On the surface, Etched seems no different from many other hot AI startups: founded by three Harvard dropouts in their early twenties, it quickly raised nearly $2 billion in funding with a product that was just an idea four months ago.

However, unlike most other startups, Etched doesn't produce software; it produces semiconductors. The high capital costs and long development cycles of the semiconductor product category mean it can only thrive in a corner of Silicon Valley where experience still trumps youthful ambition.

Sonya Huang, general partner at Sequoia Capital, said that tech investors generally believe "don't invest in young people in the chip industry." Sequoia Capital led Etched's $300 million Series C funding round. "People generally think these young people are omnipotent, and maybe they are in the internet industry, but in the chip industry, startups are presumed guilty until they are proven innocent."

Key Milestones for Etched

After a period of intensive research and development, including testing in a small data center set up in the company's offices, Etched has finally reached a key milestone: successfully signing its first client—the mysterious Wall Street quantitative trading giant Jane Street. Jane Street purchased its product—a server rack filled with AI processors optimized for fast inference computation. The startup has secured over $1 billion in orders and has begun shipping.

Meanwhile, Jane Street led a new $700 million funding round for Etched, valuing the company at $21 billion (approximately 140 billion RMB). This valuation is higher than the $20 billion Nvidia paid last December to license its technology and hire the leadership of Groq, a similar startup. Other investors in this round include Kleiner Perkins, Sequoia, Andreessen Horowitz, Tiger Global, Bain Capital Ventures, Neo, and Blackstone.

The startup stated that after retrieving the test chip from TSMC, it completed the deployment of inference workloads—the computational processes by which AI models respond to user queries—in just 44 days, a process that typically takes six months or more.

"We may be the only AI chip built by a startup that succeeded on the first try," said Robert Wachen, co-founder and president of Etched.

While its co-founders may be relatively new, the company boasts a wealth of chip industry experience. About 15% of Etched's approximately 400 employees previously worked at Nvidia, and the company actively recruits talent from the market leader and other prominent semiconductor companies. The company stated that Jimmy Daley, former VP of Data Center Systems Engineering at Nvidia, was also an investor; he retired in July.

Systems engineer Anthony Gamerman moved from Nvidia to Etched in April, partly because many of the people he admired most had left Nvidia. "One left, then another a few weeks later, and you'd hear, 'Wow, I can't believe so-and-so's going to Etched,'" he said.

Etched originated at Harvard University, where its other two founders, Gavin Uberti and Chris Zhu, met in a Math 55 course. This course is notoriously difficult, a theoretical mathematics undergraduate course with a pass rate of less than 50%. All three founders are 24 years old.

Wachen, a programmer who lived on the same floor as Uberty, was running an edtech company at the time. He later founded Prod, a student-run, non-profit startup accelerator in Cambridge, Massachusetts, which incubated the coding assistant Cursor and the AI-powered recruiting platform Mercor. He then co-founded Etched.

The startup initially struggled to raise funds because potential investors and customers believed the founders lacked experience and would not succeed in the complex field of AI computing hardware.

To address this challenge, the company adopted a "production-as-product" philosophy and a capital-intensive chip development strategy, aiming to bring products to market as quickly as possible.

This meant investing tens of millions of dollars in diagnostic equipment, writing software, and designing custom server racks before the first silicon wafers were even manufactured, working around the clock for two consecutive months in the spring.

In addition, the company also had a 2-megawatt office data center. Inside the server room, refrigerator-sized server systems hummed with the low rumble of their cooling systems. Potential customers could remotely access and test Etched's chips through this data center.

Another unique aspect of Etched is its level of autonomy over its supply chain. The company has a 20-person team in Taiwan and owns a factory there that manufactures server components. While most chip startups outsource most of their finished product testing, Etched conducts almost all of its testing internally.

Company executives say they use a unique design methodology called "cluster-scale memory" to design their chips. This allows multiple chips within a server to communicate much faster, working collaboratively like a single processor. They state that a communication task that takes approximately 4,000 nanoseconds for an Nvidia Blackwell processor can be completed in just 700 nanoseconds with Etched's chips, thanks to its chip architecture and custom interconnect technology.

In 2023, the company hired Saptadeep Pal as its chief chip architect. Pal previously researched multi-GPU parallelization (i.e., methods of connecting multiple chips) at Nvidia, then co-founded the Bitcoin mining rig Auradine and designed what he calls the lowest-voltage chip ever made.

When Etched needed help designing servers and racks to integrate its chips into a large system, the company hired Brian Loiler in 2024, a top systems engineer who had worked at Nvidia for nearly 23 years. Loiler stated that he subsequently recruited about a dozen engineers from Nvidia to join Etched, some of whom even turned down highly attractive renewal offers from other companies.

The startup sees the quantitative trading industry as a hub for its clients and investors. Around early 2025, the company poached Tim Perevozchikov, a former quantitative trader from Wall Street firm Two Sigma, to head Etched's finances.

Perevozchikov stated that quantitative trading firms typically require the most powerful processors to run their investment models quickly, gaining even a fraction of a second of trading advantage.

Patrick O'Shaughnessy, a board member of Etched and founder and CEO of venture capital firm Positive Sum, said Etched was the first inference-focused semiconductor startup founded after the launch of ChatGPT. Positive Sum has been backing the company since its early funding stages.

"They're moving faster than anyone I've ever met, especially in the hardware space, where progress is typically very slow," O'Shaughnessy said.

Source: Compiled from WSJ

View more at EASELINK