Draft:Etched
This draft is not written from a neutral point of view. Wikipedia articles must be written neutrally in a formal, impersonal, and dispassionate way. They should not read like a blog post, advertisement, or fan page. Rewrite the draft to remove:
Where to get help
How to improve a draft
You can also browse Wikipedia:Featured articles and Wikipedia:Good articles to find examples of Wikipedia's best writing on topics similar to your proposed article. Improving your odds of a speedy review To improve your odds of a faster review, tag your draft with relevant WikiProject tags using the button below. This will let reviewers know a new draft has been submitted in their area of interest. For instance, if you wrote about a female astronomer, you would want to add the Biography, Astronomy, and Women scientists tags. Editor resources
|  |
Etched is an American semiconductor company that develops rack-scale computing systems for artificial intelligence inference. The company designs its own processors, memory interconnects, server racks and supporting software. It was founded in 2022 by former Harvard University students Gavin Uberti, Chris Zhu and Robert Wachen, and is headquartered in San Jose, California.[1]
By 2026, the company announced its products are complete inference systems rather than stand-alone chips and said they could run transformer, mixture of experts and state-space model architectures.[1] Its first production silicon was fabricated by TSMC on the N4P process.[2]
History
[edit]Founding and Sohu
[edit]Uberti, Zhu and Wachen founded Etched while studying at Harvard. Uberti and Zhu left the university in 2023 after the company raised a $5.36 million seed round led by Primary Venture Partners; the round valued Etched at $34 million.[3] The founders later became Thiel Fellows.[4]
Etched's original product was a processor code-named Sohu. The company sought to improve inference performance by omitting circuitry and software support not needed by transformer models. This specialization distinguished Sohu from general-purpose graphics processing units, but also meant that the processor could not execute other kinds of neural-network workloads.[4] In June 2024, Etched announced a $120 million Series A round co-led by Primary Venture Partners and Positive Sum Ventures. The financing brought the company's total capital raised at the time to about $125 million and was intended in part to fund fabrication at TSMC.[4] Etched had 35 employees when the round was announced.[4]
Rack-scale systems
[edit]Etched subsequently expanded the scope of its product from a transformer ASIC to a vertically integrated inference system. In 2026, it described its products as "frontier inference clusters" comprising processors, circuit boards, memory, networking, cooling, racks and software.[5] The company said its systems could operate models including DeepSeek, Qwen, Mamba and Llama, encompassing mixture-of-experts and non-transformer designs as well as conventional transformer models.[1]
In June 2026, Etched announced that its initial, or A0, silicon had been manufactured successfully on TSMC's N4P process and that complete systems were undergoing customer validation.[6] Etched reported more than $1 billion in signed customer contracts and said it planned to begin shipping its first racks during the summer of 2026.[7][6] The identities of the customers and the terms of the contracts were not disclosed publicly.[6]
By July 2026, the company employed more than 400 people. It operated a 2-megawatt data center at its San Jose headquarters and had opened an 80,000-square-foot (7,400 m2), 10-megawatt facility in Milpitas, California.[1] Etched also said it had established a manufacturing facility in Taiwan.[5]
Technology
[edit]Etched's 2026 architecture divides inference into two principal workloads. The first, prefill, processes an input prompt and its context and is comparatively compute-intensive. The second, decode, generates output tokens and is more constrained by the speed and capacity of memory.[1]
For prefill, Etched developed what it calls low-voltage inference (LVI). According to the company, the design runs its mathematical processing blocks at less than half the voltage used by most AI accelerators, reducing heat and allowing greater compute density. Etched has said that the system can sustain more than 80 percent of its peak floating-point performance on sparse, trillion-parameter mixture-of-experts models without thermal throttling.[5] These performance figures were company claims; public reporting in July 2026 noted that access to the hardware remained limited to investors and early customers and that mass production had not yet begun.[1]
For decode, the company developed cluster-scale memory (CSM), a shared memory pool connected by a proprietary high-bandwidth, low-latency interconnect. The design combines high bandwidth memory and static random-access memory in an effort to provide the capacity of the former with some of the latency advantages of the latter.[5][2] Etched has said that it co-designed the processor, packaging, power delivery, cooling and scheduling software as a single rack-scale system.[5]
Funding
[edit]Etched raised $5.36 million in seed funding in 2023 and $120 million in its 2024 Series A round.[3][4] In December 2025, it closed a $500 million financing led by Stripes, with participation from Positive Sum, Ribbit Capital, Jane Street, Hudson River Trading, Two Sigma and VentureTech Alliance, among others. That round valued the company at $5 billion.[7][6] By June 2026, Etched said it had raised a cumulative $800 million across four financings.[6]
In July 2026, Etched raised a $300 million Series C led by Sequoia Capital, with participation from Andreessen Horowitz, SK Hynix, Jane Street and Diffusion Capital. The round valued the company at $10.3 billion and brought its reported cumulative funding to approximately $1.1 billion.[6][1][2] Other investors disclosed by the company have included Peter Thiel, Andrej Karpathy, Geoffrey Hinton, Fei-Fei Li, Dylan Field and Stanley Druckenmiller.[6][1]
See also
[edit]References
[edit]- 1 2 3 4 5 6 7 8 Bort, Julie (July 23, 2026). "AI chip startup Etched defies skeptics, hits $10.3B valuation from big-name investors". TechCrunch. Retrieved August 9, 2026.
- 1 2 3 Trueman, Charlotte (July 24, 2026). "AI chip startup Etched closes $300m funding round, doubles its valuation to $10.3bn". Data Center Dynamics. Retrieved August 9, 2026.
- 1 2 Ward-Foxton, Sally (June 5, 2023). "Harvard Dropouts Raise $5 Million for LLM Accelerator". EE Times. Retrieved August 9, 2026.
- 1 2 3 4 5 Cite error: The named reference
TC2024was invoked but never defined (see the help page). - 1 2 3 4 5 "Frontier Inference Clusters". Etched. June 30, 2026. Retrieved August 9, 2026.
- 1 2 3 4 5 6 7 Bort, Julie (June 30, 2026). "Nvidia competitor Etched hits $5B valuation, $1B in sales for AI chip". TechCrunch. Retrieved August 9, 2026.
- 1 2 Bass, Dina (June 30, 2026). "AI Chip Startup Etched Says Jane Street, TSMC-Linked VC Invested". Bloomberg Law. Bloomberg News. Retrieved August 9, 2026.
- provide significant coverage: discuss the subject in detail, excluding routine coverage like product launches, staff appointments, or financial reports and listings in databases or listicles;
- are reliable: from reputable outlets with editorial oversight;
- are independent: not connected to the subject, such as press releases, the subject's own website, or sponsored content.
Please add references that meet all three of these criteria. If none exist, the subject is not yet suitable for Wikipedia.