Etched's valuation doubles to $21B in one month following first rack delivery

Written by Xiao Bing

AI chip company Etched announced on August 18 that it raised $700 million, achieving a post-money valuation of $21 billion, led by quantitative trading giant Jane Street. Sequoia, a16z, Kleiner Perkins, Tiger Global, Bain Capital Ventures, Blackstone, and Peter Thiel all participated.

The company was valued at approximately $5 billion in December 2025. In July 2026, it raised $300 million, valuing it at $10.3 billion. A month later, its valuation doubled to $21 billion. Within less than a year, its paper value increased fourfold.

On the same day, Etched announced the delivery of its first-ever inference rack to Jane Street, the lead investor and the only publicly confirmed customer to have received the hardware.

Etched was founded in 2022 by three co-founders—Gavin Uberti, Chris Zhu, and Robert Wachen—who dropped out of Harvard to start the company and received funding through the Peter Thiel Fellowship. The company’s flagship product is Sohu, an ASIC chip specifically designed for Transformer inference, manufactured using TSMC’s 4-nanometer process.

The fundamental difference from NVIDIA GPUs is that GPUs are general-purpose computing chips that execute various AI tasks through CUDA software programming; Sohu directly hardwires the attention computation of Transformers onto silicon, sacrificing generality to achieve极致 efficiency for a single task.

According to Etched's official data, a server equipped with eight Sohu chips can generate over 500,000 tokens per second on Llama 70B, while an equivalent 8-H100 GPU system achieves approximately 23,000 tokens per second.

In 2026, Etched underwent a critical transformation.

Initially, it was positioned as designing chips tailored for specific large models; now, it claims the system can run any cutting-edge model. The technical approach has evolved from a pure Transformer ASIC to a two-stage optimization: a low-voltage chip handles the Prefill stage (analyzing prompts and context), while a proprietary shared memory and inter-chip interconnect manages the Decode stage (generating responses token by token).

Co-founder Robert Wachen referred to the latter as "cluster-level memory," where multiple chips are connected to a shared memory pool to reduce latency.

According to SiliconANGLE, Etched’s interconnects reduce tasks that would take competing chips 4,000 milliseconds to just 700 milliseconds. The company has also developed its own cooling plates and voltage regulator modules (VRMs).

Etched claims to have signed customer contracts exceeding $1 billion.

The investor list shines brightly, but the tech community’s feedback paints a different picture.

George Hotz (geohot), founder of tinygrad, posted a "public service announcement" on Twitter, directly naming Etched: "They might have a great chip but use off-putting marketing. Or they might have no chip at all—or a terrible one—relying entirely on smoke and mirrors."

Hotz broke down Etched’s external marketing strategy point by point, noting that its campaigns focus entirely on dramatic hardware photography, the founder’s legendary backstory, employee educational backgrounds, investor logo walls, and funding milestones—rather than the technical documentation, benchmarking methodologies, architecture diagrams, power consumption data, and analyst validations that traditional chip companies should prioritize. His conclusion: this marketing approach is harmful to the technology industry.

Tech blogger Zach’s analysis is more specific. He points out that the astonishing tokens-per-second figures claimed on Etched’s old website were "never delivered," and while the new website’s promotion of low-voltage technology may offer advantages in power efficiency, the focus has shifted from speed to performance per watt. In other words, Etched’s core selling point underwent a silent pivot over the past two years.

Zach’s judgment is: "Early skeptics had every reason to question the idea, even if they eventually ended up with a chip that had real advantages." He also points out a common path in the semiconductor industry: founding teams secure funding based on an ambitious initial concept, use that money to hire actual silicon engineers, and then those engineers develop new, better technical solutions. The concept changes, but the money is not returned.

Another frequently cited fact is that, as of mid-2026, no independent third-party benchmarking organization has published throughput test results for Sohu hardware under production conditions; all performance data comes solely from materials published by Etched itself.

Among all the doubts surrounding Etched, Jane Street’s role represents a delicate balance.

Jane Street is one of the world’s largest quantitative trading firms, with over $2 billion in net trading revenue in 2024. It is both the lead investor in Etched’s current funding round and the company’s only publicly confirmed hardware customer. Jane Street stated: “We tested the chip and are satisfied with the early results. Etched’s unique approach to inference delivers the precision we need to support our most demanding workloads. We’re excited to now have a dedicated rack running in our data center.”

A trading firm known for quantitative analysis and risk management—testing hardware before writing a check—is a stronger signal than any VC endorsement. Quantitative trading firms have extreme demands for inference latency and are willing to pay a significant premium for millisecond-level advantages, making them a highly specialized customer profile.

However, there is still a significant gap between "satisfactory testing" and "large-scale commercial validation."

A rack from Jane Street proving that the chip works does not demonstrate that the chip can be mass-produced, operate stably, or be adopted by a broader range of AI infrastructure customers. Moreover, the fact that this company appears as both a customer and the largest investor in the same funding round adds complexity to interpreting the signal.

The debate surrounding Etched reflects that AI hardware investment is currently at a unique stage.

NVIDIA's moat has never been just about chip performance. The CUDA ecosystem, network interconnects (InfiniBand/NVLink), supply chain relationships, and the developer community together form a self-reinforcing flywheel. Any new entrant must simultaneously overcome all these layers—not just achieve impressive benchmark numbers in a lab.

Etched’s chosen strategy is to bypass this flywheel: instead of selling chips, it sells complete rack systems. It designs its own chips, interconnects, cooling, and voltage regulation, delivering an integrated system so customers don’t need to use CUDA or build their own clusters. The success of this strategy hinges on one key assumption: whether Sohu’s performance advantage is sufficient to convince customers to adopt an entirely new technology stack.

The market assigned a $21 billion valuation despite the absence of independent third-party testing, publicly available revenue data, and with only the first rack having just been delivered. The list of investors in this round included nearly all of Silicon Valley and Wall Street’s top institutions, who clearly saw something compelling in the private demonstrations.

Etched’s COO, Robert Wachen, said that Andrej Karpathy, OpenAI’s Noam Brown, Geoffrey Hinton, and all investors have personally tested the hardware and are “very excited” by the results.

But between prototyping and mass production lies a long gap in yield, thermal management, system stability, and customer support. CEO Gavin Uberti himself admits: “It took us three years to deliver the first rack. The next one will be much faster.”

This is both a promise and an admission.

One unit every three years—whether the company can speed up production is the full answer to whether its $21 billion valuation can be sustained.