Etched hits $5B valuation with $1B in AI chip orders

Key Takeaways

• Etched has booked $1 billion in contract orders for its inference cluster systems, with TSMC manufacturing the chips
• The startup raised $500 million in December at a $5 billion valuation, with backing from Peter Thiel, Geoffrey Hinton, and others
• Etched sells complete systems—chips, racks, and software—targeting the inference bottleneck that plagues AI companies at scale

Etched, a two-year-old startup founded by Harvard dropouts, says it has booked $1 billion in contract orders for AI inference systems powered by its custom chips. The company also disclosed a previously unannounced $500 million funding round from December that valued it at $5 billion. TSMC manufactured the first batch of chips earlier this year, and Etched is now testing the hardware with customers.

The orders are for what Etched calls "frontier inference clusters," complete systems that bundle its chips with custom racks and software. The company claims these systems run large AI models faster, cheaper, and more efficiently than Nvidia's GPUs. That's a bold claim, and one that remains unverified by independent benchmarks.

Why inference is the prize everyone is chasing

Inference is what happens after you hit enter on a ChatGPT prompt. The model processes your input and generates a response. For AI companies serving millions of users, inference represents their largest compute cost and their biggest operational headache. Training a model is expensive but finite. Inference never stops.

This explains why investors have suddenly become obsessed with inference hardware. Cerebras launched the first major AI chip IPO this year. Groq raised $650 million. Amazon, Google, and Microsoft all build proprietary inference chips. OpenAI just announced its first custom chip, built by Broadcom. Etched is betting that the market is large enough for specialists to carve out meaningful share from Nvidia.

Who is backing Etched?

The cap table reads like a who's-who of AI and finance. Investors include VentureTech Alliance, Jane Street, Hudson River Trading, Two Sigma, and Ribbit Capital. The angel list is arguably more impressive: Geoffrey Hinton, Andrej Karpathy, Fei-Fei Li, Arthur Mensch, and Scott Wu. Billionaires Stanley Druckenmiller and Peter Thiel also participated.

Co-founders Gavin Uberti (CEO) and Robert Wachen (President) dropped out of Harvard and became Thiel Fellows to start the company in 2022. On Patrick O'Shaughnessy's "Invest Like the Best" podcast, they described pitching a 30-page memo in 2023 arguing that AI would need specialized chips, not general-purpose GPUs. Every major investor passed. The company was reportedly operating month-to-month, nearly out of cash.

Two years later, Etched has raised $800 million total. The fundraising environment shifted dramatically as investors began treating inference optimization as the next bottleneck to solve in AI infrastructure.

What makes Etched's approach different?

Etched builds ASICs (application-specific integrated circuits) designed exclusively for transformer models. This is a deliberate trade-off. Nvidia's GPUs can run any workload. Etched's chips can only run transformers. In exchange for that constraint, Etched claims dramatically better performance on the one architecture that powers ChatGPT, Claude, Gemini, and virtually every frontier model.

The company calls its chip "Sohu." Previous statements from Etched have claimed 8x faster inference than Nvidia's H100 and roughly 20x better power efficiency for transformer workloads. Those are company claims, not independently verified benchmarks. The real test comes as customers put the systems into production.

By selling complete systems rather than chips alone, Etched controls the full stack. This mirrors Nvidia's strategy with DGX systems and allows Etched to optimize software and hardware together. It also means higher revenue per sale and, presumably, fatter margins.

The competitive landscape is getting crowded

Etched faces competition from multiple directions. Nvidia dominates the market and continues to iterate. Cerebras, Groq, and SambaNova all target inference with different architectural bets. Hyperscalers build custom chips for their own data centers. And the AI labs themselves, starting with OpenAI's Broadcom partnership, are moving toward vertical integration.

The question is whether the inference market is large enough for multiple winners. If AI inference costs continue growing as usage scales, there may be room for several specialized vendors serving different niches. If Nvidia manages to close the efficiency gap with its next GPU generation, the window for startups narrows.

Frequently Asked Questions

What is Etched and what does it sell?

Etched is a startup founded in 2022 that builds AI chips specifically designed for transformer model inference. It sells complete systems including chips, racks, and software rather than standalone chips.

How does Etched compare to Nvidia?

Etched claims its specialized chips are 8x faster and 20x more power efficient than Nvidia's H100 for transformer workloads. These are company claims that have not been independently verified.

Who are Etched's investors?

Etched has raised $800 million from investors including VentureTech Alliance, Jane Street, Two Sigma, Ribbit Capital, Peter Thiel, and AI researchers including Geoffrey Hinton and Andrej Karpathy.

When will Etched's chips be available?

TSMC manufactured the first chips earlier in 2026. Etched is currently testing systems with customers and has booked $1 billion in contract orders.

Source: TechCrunch / Julie Bort