Etched weighs $40B–$50B valuation offers — four times its July price in three months
Etched, a U.S. AI chip startup, is reviewing investment offers that would value it at $40 billion to $50 billion, TechCrunch reported on October 5, 2026. That compares with $10.3 billion in a Sequoia-led round in July and $21 billion when it raised $700 million in September — roughly a fourfold rise in three months. Etched designs chips used only for AI inference, not training, and sells complete server systems built around them. It reported $1 billion in customer orders as of July, and trading firm Jane Street has received an early system. Talks are at an early stage and terms could change
The three lines
- Numbers — $10.3B in July → $21B in September → $40–50B offers in October
- Pitch — inference-only chips sold as complete systems; $1B in orders by July; Jane Street has an early system
- Caveats — offers, not a closed deal; revenue, volumes and chip performance are undisclosed
Key questions
- What is Etched valued at
- **Offers range from $40 billion to $50 billion.** | Date | Raise | Valuation | |---|---|---| | July 2026 | $300M (Sequoia-led) | $10.3B | | September 2026 | $700M | $21B | | October 2026 | Offers under review | $40–50B | Top-tier investors reportedly offered $40B; lesser-known backers offered $50B.
- Etched vs Nvidia
- **Etched is inference-only; Nvidia GPUs do both training and inference.** | Item | Etched | Nvidia GPU | |---|---|---| | Use | AI inference only | Training and inference | | Sold as | Chip plus full server | Chips, boards, racks | | Claimed edge | More tokens, faster, cheaper | Flexibility, software ecosystem | | Risk | Model architecture shifts | Price, power |
- Who founded Etched
- **Gavin Uberti and Chris Zhu, who met at Harvard, with co-founder Robert Wachen.** | Item | Detail | |---|---| | Staff | ~400, about 15% ex-Nvidia | | Facilities | 10MW data center in Silicon Valley; site in Taiwan near TSMC | | Customer | Jane Street (early system, also an investor) |
AI chip startups are now doubling in value by the month, not the year. Etched, which builds chips for AI inference only, is reviewing investment offers that value it at $40 billion to $50 billion, TechCrunch reported on October 5, 2026. In July it was valued at $10.3 billion. Talks are early, but the price shows where AI money is heading.
1. Three months of numbers
| Date | Raise | Valuation | vs July |
|---|---|---|---|
| July 2026 | $300M (Sequoia-led) | $10.3B | 1x |
| September 2026 | $700M | $21B | ~2x |
| October 2026 (offers) | TBD | $40–50B | ~4–5x |
Top-tier firms reportedly offered $40 billion and lesser-known backers $50 billion — only weeks after the September round closed.
2. What Etched sells
Etched does not just sell chips. It designs inference-only silicon and the complete servers around it. Inference is the stage where a trained model answers users; the company says its systems process more tokens, faster and more cheaply than Nvidia GPUs.
| Item | Detail |
|---|---|
| Founders | Gavin Uberti and Chris Zhu (met at Harvard); co-founder Robert Wachen |
| Staff | ~400, about 15% from Nvidia |
| Product | Inference chips plus full AI server systems |
| Facilities | 10MW data center in Silicon Valley; Taiwan site near TSMC |
| Orders | $1 billion (as of July 2026) |
| Customer | Jane Street — early system delivered; also an investor |
The investor logic: as AI usage explodes, spending is shifting from training to inference, and inference is repetitive enough that purpose-built chips can beat general-purpose GPUs on efficiency.
3. The risks
Specialised chips cut both ways. Hard-wiring today's model structure makes them fast, but a major architectural shift could leave them behind. Building full systems brings bigger ticket sizes and much bigger capital needs. Etched also competes with Google, Amazon and others designing their own accelerators.
There is a Korean angle: inference chips still need high-bandwidth or fast DRAM, so every non-Nvidia chipmaker that scales is another potential customer for Korean memory makers. Etched has not said whose memory it uses.
4. What remains unconfirmed
- The valuation range reflects offers; no deal has been confirmed.
- Revenue, production volumes and independent performance data are undisclosed; "cheaper and faster than Nvidia" is a company claim.
- The bidders were not named.