AI inference chip startup Etched is preparing a new funding round of $700 million at a valuation of around $21 billion, more than double the $10.3 billion valuation it received in its late-July Series C — a $300 million round led by Sequoia — according to a Wall Street Journal report published on August 18.

The new figure represents a fourfold increase from the company's $5 billion valuation in December 2025, and comes barely a month after the Series C closed on July 31, with Andreessen Horowitz (a16z), SK Hynix, Jane Street and Diffusion Capital also participating. SK Hynix is one of the world's largest memory-chip makers and a major supplier of the high-bandwidth memory used in AI accelerators, giving Etched a backer with direct ties to the memory supply chain that rack-scale AI systems rely on.

Etched has also signed Jane Street as its first customer, delivering a server rack equipped with AI processors optimized for rapid inference computing to the Wall Street quantitative trading giant. Landing a paying deployment this quickly is significant validation for a young chip startup, since large customers typically want evidence that new silicon can run reliably at scale before committing to it. The pace of the company's growth — it only emerged from stealth on June 30 — underscores the broader rush to find alternatives to Nvidia, which holds a commanding lead at the head of the supply chain providing the silicon powering the AI boom.

"This round reflects a growing industry conviction that the challenge demands a new entrant willing to rebuild the stack from first principles," co-founder and CEO Gavin Uberti said in a statement.

What Etched's chips do

Etched sells full rack systems optimized for the inference portion of the AI compute stack, the step that occurs after users submit prompts. The company argues that its inference-only focus allows it to build specialist chips, in contrast to Nvidia, which builds all-purpose GPUs for both AI training and inference.

Etched splits the work into two phases. The prefill phase, which reads and interprets prompts, uses what co-founder Robert Wachen calls low-voltage inference — a design choice that allows the chip to run cooler, accommodate more transistors, and operate at a higher clock speed. The decode phase, in which the model writes its answer one token at a time, relies on a Cluster Scale Memory design. That shared memory architecture allows accelerators inside a rack to draw data from another memory pool instead of having to copy the data themselves.

Etched said its systems are already inside DeepSeek, Qwen, Mamba, and Llama models.

Big money is flowing into inference

The raise comes as forecasters identify inference as the fastest-growing slice of an already hot AI sector. Etched delivered on its promise to start shipping chips by the summer and has filled more than $1 billion in booked orders, as reported by Cryptopolitan.

Bloomberg Intelligence has set a $1.3 trillion target for the inference sector by 2032, a figure that would double the size of the AI training market. Iron Mountain and Structure Research expect inference capacity to overtake training capacity this year and to continue growing until it accounts for 80% of AI compute load by 2030. The underlying economics explain the divergence: training a model is a discrete, one-time outlay, while inference runs every time a user sends a prompt, performs a search, or triggers an AI agent — a recurring cost that scales with how widely the technology is actually used.

The rush to build AI data centers has created a steady demand route fed by players such as Nvidia, Cerebras, and AMD, which are also building inference-specific systems based on the same prefill-and-decode split that Etched describes.

Etched now employs more than 400 people and has reported first-pass silicon success on TSMC's N4P, a 4-nanometer-class manufacturing node. The company runs a 2-megawatt data center at its San Jose headquarters and has opened a new 80,000-square-foot, 10-megawatt facility in nearby Milpitas, TechCrunch reported. What to watch from here is execution: whether the booked order book converts into systems running in customer data centers as rival inference offerings reach the market at scale.