Etched Valuation Reaches $21 Billion in One Month and Delivers First Rack to Jane Street
The American startup specializing in dedicated ASIC chips for transformer models raised $700 million led by the quantitative desk that transitioned from investor to first client, signaling the next move in the competition against Nvidia in inference.
Etched announced on Tuesday, August 18, a $700 million Series D round led by Jane Street, valuing the startup at $21 billion. This value doubles the valuation from the previous round, which was $10.3 billion closed on July 23 under the leadership of Sequoia. In December, the same company was valued at $5 billion. No other chip manufacturer for inference has accelerated the pricing curve with such intensity in this cycle.
The round solidified a list of investors that combines pure-bred venture capital and sophisticated financial capital: Kleiner Perkins, Sequoia, Andreessen Horowitz, Tiger Global, Bain Capital Ventures, Neo, Primary, Stripes, Positive Sum, and Blackstone. In total, Etched has raised $1.9 billion in capital. The point that distinguishes this news from another headline of inflated venture deals is the operational gesture: Jane Street tested Etched's hardware in its own data center before becoming the leader of the round and is now operating a rack within the production environment of the quantitative treasury.
What the Sohu Chip Proposes
The Sohu is a transformer-only ASIC manufactured by TSMC using the N4P process, with 144 gigabytes of HBM3 memory per chip. The technical thesis is radical: sacrificing the general programmability that characterizes GPUs and dedicating every bit of silicon to the attention operations, projections, and feed-forward processes that dominate the inference of transformer models. According to the numbers released by the company, a server with eight Sohus exceeds 500,000 tokens per second using Llama 70B, compared to about 23,000 tokens in an equivalent cluster with eight Nvidia H100s. The claimed throughput is twenty times that of an H100 under transformer loads, with sustained FLOP utilization close to 80%.
Manufacturer numbers do not stand alone against production reality. What the August round adds is precisely the counter-proof: a client with its own skepticism and funding, Jane Street, invested only after running real workloads. This is a different signal than what MLPerf allows, because benchmark tests are optimized; trading workloads cannot forgive residual latency. Etched also confirmed more than $1 billion in signed orders, disclosed in June, and the Series D is the necessary check to scale production ahead of contractual deliveries.
The Landscape Outside the United States
In Taiwan, every new order for inference accelerators tightens TSMC's schedule on the N4P node, the same process used by the Sohu and contested by Nvidia, AMD, Apple, and now a growing legion of dedicated silicon startups. Etched enters this queue with $700 million in fresh funding and the contractual guarantee of more than $1 billion in orders announced in June. This compresses timelines for all other wafer takers and alters the negotiation benchmark for contracts through 2027, at a time when the U.S. Department of Commerce is still calibrating the rules for advanced silicon exports.
In South Korea, SK Hynix, the supplier of the HBM3 memory that populates each Sohu with 144 gigabytes per chip, was already a shareholder in previous rounds and is increasing its relative exposure against Samsung and Micron. The concentration of HBM in inference ASICs mitigates some of the commodity risk for the South Korean company, but ties its revenue to the fortunes of a handful of specific architectures. Rounds like this reinforce the thesis that HBM today is more akin to a military aircraft component than to traditional DRAM.
In China, lacking access to TSMC's N4P due to U.S. export controls, local manufacturers such as Cambricon and Biren must compete in less advanced nodes while domestic demand for inference grows under the push from open models Qwen3.8-Max, published by Alibaba on August 12 with open weights, and Kimi K3 from Moonshot. The gap with Etched is not one of talent; it is one of foundry capabilities. This reinforces the political pressure in Beijing for domestic capacity in advanced silicon.
The critical takeaway that avoids euphoria: a dedicated ASIC for transformers is an explicit bet against the evolution of architecture. If the market migrates within two years to state-space models, hyper-sparse mixtures of experts, or variants that reduce the fraction of pure attention, the Sohu could become a depreciated asset. Voices like Chris Ré at Stanford and the Mamba team at Princeton have been publishing papers arguing that the next generation of models will leave the classic transformer behind. Etched bets that the next three years of corporate deployment will still be dominated by the current architecture, a timeframe sufficient to amortize capex and capture margin that Nvidia currently concentrates in H100 and B200. The August Series D effectively buys that time. If the architecture moves too quickly, $21 billion may seem like the exact peak of a cycle.