AI chip startup Etched defies skeptics, hits $10.3B valuation from big-name investors

The artificial intelligence landscape is witnessing a seismic shift, with specialized hardware emerging as a critical battleground. In a striking development that underscores this trend, AI chip startup Etched has defied early skepticism to reach an impressive $10.3 billion valuation. This milestone comes on the heels of a $300 million Series C funding round led by heavyweight investors, signaling strong confidence in the company's innovative approach to AI inference.

Etched, a company born from the vision of three Harvard dropouts – Gavin Uberti, Chris Zhu, and Robert Wachen – is making waves with its claim of new chips and memory components that dramatically speed up inference on any AI model, all without the need for traditional Graphics Processing Units (GPUs).

Etched's Ascent: From Harvard Dorms to a Multi-Billion Dollar Valuation

Founded in 2022, Etched emerged from stealth mode in late June with working silicon and a staggering over $1 billion in signed customer contracts. The journey began when Gavin Uberti, Chris Zhu, and Robert Wachen, then Harvard students, recognized a significant gap in the burgeoning AI infrastructure. They became Thiel Fellows, an unusual concentration for the program, which provided early validation for their ambitious idea to build better inference hardware.

Their initial vision faced considerable skepticism. In 2023, as the AI chip market was largely fixated on training larger models, Etched struggled to attract investors for their focus on the "unglamorous" business of running deployed models, known as inference. However, the founders held a strong conviction: frontier AI would become one of the most economically significant technologies, but the existing infrastructure wasn't equipped to serve these models sustainably and economically.

This conviction has now paid off handsomely. The recent $300 million Series C round, led by Sequoia Capital with participation from Andreessen Horowitz (a16z), Jane Street, SK Hynix, and Diffusion, catapulted Etched's valuation to $10.3 billion. This marks Sequoia Capital's highest-valued Series C investment. This round more than doubled Etched's valuation from approximately $5 billion in late 2025, when it secured a $500 million round led by Stripes. The company has now raised over $1 billion in total funding, with earlier investors including Peter Thiel, Jump Trading, Two Sigma, Ribbit Capital, and Radical Ventures.

The involvement of SK Hynix is particularly noteworthy, as the memory manufacturer supplies the HBM3E chips utilized in Etched's hardware, indicating a strategic alignment within the supply chain. Etched is reportedly already in discussions for a subsequent funding round at an even higher $20 billion valuation, led by existing investor Jane Street, though these terms are not yet finalized.

The Sohu Chip: Redefining AI Inference Hardware

At the heart of Etched's groundbreaking claims is its core product, the Sohu chip. This isn't a general-purpose processor; it's a fixed-function Application-Specific Integrated Circuit (ASIC) designed specifically for one task: running transformer-based AI models. This laser focus on the transformer architecture, which underpins nearly all modern large language models (LLMs) and generative AI applications, is Etched's boldest bet.

Unlike general-purpose GPUs from companies like Nvidia, which are designed for flexibility across a wide range of computational tasks, Sohu "burns the transformer architecture directly into silicon." This means the complex matrix multiplication patterns specific to transformer inference are hard-wired into the chip's physical design, rather than being orchestrated by software on a programmable compute unit.

This specialized approach offers significant advantages in performance and efficiency for its intended workload. Etched claims a single 8-chip Sohu server can process around 500,000 tokens per second on Meta's Llama 70B model. To put this in perspective, an equivalent 8-GPU Nvidia H100 system reportedly runs the same workload at roughly 23,000 tokens per second, implying a single Sohu server could replace about 160 H100s for transformer inference.

