Etched and the chip market

Data centers and AI accelerator market is nearing 1T a year. Investments in Cerebras, SambaNova, Groq, and others reveal a broad bet that specialty hardware will fragment away from Nvidia's monopoly.

Etched was founded in 2022 by Harvard dropouts Gavin Uberti, Chris Zhu, and Robert Wachen with the goal of creating specialized chips designed to accelerate AI inference. Their core product is the Sohu chip — an AI chip optimized specifically for transformer models, embedding the transformer architecture directly into silicon to pioneer servers for transformer inference.

Competitors

• Groq — Developed the Language Processing Unit (LPU) focused exclusively on delivering high speed and energy efficiency for AI inference. Was valued at $6.9B before being acquired by Nvidia.
• Cerebras — Known for its massive Wafer-Scale Engine (WSE-3) with 4 trillion transistors and 900,000 AI cores on a single silicon wafer. Raised $1B at a $23B valuation and is refiling for an IPO.
• Tenstorrent — Led by legendary chip architect Jim Keller (AMD Zen, Tesla self-driving). Uses open-source RISC-V architecture. Raised $700M Series D at a $2.6B valuation, backed by Samsung, LG, and Jeff Bezos.
• SambaNova — Unveiled the SN50 chip in February 2026, claiming 5x more compute per accelerator and 4x more network bandwidth than its previous generation, with support for 10 trillion+ parameter models.

Incumbent giants: Nvidia (dominant, ~70–95% AI accelerator market share), Google TPU, AWS Trainium/Inferentia, and AMD.

As of March 2026, Sohu hasn't shipped to customers. No third-party benchmarks exist. No inference provider has published production throughput numbers. The 500,000 tokens/second claim comes entirely from Etched's own marketing materials. This makes it a pre-revenue hardware bet valued at $5 billion.

Valuation and future

Etched's specific niche is transformer inference ASICs — purpose-built silicon for running LLMs, not training them from scratch. Etched's CEO says their first product has an order of magnitude more throughput and lower latency than Nvidia's B200, and that with Etched ASICs, you can build products that would be impossible with GPUs, like real-time video generation and extremely deep chain-of-thought reasoning.

These massive performance claims, if true, are transformative. The $500M funding backs Sohu's claims of 10–20x better performance-per-watt and raw throughput versus Nvidia's H100 — on TSMC's 4nm process.

Etched has gone through a remarkable valuation escalation in a short time:

• 2023 — Seed round, modest valuation
• Mid-2024 — Raised $120 million in a Series A led by Positive Sum and Primary Venture Partners, with backers including Peter Thiel and Replit CEO Amjad Masad
• Early 2025 — Secured $85 million at a $1.5 billion valuation, following two stealth rounds at $500M then $750M valuations in rapid succession
• Early 2026 — Raised $500 million in a Stripes-led round at a ~$5 billion valuation, bringing total funding to nearly $1 billion

The company's strategy involves breaking backward compatibility to optimize solely for LLM workloads. The startup targets completing its final chip design in 2026, with manufacturing and shipping anticipated by 2027

Etched has a lot of risks however:

1. Transformer-only is an existential single point of failure. Sohu can't run CNNs, LSTMs, SSMs (Mamba), or any non-transformer architecture — a deliberate and permanent trade-off. Unlike general-purpose GPUs, which can be reprogrammed via software to handle new architectures, Etched's rigid hardware design leaves it highly vulnerable to sudden shifts in AI development.
2. Hyperscaler competition. Being out of the research/model serving loop itself, Etched doesn’t really have an advantage on research, even if we continue sticking with transformers. Google TPUs for example can be designed to fit their own models better.
3. The software ecosystem problem. Winning adoption among hyperscalers and major model labs requires not just raw performance, but a mature software stack, deep framework integrations, and sustained support — areas where Nvidia's CUDA has a decade-long moat.

The $5B valuation with ~zero ARR is justifiable only if you believe (a) transformers dominate AI for the next decade, (b) Sohu ships on schedule and performs as claimed, and (c) they can win customers away from Nvidia's deeply entrenched ecosystem.