Etched at $10.3B: Taking on Nvidia with Custom AI Chips

More Than Just a Funding Round

On July 23, Etched — the AI chip startup founded by three Harvard dropouts in 2022 — announced a $300 million Series C at a $10.3 billion valuation. That's double its $5 billion valuation from December 2025, achieved in just seven months. It's also the highest valuation ever for a Sequoia-led Series C.

But the story goes deeper than the numbers. This signals that the AI race is shifting from "who can train the biggest model" to "who can run inference the cheapest." And Nvidia, after years of dominating AI hardware, has another serious contender at its gates.

Who Is Etched?

Etched was founded by Gavin Uberti (CEO), Robert Wachen (COO), and Chris Zhu — three Harvard dropouts who pursued an idea many considered absurd: designing a chip that does nothing but run inference for transformer models.

Their first product is Sohu — an ASIC (Application-Specific Integrated Circuit) purpose-built for the transformer architecture powering virtually every modern AI model: ChatGPT, Claude, Gemini, and more. Unlike Nvidia's GPUs — general-purpose chips carrying unnecessary overhead for AI workloads — Sohu does exactly one thing: transformer inference. And it does it extremely fast.

Notably, Etched doesn't sell standalone chips. They sell complete rack systems, mirroring Nvidia's DGX strategy. Smart move — enterprise customers don't want to assemble their own AI infrastructure.

Two Chips, Two Radically Different Approaches

Etched's differentiation lies in its dual-chip architecture for the two stages of inference:

Prefill: The phase where the model "reads" the user's prompt, including all context. This is compute-intensive. Etched's prefill chip operates at ultra-low voltage — "low-voltage inference." Less voltage = less heat = more transistors in the same die area, yielding significantly higher FLOPs density than traditional GPUs.

Decode: The phase where the model generates output tokens one by one. This needs less computation but demands massive memory bandwidth — every generated token must access the model's full parameter set. Etched built "cluster-scale memory" — a new memory and interconnect technology letting multiple chips share a single memory pool at extremely low latency.

Combined, Etched claims these two chips deliver an order of magnitude better inference performance than Nvidia GPUs at lower cost. Third-party analysts estimate Sohu could achieve up to 10x better performance and efficiency for large-model inference.

Who's Betting on Etched?

The investor list reads like a tech industry who's-who:

Sequoia led the Series C — their highest-valuation Series C ever
Andreessen Horowitz (a16z), SK Hynix (leading HBM memory manufacturer), Jane Street, and Diffusion Capital joined the round
Angel backers include Peter Thiel, Andrej Karpathy (Anthropic), Dylan Field (Figma), and Amjad Masad (Replit)

Wachen revealed how they convinced these names: by letting them use real hardware. "Andrej Karpathy, Noam Brown from OpenAI, Geoffrey Hinton — these are all people who actually tried the hardware and are very excited about it." No slide decks, no demo videos — actual silicon running real workloads.

The company also announced $1 billion in booked orders back in June. Etched now employs around 400 people, operates a 2MW data center at its San Jose headquarters, and just opened an 80,000-square-foot, 10MW facility in nearby Milpitas.

Not Just Transformers

A common misconception: Sohu only runs one type of model. In reality, Wachen confirms the systems support any modern AI architecture — including Mixture of Experts models (DeepSeek, Qwen) and even non-transformer designs like Mamba (a state-space model). "Transformer-only" means optimized for transformers, not locked to a specific model.

Here's the interesting part: etching model architectures into silicon is no longer considered crazy. Google is reportedly developing its Frozen v2 chip with Gemini's architecture etched directly into the silicon. When Google pursues the same approach, ASICs for AI clearly aren't a dead end.

What This Means for Developers

The rise of specialized inference chips has three direct implications for software engineers:

1. API costs will drop — significantly

Inference dominates the operating costs of AI services. If specialized chips can reduce per-token costs by an order of magnitude, API providers will have room to cut prices. This especially benefits high-throughput applications: semantic search, customer support chatbots, real-time code completion.

2. Self-hosting open-weight models becomes viable

At current Nvidia H100/B200 prices, self-hosting a model like Llama or DeepSeek at production scale is economically challenging for most startups. If Sohu or similar chips reduce inference costs by 5-10x, the balance tilts toward self-hosting — less dependency on Big Tech APIs.

3. A more diverse hardware ecosystem

Monopolies are never good for developers. An AI GPU market with only Nvidia means high prices, supply shortages, and no alternatives when you need different trade-offs. Etched's emergence, alongside Groq, Cerebras, and d-Matrix, is creating a competitive ecosystem — and developers are the ultimate winners.

The Road Ahead Isn't Paved

For all the excitement, significant challenges remain:

No independent benchmarks yet: All performance claims currently come from Etched or its invited partners. The technical community needs third-party benchmarks for objective assessment.
The ecosystem problem: Nvidia doesn't just sell chips — it sells CUDA, a software platform that has been the industry standard for over 15 years. For PyTorch, vLLM, and TensorRT-LLM to run smoothly on Sohu, Etched needs to build its own software stack — and convince the community to adopt it.
Mass production at scale: TSMC successfully manufactured the first batch of chips, but scaling to mass production for thousands of systems is an entirely different challenge. Wachen admits: "I think we still have to be humbled by what it will take to actually get to scale."

Takeaways

Etched raised $300M Series C at a $10.3B valuation, doubling in 7 months — the highest valuation ever for a Sequoia-led Series C
Sohu is an ASIC purpose-built for transformer inference, using a dual-chip architecture (low-voltage prefill + cluster-scale memory decode) to deliver an order of magnitude better performance than Nvidia GPUs
$1B in booked orders, first silicon successfully manufactured by TSMC, systems now being tested by customers
Developer impact: lower inference costs → cheaper APIs, viable self-hosting for open-weight models, reduced Nvidia dependency
Challenges: no independent benchmarks, must build a CUDA-competitive software stack, need to scale manufacturing

The AI race is entering a phase where inference cost — not training cost — is the deciding factor. If Etched and other specialized chip companies can deliver on their performance promises, the AI hardware market will fundamentally restructure within 2-3 years. If not, this $300 million round may become just another footnote in the graveyard of startups that dared to challenge Nvidia.

Either way, developers win. Competition in hardware means cheaper, faster inference — and more options to choose from.

Content assisted by AI (Amy 🌸). Reviewed by the author.