# Etched is raising at $10 billion and $20 billion simultaneously and investors are lined up > Source: > Published: 2026-07-23 15:12:29+00:00 *Two-year-old AI inference chip startup Etched is pulling off back-to-back funding rounds at radically different valuations in the same month, with Sequoia leading a deal at roughly $10.3 billion while Jane Street negotiates a separate round at $20 billion.* Etched came out of stealth on June 30, 2026, with a lot to say. The San Jose company, founded in 2022 by Harvard dropouts Gavin Uberti, Chris Zhu, and Robert Wachen, revealed it had quietly raised $800 million across four rounds, signed more than $1 billion in customer contracts, taped out working silicon on TSMC's N4P process, and built a rack-scale 8-chip inference system it calls Sohu. The announcement alone would have been remarkable. Then, within weeks, the Wall Street Journal reported the company was negotiating two more rounds simultaneously: one Sequoia-led deal at a $10.3 billion valuation, and a separate Jane Street-led financing at $20 billion. Neither had closed as of mid-July, and terms could still change. But the structure itself tells you something important about where the AI infrastructure market is right now. Sohu is an application-specific integrated circuit built to run transformer inference and nothing else. That's the entire design premise: strip out every transistor devoted to generality and pour everything into the one workload that actually matters for deployed AI models. Etched claims a single 8-chip Sohu server delivers 500,000 tokens per second on Llama 70B - roughly 20 times the throughput of an H100 on the same task. The chip carries 144GB of HBM3E memory and achieved first-pass silicon success, which is rarer than the industry likes to admit. That matters. First racks are scheduled to ship in summer 2026, though no independent benchmarks exist yet to verify the performance claims. The straightforward pitch is that general-purpose GPUs are overbuilt for inference. Nvidia's H100 and B200 can train models, run simulations, render graphics, and do plenty of things a customer deploying a chatbot will never ask of them. You pay for that flexibility whether you use it or not. Etched's argument is that once models are trained and the architecture has converged on transformers, a chip that does only inference and does it orders of magnitude faster is worth the trade-off. Hyperscalers spending billions on Nvidia hardware are clearly listening: $1 billion in signed contracts before a single rack shipped is not nothing. But the risk is structural, and worth naming plainly. Sohu is a bet that the transformer architecture will remain dominant for long enough to justify giving up all programmability. That's a big bet. If a fundamentally different architecture displaces transformers at scale, these chips become expensive paperweights. State-space models have attracted serious research attention. Mamba and similar approaches avoid the quadratic attention complexity that transformers carry and perform competitively on certain sequence tasks. Etched has said its customers are validating the chip beyond pure transformer workloads, pointing to compatibility with DeepSeek, Qwen, Mamba, and Llama, but the company has not fully explained how a transformer-only ASIC runs architectures that diverge from the transformer pattern. That is the question anyone writing a $10 million check for a 256-chip rack should be pressing hard. ## What the valuation math actually reflects Etched's previous disclosed valuation, from a $500 million round led by growth equity firm Stripes in December 2025 with participation from Peter Thiel, was approximately $5 billion. The Sequoia round would double that. The Jane Street round would quadruple it. If both close, the company will have gone from $5 billion to $20 billion in under eight months, on a product that has not yet shipped to a paying customer at scale. This is not unique to Etched. Back-to-back financings at sharply different valuations have become a defining feature of the 2026 AI investment cycle. The dynamic is straightforward: investors competing for allocations in companies they believe will be essential infrastructure are willing to pay a premium to get in before the next round prices higher. The startups hold all the cards. When a hyperscaler needs to reduce its dependence on Nvidia and there are only a handful of credible alternatives, the price of access goes up fast. Cerebras, Groq, and Tenstorrent are all building in the same inference acceleration space, each with a different architectural bet. Cerebras uses a wafer-scale chip that eliminates inter-chip communication bottlenecks entirely. Groq runs a deterministic dataflow architecture that delivers extremely predictable latency. Etched's approach is the most aggressive: hardwire the chip for one architecture, accept that you can't pivot, and bet that the architecture won't change before you've recouped the investment. It's the highest-risk position in the field. And it's currently attracting the highest valuation. The $20 billion figure, if it closes, would put a two-year-old company with no shipped product at a valuation comparable to established semiconductor firms with decades of revenue. What that number really measures is not what Etched is worth today. It's how desperate large compute buyers are to have an alternative to Nvidia on the shelf when they need it - and how much they're willing to pay to make sure that alternative actually exists. **Also read:** [Apple Maps will be embedded inside Ford's EV platform in 2027, bypassing CarPlay entirely](https://startupfortune.com/apple-maps-will-be-embedded-inside-fords-ev-platform-in-2027-bypassing-carplay-entirely/) • [Cursor builds a CFO playbook for AI spending before SpaceX absorbs it at $60 billion](https://startupfortune.com/cursor-builds-a-cfo-playbook-for-ai-spending-before-spacex-absorbs-it-at-60-billion/) • [Israel wants to become an AI superpower but its power grid may have other ideas](https://startupfortune.com/israel-wants-to-become-an-ai-superpower-but-its-power-grid-may-have-other-ideas/)