other half worth roughly $200 billion in annual chip spending by 2028 open to challengers. That is the prize Etched and a handful of other startups are positioning for.
Nvidia is not standing still. Its newest chip generations deliver meaningfully faster inference than previous ones. It has a deep software network that companies have spent years building around, and switching away from it involves real cost and friction. Etched will need its performance advantages to be large enough to justify customers changing their entire software toolchain.
For business operators, the practical takeaway is not about buying Etched's chips: they are not publicly available yet. The takeaway is about what the cost of AI processing is doing. Every company that uses AI tools pays for inference, whether directly through API fees or indirectly through cloud computing bills. Competition in specialized chips is already pushing those costs down. The cost of running AI models has dropped roughly 1,000 times over three years, and that trend has further to go as more purpose-built hardware comes to market.
Etched is also pursuing what appears to be a second, much larger funding round that could push its valuation toward $20 billion. That round has not closed. If it does, Etched would be one of the most valuable private chip companies in the world before shipping a single product at commercial scale. That is either a sign of exceptional technology or exceptional enthusiasm in a very hot market. Probably some of both.