Etched Doubles Its Valuation to $21 Billion in a Single Month

In a scene reflecting the scale of unprecedented investor enthusiasm toward AI hardware companies, the startup «Etched» announced a new funding round worth $700 million, lifting its market valuation to $21 billion. What draws attention is not only the size of the amount, but the record speed at which the company's value doubled in just a few weeks.

An Exceptional Leap in Market Valuation

Even by the standards of the AI sector, known for its lofty valuations, Etched's trajectory seems astonishing. Last December, the company was valued at around $5 billion, then in July it raised a Series «C» round worth $300 million at a valuation of $10.3 billion. Now, within roughly one month, investors have doubled that figure, jumping by about $11 billion all at once.

The new funding round was led by «Jane Street», the fund specializing in quantitative trading, after it tested the company's hardware and actually acquired it to run within its own data centers.

Integrated Solutions for the Inference Process

Etched delivers its technologies as integrated systems it calls «frontier inference clusters», a concept that parallels what its major competitor «Nvidia» refers to as AI factories.

Co-founder and COO Robert Wachen attributes this investor appetite to the fact that the company designed from scratch two new components aimed at accelerating the inference process, that is, the processing that occurs after a user sends their request to the model.

The Two Phases of Inference: Prefill Then Generation

Wachen explains that the inference process is divided into two main phases:

• Prefill phase: A computationally intensive phase, in which the system needs to fully understand the request and its context.
• Decode phase: A phase heavily dependent on memory, where the system generates the tokens that form the final answer the user sees.

To handle each phase separately, the company developed a dedicated chip for the prefill phase that operates at low voltage, allowing more transistors to be integrated without the high-heat problems that other chips suffer from, thereby processing more tokens more quickly.

Cluster-Level Memory

As for the generation phase, Etched invented a new type of memory alongside an interconnect it named «cluster-level memory». Wachen explains that this solution allows a large number of chips to be linked together and used as a shared memory pool with very low latency, which translates into higher speeds and lower operating costs.

Correcting a Common Misconception

The company is still working to correct an impression that prevailed in its early days, namely that it «etches» a specific model into each chip so that it becomes dedicated to running only one model. While this was the original intention, that is no longer the case; Etched's systems have become capable of running any advanced AI model.

The Confidence of Major Investors

The «Jane Street» fund expressed in an official post its satisfaction with the initial results after testing the chip, noting that Etched's unique approach to inference provides the precision needed to support its most complex computational workloads, and that it now runs a dedicated unit within its data center.

The company's list of investors includes prominent names in the world of venture capital, among them «Kleiner Perkins», «Sequoia Capital», «Andreessen Horowitz», «Tiger Global», «Bain Capital Ventures», and «Blackstone», in addition to the famous investor Peter Thiel.

This investment momentum reflects a growing trend toward specializing hardware specifically for the inference phase, which has come to represent the largest portion of operating costs for generative AI models, opening the door to escalating competition over who can deliver the highest performance at the lowest cost.

✦ بقلم فريق دروب أيديا