Etched has raised $300 million in new financing at a $10.3 billion valuation to increase production and customer deployments of its purpose-built artificial intelligence inference systems. The financing was led by Sequoia Capital, with participation from Andreessen Horowitz, Jane Street, Diffusion and SK Hynix. Etched described the transaction as the highest valuation achieved by a Sequoia-led Series C financing.
The new round comes less than one month after Etched emerged from stealth and brings the company’s total funding to more than $1 billion.
Etched plans to use the capital to expand manufacturing capacity, accelerate product development and deploy additional frontier-scale inference clusters for customers. The company said demand for its systems is continuing to exceed available supply as customers move from technical evaluations into production deployments.
AI inference is the process of running a trained model to generate an answer, image, prediction or other output. While model training requires substantial computing resources, inference becomes the larger recurring workload as millions or billions of people and software applications continuously use AI systems.
Etched believes conventional AI hardware cannot provide the speed, power efficiency and economics needed to operate advanced models at a global scale. The company is developing chips, memory systems, networking and rack-scale infrastructure specifically for inference rather than adapting general-purpose processors to the workload.
The company said its hardware is architecture-agnostic, meaning customers can run AI models with different architectures and parameter counts without committing their infrastructure to a single model design.
Etched’s inference clusters currently run large mixture-of-experts models, including DeepSeek and Qwen, as well as non-transformer systems such as Mamba.
Mixture-of-experts models divide a neural network into specialized components and activate only the experts needed for a particular task. This can reduce the computing resources required for each request, but it also creates significant memory and networking demands as information moves among the different components.
Mamba uses a state space architecture that differs from the transformer design underlying many prominent generative AI models. Supporting both approaches is intended to help Etched’s customers prepare for changes in how frontier models are built.
Etched is developing two computing technologies for its rack-scale system to address what it considers the main limitations affecting AI inference: electricity consumption and memory performance.
The first, Low Voltage Inference, is designed to increase the amount of computing performance available within a fixed power envelope.
AI processors generate substantial heat when operating at high speeds. Thermal limits can force chips to reduce their clock rates before reaching their theoretical maximum performance.
Etched said Low Voltage Inference combines hardware and systems design to increase floating-point operations density while remaining within the power and cooling constraints of a data center. The company expects the technology to improve the number of tokens generated per watt and per dollar.
Tokens are the pieces of text or other information processed and generated by an AI model. Increasing tokens per watt can allow a data center to serve more AI requests without a corresponding increase in electricity consumption.
Etched’s second technology, Cluster Scale Memory, is a hybrid memory architecture combining static random-access memory and high-bandwidth memory.
Instead of designing the memory system around an individual processor, Cluster Scale Memory treats the entire computing cluster as one domain. It creates a shared pool of fast SRAM connected through Etched’s proprietary low-latency, high-bandwidth interconnect.
The approach is intended to give processors faster access to the model parameters and intermediate information required during inference.
Etched said the hybrid system can improve latency while avoiding some of the cost, manufacturing yield, reliability and thermal limitations associated with using SRAM-only systems or three-dimensional DRAM designs.
The company is expanding its physical infrastructure to support production and prototyping. Earlier in 2026, Etched established a manufacturing facility in Taiwan.
Etched also opened an 80,000-square-foot facility near its San Jose headquarters. The 10-megawatt site in Milpitas, California, is expected to house a new product introduction laboratory, an internal surface-mount technology line and additional customer deployment capacity.
A new product introduction laboratory helps a semiconductor company move designs from engineering prototypes into repeatable manufacturing. An internal surface-mount technology line allows Etched to assemble and test electronic components on circuit boards without relying entirely on external suppliers during early production.
The Milpitas expansion is intended to shorten development cycles and give the company more control over how quickly it can produce, validate and improve its systems.
Etched’s San Jose headquarters also contains a prototyping facility and a two-megawatt data center that operates the company’s hardware continuously.
The company has grown to approximately 400 employees. Its team includes engineers with experience at NVIDIA, Broadcom, Google’s tensor processing unit organization, SK Hynix and quantitative trading firms.
Etched said it operates as an in-person organization, with employees working from its San Jose location. The company is continuing to recruit engineers and other employees as it works toward deploying inference infrastructure at gigawatt scale.
Etched was founded on the view that inference will become one of the world’s largest markets as AI is integrated into software, consumer services, scientific research and industrial operations.
Its strategy differs from semiconductor companies building processors that support both model training and inference. Etched is focusing its hardware and broader system design on the specific requirements of serving trained models at high volume.
The company believes this specialization can provide lower latency and better power and cost efficiency than incremental improvements to existing general-purpose AI infrastructure.
Etched’s investors include Sequoia Capital, Andreessen Horowitz, SK Hynix, VentureTech Alliance, Peter Thiel, Jane Street, Hudson River Trading, Jump Trading, Two Sigma, Ribbit Capital, Stripes, Radical Ventures, Primary Venture Partners and Positive Sum.
KEY QUOTES:
“Inference is on the path to becoming the largest market in the world, and purpose-built compute will power the majority of the world’s inference.”
“What the team has accomplished in less than three years is extremely rare in the semiconductor world and a testament to the company and the culture they have built. We could not be more thrilled to partner with them as they build the defining company in inference hardware.”
Sonya Huang, Partner at Sequoia Capital
“Now is the time to be aggressive. Our chips work, people want them, and it’s time to ship.”
“The infrastructure required to serve frontier AI sustainably and economically was never going to come from incremental improvements to existing hardware. This round reflects a growing industry conviction that the challenge demands a new entrant willing to rebuild the stack from first principles and deliver frontier inference without the trade-offs between speed, cost, and scale that have hampered this market until now.”
Gavin Uberti, Co-Founder and CEO of Etched
“We tend not to celebrate fundraises. We have a lot of work to do to get to gigawatt scale. We’re excited to partner with the best AI infrastructure investors in the world to get there faster.”
“The team is working around the clock with our early customers to bring our first product to life.”
Rob Wachen, Co-Founder and President of Etched
