- Etched nearly ran out of cash in 2023 and has since raised $800 million at a $5 billion valuation.
- The startup has booked over $1 billion in contracts for Sohu, its transformer-only inference chip.
- Peter Thiel, Jane Street and a TSMC-linked fund are betting it can beat Nvidia on inference costs.
In 2023, Etched’s founders could not get investors on the phone. The startup was running out of money, and the AI chip market was fixated on training bigger models, not the unglamorous business of running them once they were built. Two years later, Etched is worth $5 billion.
The company, founded by Harvard dropouts and Thiel Fellows Gavin Uberti and Rob Wachen, has raised $800 million in total and booked more than $1 billion in signed customer contracts for a chip that does one job: run inference, the computation that happens every time someone sends an AI model a prompt and waits for an answer.
“We recognised early on that frontier AI would become one of the most economically significant technologies ever created, but that the infrastructure needed to serve those models in a sustainable and economically viable way simply did not exist,” Uberti says.
A chip that only does one job
Training a model happens once or a handful of times. Running it happens millions of times a day, for every query a customer sends. As AI products scale to more users, inference is quietly becoming the industry’s biggest cost centre, ahead of training itself.
That shift is why investors who once chased training-focused chipmakers are now chasing companies like Etched. The global AI inference market was valued at $106 billion in 2025 and is projected to reach $255 billion by 2030, according to MarketsandMarkets.
Etched’s chip, called Sohu, is a fixed-function application-specific integrated circuit, or ASIC, a piece of silicon built for exactly one task, running transformer-based AI models, rather than a flexible, general-purpose graphics processing unit, or GPU, like the ones Nvidia sells.
The trade-off is real: Sohu cannot train models or handle non-transformer workloads. In exchange, Etched says the chip runs inference far faster. The company claims a single eight-chip Sohu server can process around 500,000 tokens per second on Meta’s Llama 70B model, and that one Sohu-equipped server can replace up to 160 Nvidia H100 GPUs.
Those figures come from Etched’s own internal testing and, according to TechCrunch, have not yet been independently verified through production deployments.
Etched has also achieved first-pass, or A0, silicon success on Taiwan Semiconductor Manufacturing Company’s N4P manufacturing process, and its systems are already running frontier models, including DeepSeek, Qwen, Meta’s Llama, and Mamba. The company plans to ship its first rack-scale systems in the summer of 2026.
The AI chip race
Nvidia still dominates AI infrastructure through its Blackwell platform and CUDA software ecosystem. Cerebras, which completed its IPO earlier this year, and Groq, which raised $650 million, are chasing the same inference-efficiency opportunity, as are in-house chips from Amazon, Google and Microsoft.
Most of those competitors build chips flexible enough to handle several types of AI workloads. Etched’s bet is narrower and riskier: it wagers that the transformer architecture, which underpins nearly all of today’s large language models, stays dominant long enough to justify hardwiring a chip to it permanently.
Manufacturing AI infrastructure at scale
Etched’s total funding of $800 million includes a $500 million round that closed in December 2025 at a $5 billion post-money valuation, led by Stripes.
The round included a strategic investment from VentureTech Alliance, a fund tied to TSMC, as well as Peter Thiel, Jane Street, Hudson River Trading, Jump Trading, Two Sigma, Ribbit Capital, Radical Ventures, Primary VC, and Positive Sum.
Jane Street has invested more than $100 million in Etched across multiple rounds, according to Startup Fortune’s reporting on Bloomberg’s disclosure. Angel investors include Andrej Karpathy, Geoffrey Hinton, Fei-Fei Li, Stanley Druckenmiller, Arthur Mensch, and Scott Wu.
“Gavin and Rob have resourced an incredible team around the ambition, including leaders from NVIDIA, Google’s TPU program, and Broadcom, who left the best roles in technology to build the entire system rather than just one piece of it,” says Ken Fox, founder and partner at Stripes.
The capital is funding physical infrastructure, not just chip design. Etched has built a factory in Taiwan and a data centre, test facility, and prototyping lab at its San Jose headquarters, with a team of more than 400 engineers pulled from Nvidia, Google’s TPU program, Broadcom, and quantitative trading firms.
Etched’s chip has not shipped in volume, and its most aggressive performance claims remain unverified by outside benchmarking. Whether a single-purpose chip beats Nvidia’s flexibility once real racks are running in production will determine whether this is the year specialised inference silicon actually cracks Nvidia’s grip, or simply narrows the gap on paper.
