Key points
- Michael Burry, one of the loudest Nvidia (NVDA) bears this year, called AI chip startup Etched "serious competition" on Tuesday.
- Etched's Sohu chip claims roughly 10x to 20x an Nvidia GPU on one job, but it can't run the mixture-of-experts models many teams now deploy.
- Etched is reportedly raising at up to $21 billion, about 4x its $5 billion value in December, with chips still in customer testing.
Michael Burry spent much of 2026 betting against Nvidia (NVDA). On Tuesday, August 18, he shared a Wall Street Journal article about Etched and called the startup "serious competition for NVDA." Etched makes Sohu, a processor hardwired to run only transformer AI models, and it is reportedly raising money at a valuation of up to $21 billion. Sohu could pose a meaningful challenge to one piece of Nvidia's business because it is extremely fast at a narrow task. It cannot run several model types that are widely used today.
Burry rarely has anything complimentary to say about the AI trade, which is why that phrase caught my attention. He's been short Nvidia for most of the year, and most of his warnings have been about money, not chips. The circular financing, the $500 billion deals he compared to a Wall Street stunt, the Enron references. This one was different. He pointed at an actual product.
On Tuesday afternoon, he shared a Wall Street Journal piece on Etched, a four-year-old startup out of San Jose, under a headline about a $21 billion "kids in chips" company scooping up Nvidia's engineers. The detail that stuck with him was the speed. Etched said it took just 44 days from getting its test chips back from Taiwan Semiconductor Manufacturing (TSM) to running real inference on them. In chip terms, that's quick.
Here's what Etched actually built. Its chip is called Sohu, and it's an ASIC with the transformer architecture, the foundation of basically every large language model, built directly into the silicon rather than executed as software on a general-purpose processor. An Nvidia GPU can handle almost any workload. Sohu handles transformers and nothing else. Etched is betting that specialization will make it much faster at the one job it supports.
The company says a server containing eight Sohu chips produces about 500,000 tokens per second on Llama 70B. Etched presents that result as roughly 10 to 20 times the performance of an Nvidia H100 on the same workload. No independent laboratory has verified those figures. They remain company claims, not established facts.
The money is the part that makes you blink. Etched came out of stealth at the end of June with working silicon, more than $1 billion in signed orders and a valuation of about $5 billion from a round late last year. It raised another $300 million in July at $10.3 billion. Sequoia led that round, with a16z, Jane Street and SK Hynix (SKHY) participating. The Journal now reports that Etched is seeking more capital at a valuation as high as $21 billion. In eight months, that would put the company at roughly four times its December value while its processors remain in customer testing. Its investors range from Peter Thiel and Stanley Druckenmiller to Andrej Karpathy and Geoffrey Hinton.
So is Nvidia's moat cracking? I don't think this is the crack. Nvidia's real edge was never raw speed. It's CUDA, the software layer millions of developers already know, plus the fact that a GPU runs whatever model ships next. Sohu gives that up on purpose. It can't run mixture-of-experts models like DeepSeek V4 or Qwen3, which are some of the most-used open models out there right now. It can't do multimodal, image generation or training. And customers have to move off standard tools like vLLM onto Etched's own software to run it. That's a lot to give up for speed on a single workload.
It's also worth remembering Burry is short Nvidia, so "serious competition" is a convenient thing for him to post. His shorts haven't all worked. Nebius soared 27% last week and torched one of them. Nvidia (NVDA) closed at $219.80 on Tuesday, down about 2.3% from Monday's $225.01 close, but that was a soft day for chips across the board, not the market pricing in a startup that hasn't shipped at scale.
The honest read sits in the middle. Inference, the running-the-model half of AI, is where more of the money is going now, and specialists like Etched, Groq, and Cerebras are real players in it. That's a genuine threat to the slice of Nvidia's business that is just raw inference. It is not a threat to the part where Nvidia sells the one chip that runs everything. Burry's right that competition is showing up. He's just selling you the half of the story that fits his short.
This column is my personal opinion. I am not a financial advisor, and nothing here is investment advice.
