AI chip startup Etched, founded in 2022 by three Harvard dropouts, has successfully closed a $300 million Series C funding round, achieving a valuation of $10.3 billion. The company's co-founder and Chief Operating Officer, Robert Wachen, officially confirmed the landmark deal.

The funding round was spearheaded by Sequoia Capital, featuring robust participation from major venture capital and technology entities including Andreessen Horowitz, SK Hynix, Jane Street, and Diffusion Capital. The inclusion of SK Hynix—one of the world's leading manufacturers of high-bandwidth memory (HBM)—carries strategic significance given that memory bandwidth is a primary bottleneck in AI inference, the exact problem Etched's architecture targets. The Series C also drew continued support from earlier backers, notably Peter Thiel, Andrej Karpathy, Dylan Field, and Amjad Masad.

This massive injection of capital doubles Etched's worth from a $5 billion valuation recorded in December 2024, a milestone achieved during a $500 million fundraising round. Etched claims that its new $10.3 billion valuation represents the highest valuation ever achieved in a Sequoia-led Series C funding round.

From Garage Servers to $1 Billion in Orders

Etched's trajectory to a multi-billion dollar enterprise has been anything but conventional. The startup specializes in designing specialized chips—sold as comprehensive full rack systems—that are explicitly optimized for transformer-based AI models. This specific architecture serves as the foundational technology behind prominent AI systems such as ChatGPT and Claude.

The company's origins are characterized by extreme frugality and determination. Wachen recalled the startup's humble beginnings, noting that the founders slept on a friend's floor in the Bay Area after dropping out of Harvard University. In those early days, the team ran their chip design tools on servers physically located in an early employee's garage. The setup was so rudimentary that it required the employee's wife to manually reboot the servers whenever the system encountered issues.

Today, Etched's operational scale has expanded significantly. The company currently employs 400 people and operates a massive 2-megawatt data center. The transition from running rudimentary design operations in a garage to overseeing advanced silicon manufacturing underscores the rapid scaling of the startup. Just last month, Etched announced several critical milestones: it had successfully manufactured its first chips through Taiwan Semiconductor Manufacturing Company (TSMC), initial systems were actively being tested by client companies, and the firm had already secured $1 billion in booked orders. The company notes it is working with some of the largest AI companies in the world, though specific customer names have not been publicly disclosed.

Technical Approach: Prefill and Decode Phases

Etched's core technology focuses on optimizing two distinct phases of AI inference—the process of running trained models to generate outputs for users, which represents the dominant share of real-world AI computing costs as models are deployed at scale. By separating the inference process into these two distinct computational workloads, Etched effectively customizes the hardware architecture to handle the specific bottlenecks associated with each phase of AI generation. The first is the "prefill phase," which involves the system comprehending the user's prompt through a highly compute-intensive mathematical process. The second is the "decode phase," which is responsible for generating the actual output tokens that the end-user sees. While the decode phase requires less raw computation, it demands massive memory bandwidth.

To address the computational demands of the prefill phase, Etched engineered a specialized chip that operates at a significantly lower voltage than competing AI chips on the market. This reduction in voltage notably decreases heat generation, which in turn allows a higher density of transistors to be packed into the hardware. The company claims this approach allows for faster inference at lower costs compared to general-purpose AI accelerators.

For the decode phase, Etched innovated a new memory and interconnect technology dubbed "cluster scale memory." This architecture enables numerous chips to seamlessly share a unified memory pool with extremely low latency. According to Wachen, the combination of these hardware innovations results in significantly higher processing speeds at lower operational costs.

Addressing Hardware Specialization Skepticism

When Etched initially launched, the concept of building a chip specifically tailored for transformer-based AI models was widely considered a risky venture by industry observers. The company continues to navigate the prevailing perception that its specialized systems are only capable of running specific large language models.

Wachen directly addressed this skepticism, clarifying that Etched's systems possess the versatility to run any AI model. This includes compatibility with Mixture of Experts (MoE) architectures, such as DeepSeek and Qwen, as well as non-transformer designs like Mamba, which utilizes a state-space model architecture.

The broader technology sector appears to be gradually validating Etched's specialized approach. For instance, Google is reportedly developing its Frozen v2 chip specifically tailored for its Gemini AI model, building on its established TPU (Tensor Processing Unit) program. Amazon has similarly invested in custom silicon through its Trainium and Inferentia chip lines for AWS AI workloads, and Meta has developed its MTIA inference accelerator. This movement suggests that the concept of etching model-specific features directly into silicon is steadily gaining mainstream industry acceptance.

Investor Confidence Through Hands-On Demonstrations

Etched successfully secured its highly prominent list of investors by inviting them to private, hands-on hardware demonstrations hosted at its own offices. Wachen highlighted that key industry figures, including Andrej Karpathy from Anthropic, Noam Brown from OpenAI, AI pioneer Geoffrey Hinton, and all the financial backers participating in the funding round, physically tested the hardware before committing their capital.

“These are all people who actually tried the hardware and are very excited about it,” Wachen remarked regarding the investors and testers. The fact that industry veterans and collaborators from competing AI labs were willing to publicly back the hardware demonstrates a significant vote of confidence in Etched's underlying architecture.

Despite the substantial progress and formidable backing, Wachen acknowledged that the company still faces significant hurdles in scaling its operations. “We had no idea how hard it was going to be,” he admitted. “I think we still have to be humbled by what it will take to actually get to scale.”

Implications for the AI Industry

Etched's rapid growth trajectory and its ability to secure high-level investor confidence signal a broader shift in the AI hardware landscape. As AI models continue to grow increasingly complex, the demand for specialized inference chips capable of simultaneously reducing power consumption and lowering operational costs is expected to surge. These specialized hardware components could very well become critical infrastructure for the next generation of artificial intelligence deployment.

If Etched successfully delivers on its ambitious technical promises and scales its manufacturing capabilities, it could present a direct challenge to Nvidia's current dominance in the AI compute market. Nvidia controls an estimated 80% or more of the AI chip market, and the emergence of well-funded competitors focused on purpose-built inference hardware marks one of the most significant competitive dynamics to watch in the semiconductor and AI infrastructure industries.