Etched said on Tuesday that it has raised another $700 million at a $21 billion valuation, a jump that puts the young AI hardware company in a different league from where it stood just months ago. Jane Street led the round after testing and buying Etched’s hardware.
The raise matters because it comes with a valuation that has climbed by nearly $11 billion in a month. Etched was valued at $5 billion in December and at $10.3 billion when it raised a $300 million Series C in July, a pace that suggests investors are assigning extraordinary value to the company’s pitch for speeding up AI inference.
Robert Wachen said investors are excited because Etched built two new components from scratch to make inference faster. He said inference is built in two stages, prefill and decode. Prefill is the mathematically heavy part, while decode is the memory-heavy part, and Etched says it designed a low-voltage prefill chip, plus a new memory system and interconnect for decode that it calls cluster-scale memory.
That technology is packaged as full systems the company calls frontier inference clusters, not just chips. Wachen said cluster-scale memory lets many chips connect together and use a shared memory pool at very fast, low latency. The company had once been viewed as a chip maker that etched one particular model into each chip, but it now says its systems can run any frontier model.
Jane Street’s backing adds a sharper edge to the story because the firm did not just write a check. It said it tested the chip and was pleased with the early results, then said it now has its own rack running in its datacenter. The firm also said Etched’s approach delivers the precision it will need to support its most demanding workloads, though it did not spell out exactly which workloads were used in testing.
The unanswered question is whether Etched can keep proving that advantage as the workloads get harder. For now, the deal shows that a startup focused on inference rather than training can still command an aggressive price if a serious buyer decides the hardware is already good enough to run inside its own datacenter.
