Etched AI chip startup has closed a $300 million Series C funding round at a $10.3 billion valuation, doubling its worth in just seven months. The round, led by Sequoia, marks the highest valuation ever for a Sequoia Series C and signals strong market confidence in specialized AI inference hardware.
Andreessen Horowitz, SK Hynix, Jane Street, and Diffusion Capital also participated in the round, alongside earlier investors. The company, founded by three Harvard dropouts in 2022, previously raised $500 million at a $5 billion valuation in December. Etched announced last month that it had successfully manufactured its homegrown chips via TSMC, and it has already booked $1 billion worth of orders. High-profile individual backers include Peter Thiel, Andrej Karpathy, Dylan Field, and Amjad Masad.
Private Demos Win Over Top AI Researchers
Much of the skepticism surrounding Etched stemmed from limited access to its hardware. The company countered this by inviting prominent figures to its office for private demonstrations. “Andrej Karpathy from Anthropic, Noam Brown from OpenAI, Geoffrey Hinton, as well as all the investors in the funding round — these are all people who actually tried the hardware and are very excited about it,” co-founder and COO Robert Wachen says. This hands-on approach proved crucial for a startup trying to convince the tech elite that its silicon could outperform established giants in running transformer models.
Etched AI Chip Breakthroughs Drive Investor Confidence
The core of Etched’s pitch is a pair of custom components designed from scratch to accelerate inference. “Inference is built in two stages,” Wachen explains. “Prefill and decode.” The prefill phase involves understanding the prompt and context, which is mathematically intensive. The decode phase generates the output tokens and requires massive memory capacity. To tackle the prefill phase, Etched created a chip that operates “dramatically” faster by running at a much lower voltage than any other AI chip. “We call this low-voltage inference,” Wachen says. Lower voltage generates less heat, allowing the chip to pack in more transistors. For the decode process, the company developed a new type of memory and interconnect technology called cluster-scale memory. This allows many chips to connect together and use a shared memory pool at very fast, low latency. The result is high speeds but lower costs for AI inference hardware.
Running Transformer Models and Beyond
Etched launched when building chips specifically for transformer models was considered a wild concept. The company still battles the perception that its systems only run specific large language models. Wachen clarifies that the systems can run any AI model. This includes Mixture of Experts models like DeepSeek and Qwen, which split tasks across specialized sub-models, as well as non-transformer designs like Mamba, built on a state-space model architecture. The startup’s journey began in a garage with servers that required manual reboots by an employee’s wife. Today, the company employs 400 people and operates a 2-megawatt data center. It recently opened a new 80,000 square-foot, 10MW facility in Milpitas, down the road from its main San Jose office. “We’re running tokens in our lab today, working with some of the largest AI companies in the world,” Wachen notes.
What Happens Next
The massive $10.3B valuation sets high expectations for Etched as it transitions from testing to mass production. The company must prove it can scale its manufacturing and deliver rack systems to clients who have placed $1 billion in orders. While the founders have secured top-tier backing and demonstrated low-voltage inference in private demos to figures like Geoffrey Hinton and Noam Brown, the path to mass delivery remains challenging. “We had no idea how hard it was going to be,” Wachen admits. “I think we still have to be humbled by what it will take to actually get to scale.” Competitors like Nvidia and emerging players like Google’s Frozen v2 chip for Gemini will not cede the market easily. The next few quarters will determine whether Etched can fulfill its orders and establish its AI inference hardware as a standard in data centers running transformer models and other architectures.
— David Kim, technology desk, AXO News