Just today, computing power titan Nvidia was highlighted by "The Big Short" investor Michael Burry as confronting "intense competition" from a mere four-year-old startup.

This up-and-coming contender is Etched, a chip firm established four years ago and based in San Jose, California. On August 18, Etched successfully secured a fresh round of funding totaling $700 million, propelling its post-money valuation to a staggering $21 billion. Leading the investment is quantitative trading giant Jane Street, with renowned names like Sequoia, a16z, Tiger Global, and Blackstone also joining as co-investors.

Its valuation stood at a mere $5 billion in December 2025. In a span of just eight months, Etched's worth has surged fourfold.

Several concrete facts do lend credence to Burry's assertions. Etched, headquartered in San Jose and founded in 2022, boasts a workforce of around 400, with 15% hailing from Nvidia. What's even more remarkable is that their chip transitioned from test silicon to successfully executing inference workloads in a mere 44 days. Typically, such endeavors in the industry demand six months or more.

While confidence is high, the origins of the most pivotal data points warrant separate scrutiny. Etched asserts that an eight-chip server can churn through 500,000 tokens per second on Llama70B, dwarfing the approximately 23,000 of an eight-card H100 and around 45,000 of a B200.

All these figures stem from the company's internal tests, yet as of now, no independent third party has verified them in a real-world production setting. The "10x" performance claim cited by Burry falls within the realm of the company's promotional rhetoric, but it remains unverified by an outside party, and the benchmark for comparison remains ambiguous.

Furthermore, an unavoidable aspect is that Jane Street, the company's "backer," serves as both the lead investor and Etched's inaugural customer, with the first batch of racks already delivered. Following testing, Jane Street expressed satisfaction with the preliminary outcomes. Such an arrangement, where the lead investor and first customer are one and the same, is a rarity in the venture capital landscape.

Delving deeper into the chip-level intricacies, Sohu, Etched's chip, is specifically tailored for the transformer architecture, with the attention mechanism seamlessly integrated into the silicon. This design does enhance efficiency in inference tasks but at the expense of architectural flexibility. It struggles to efficiently handle MoE architectures, which are the current path for large models like DeepSeek-V4, and diffusion models are also beyond its purview.

In essence, it excels solely in inference, not training. Even if all the purported performance figures hold true, it would only outshine in the singular scenario of transformer inference.

Moreover, since Burry's re-emergence on the scene, he has consistently maintained a bearish stance on Nvidia. By his own admission, up to 80% of his portfolio is allocated to put options on Nvidia and Palantir, with a notional value totaling roughly $1.1 billion, including contracts for 1 million Nvidia shares with a $110 strike price and a 2027 expiration.

At the time of writing, Nvidia's market capitalization hovers around $5 trillion, with its stock price nearing all-time highs, rendering these options significantly out-of-the-money.

Evidently, Burry is ultimately spinning a narrative about Nvidia's AI bubble, with Etched merely serving as the latest piece of evidence he has latched onto.

Since the AI large model craze took hold, Nvidia has consistently been the subject of "copycat" accusations. Yet, Nvidia has been consistently imitated but never surpassed.