Jane Street, the quant trading firm famous for keeping its strategies close to its chest, has spent months testing a new kind of AI chip. On Tuesday it stopped testing and started paying. The firm led a $700 million round for chip startup Etched at a $21 billion valuation, and the first production rack now sits inside a Jane Street data center.
The number moved fast even for a market used to fast numbers. Etched was worth $5 billion in December. A $300 million Series C in July put it at $10.3 billion. Tuesday's round more than doubles that, a jump of nearly $11 billion in under a month. Co-founder and COO Robert Wachen told TechCrunch the stampede comes down to one thing: investors watched real hardware work, instead of trusting slides.
Why inference, and why now
Etched sells what it calls "frontier inference clusters" — full systems, not loose chips. Nvidia sells the same idea under a different name: AI factories. The difference is where the work happens inside the machine.
Wachen explains inference as two stages. The prefill phase reads the prompt, digests the context, and prepares an answer. It is compute-heavy and math-heavy. The decode phase then emits the answer token by token, and it is memory-heavy. Most chips on the market split the difference and do both jobs at once. Etched built dedicated silicon for each.
Its prefill chip runs at low voltage, which lets the company pack in more transistors without the heat problems that throttle high-end parts. More transistors, the company argues, means more tokens processed per second. For decode, Etched created something it calls cluster-scale memory — new memory hardware plus a new interconnect that lets many chips share one memory pool at very low latency.
"Many chips connect together and use a shared memory pool at a very, very fast, low latency," Wachen said. The payoff, if the benchmarks hold, is higher throughput and lower cost per query.
That economics matters because inference bills are growing faster than most forecasts. During training, a model maker spends once on compute to build the weights. During inference, every single request pays the bill again — a chatbot session can run hundreds of thousands of tokens through the same model. A small efficiency gain on that path compounds into very large savings for anyone running models at scale. That is the entire business case in one paragraph.
From niche bet to serious order book
Etched still fights its origin story. The company started in 2024 with Sohu, a 4nm ASIC designed only for transformer models — a single-purpose part that critics said would be obsolete the moment the field moved past transformers. That perception has faded as transformers stayed dominant and as the company dropped the single-model design.
Today the systems run any frontier model, and the hardware has a real customer list behind it. Jane Street said: "We tested the chip and are pleased with the early results. Etched's unique approach to inference delivers the precision we will need to support our most demanding workloads. We're excited to now have our own rack running in our datacenter."
That quote, published on Etched's blog, explains why the round closed so quickly. Quant trading runs on latency. Every microsecond of delay on a pricing calculation is money left on the table, and Jane Street is among the largest users of compute in that business. Its market-making desks price tens of millions of options and ETF trades a day, and the firm has a reputation for taking exotic hardware bets — it ran some of the largest FPGA deployments in finance long before AI entered the picture.
Beyond Jane Street, Etched says it carries roughly $1 billion in existing customer contracts. It also disclosed an extra $800 million gathered across four earlier rounds that were never announced, with strategic backing from VentureTech Alliance to deepen its manufacturing partnership with TSMC. The engineering team of more than 400 people has been recruited heavily from Nvidia, Google, and other chip veterans.
The investors lining up
The cap table reads like a who's who of Silicon Valley money. Kleiner Perkins, Sequoia Capital, and Andreessen Horowitz all hold stakes, joined by Peter Thiel, Tiger Global, Bain Capital Ventures, Neo, Stripes, Primary, Positive Sum, Diffusion, Argo, and Blackstone.
The production side is scaling too. Etched has opened a new factory in Taiwan to support TSMC-based manufacturing, plus an 80,000-square-foot prototyping hub near its San Jose headquarters. The company talks openly about ramping toward gigawatt-scale deployment, which in chip terms means a very large installed base. For comparison, a single gigawatt of continuous compute is roughly what some hyperscaler campus expansions are being sized for, and Etched is planning that as an internal milestone rather than a distant dream.
The field it is running against
Etched is not alone in betting that inference hardware needs a rethink. Groq has built a business on language-processing units that skip many of the memory bottlenecks GPUs face, and Cerebras sells wafer-scale engines that keep data on one enormous piece of silicon. Both have won real customers and both remain far behind Nvidia in volume. Specialized silicon has a long history of great benchmarks and short histories of market share.
Etched's answer to that history is full systems rather than chips. A customer does not need to redesign software stack around a new part; the cluster arrives as a unit, with interconnect and cooling solved. That is the same bundle strategy Nvidia perfected with its DGX line, and it is why the "Nvidia talent" label attached to Etched's hiring spree is not just branding — the people who built the bundle model are building the challenger.
What could still break the story is timing. Etched's architecture is tuned for today's transformer-heavy workloads. If a new architecture takes over, the dedicated circuits lose their edge. The company's answer is that its systems are general enough to run whatever frontier models look like, and that transformers have survived every obituary written for them so far.
What this says about the inference market
The valuation reset is a signal about where AI spending is heading. Training models gets the headlines, but inference is where the bills get paid. Every chatbot answer, every agent run, every automated code review executes on inference hardware, and that workload is growing faster than most companies can forecast.
Two recent stories on this site make the same point from different angles. The Stripe-OpenRouter deal, which turned AI model ordering into a payments business, shows how much money is changing hands per query. And the Grok 4.6 launch, which matched frontier scores at a fraction of the cost, shows the pricing pressure building underneath. Etched is betting it can sit in the middle of both trends: more demand, lower cost per token.
That is the same economic logic Nvidia is defending with its full-stack AI factory push. The incumbent has the distribution, the software stack, and the brand. Etched has a different architecture, a fast-moving balance sheet, and one of the most demanding customers on Earth running its first rack. Whether that is enough to dent Nvidia's grip is an open question — but Jane Street does not usually write $700 million checks on faith.
Related reading: AI news and analysis across this site, plus coverage of the broader infrastructure shift.
What could go wrong
The risks are easy to list. Etched's silicon is specialized, and specialization cuts both ways: if the industry shifts away from transformer-style architectures, the advantage shrinks. Manufacturing is another constraint — the company depends on TSMC capacity at a time when every AI chipmaker on Earth wants the same wafers. And the jump from a tested rack at a friendly customer to volume production across dozens of sites is the hardest step in hardware.
Scaling costs, too. Factories are not cheap. The company's own blog says the next stage brings "building new factories, global supply chains, fleet software," and other unglamorous work. Investors have now said, twice in a month, that they are willing to fund that work at a premium valuation.
For the rest of the industry, the message is simpler. Hardware for inference is becoming one of the most expensive and most watched categories in technology, and the bar for entry keeps rising. Etched cleared it with a partner who had already stress-tested the product where it matters most — inside a live trading operation.