Key Architectural Innovations

Etched's performance gains are attributed to two primary breakthroughs:

• Low Voltage Inference (LVI): Traditional AI chips often suffer from thermal throttling, where increased utilization leads to higher power consumption and reduced clock speeds, limiting sustained inference throughput. Etched's architecture runs its math blocks at less than half the voltage of most AI chips. This drastically reduces heat generation, enabling higher sustained speeds and greater FLOPs density without thermal throttling. The company claims this allows them to run trillion-parameter sparse Mixture-of-Experts (MoE) models at over 80% peak FLOPs.
• Cluster Scale Memory (CSM): Current AI chips using HBM (High Bandwidth Memory) face bottlenecks in memory subsystem and interconnect, hindering SRAM-level decode speeds. Etched addresses this with a proprietary ultra-low-latency, high-bandwidth interconnect that creates a shared memory pool across an entire cluster. This HBM/SRAM hybrid design aims to solve both memory capacity and memory-to-memory latency issues, enabling high throughput and interactivity simultaneously. Sohu utilizes 144GB of HBM3E per chip, offering approximately 1.8 times the memory bandwidth of an H100 SXM5.

These innovations are part of a comprehensive co-design strategy, where Etched designs not just the chips, but also the entire racks, software, and manufacturing methods to optimize for frontier models.

Industry Implications and the Future of AI Hardware

Etched's emergence and significant valuation highlight a pivotal moment in the AI hardware industry. Nvidia has long dominated the market with its powerful GPUs, particularly for AI training. However, the rapidly expanding AI inference market — the process of running trained AI models to generate predictions or responses — is becoming a distinct and massive opportunity.

As AI applications scale to millions of users, inference is quietly becoming the industry's biggest cost center, even surpassing training expenses. This shift is precisely why investors are now flocking to companies like Etched, which offer specialized solutions for this growing demand.

The "no GPUs required" claim from Etched directly challenges the established paradigm. While GPUs offer unmatched flexibility, their general-purpose nature means they are often underutilized for specific, repetitive tasks like transformer inference. Etched's bet is that this specialization will lead to superior performance per watt and lower operational costs for hyperscale data centers.

However, this specialized approach also carries inherent risks. If the transformer architecture were to be displaced by a fundamentally different AI model type, Sohu's advantage could diminish significantly. Yet, with transformers currently dominating every frontier of AI, from language and vision to video and agents, Etched's bet appears prescient.

The company plans to use its newly raised capital to aggressively scale production and customer deployments, including expanding its 80,000-square-foot facility and a 10-megawatt Milpitas site for NPI (New Product Introduction) and in-house SMT (Surface-Mount Technology) lines. With over $1 billion in signed customer contracts and first racks scheduled to ship in summer 2026, Etched is poised to make a tangible impact on the AI infrastructure landscape.

Etched's success underscores a broader trend: the AI hardware market is diversifying beyond general-purpose GPUs. Companies like Cerebras Systems, Groq, SambaNova Systems, and Fractile are also developing inference-focused hardware, but Etched stands out with its singular commitment to transformer-only architecture. This intense competition and innovation promise to drive down costs and unlock new, previously impossible AI applications, from real-time conversational agents to generative media and advanced robotics.

Frequently Asked Questions

What is Etched, and what problem does it solve?

Etched is an AI chip startup that designs specialized chips and memory components to accelerate AI inference. It aims to solve the problem of inefficient and costly AI inference on general-purpose GPUs by offering purpose-built hardware optimized specifically for transformer-based AI models.

Who are the founders of Etched?

Etched was founded in 2022 by three Harvard dropouts: Gavin Uberti, Chris Zhu, and Robert Wachen.

What is the Sohu chip, and how does it differ from GPUs?

The Sohu chip is Etched's core product, an Application-Specific Integrated Circuit (ASIC) designed exclusively for running transformer-based AI models. Unlike general-purpose GPUs, which are flexible but less efficient for specific tasks, Sohu hard-wires transformer attention patterns directly into silicon, leading to significantly higher performance and efficiency for AI inference workloads.

What is Etched's current valuation and who are its main investors?

Etched recently achieved a $10.3 billion valuation after a $300 million Series C funding round. Key investors include Sequoia Capital, Andreessen Horowitz (a16z), Jane Street, SK Hynix, Diffusion, Stripes, Peter Thiel, Jump Trading, Two Sigma, Ribbit Capital, and Radical Ventures.