{"slug": "etched-raises-300m-at-a-10-3b-valuation-on-a-bet-that-transformer-only-chips-at", "title": "Etched raises $300M at a $10.3B valuation on a bet that transformer-only chips will displace Nvidia at the inference layer", "summary": "Etched, a chip startup founded by three Harvard dropouts, closed a $300 million Series C at a $10.3 billion valuation on July 23, led by Sequoia, to produce its Sohu ASIC that hardwires transformer model inference into silicon, claiming a 20x throughput advantage over Nvidia's H100 GPUs. The company has over $1 billion in signed customer contracts and plans first rack shipments for summer 2026, betting that purpose-built transformer chips will dominate AI inference as the industry shifts from training to serving tokens at scale.", "body_md": "*The three Harvard dropouts behind Etched just closed a $300M Series C at a $10.3B valuation, doubling their worth in seven months, on a single architectural wager: that hardwiring the transformer into silicon beats running it on a general-purpose GPU every time.*\n\nThe round closed July 23, led by Sequoia in what the firm has described as its highest-valued Series C investment ever, with a16z, Jane Street, SK Hynix, and Diffusion joining alongside. That number alone would make Etched a remarkable story. What makes it a significant one is what the company is actually building, and why the math behind it is hard to dismiss.\n\nEtched's chip, the Sohu ASIC, does exactly one thing: run transformer model inference. There are no general-purpose compute paths, no fallback modes. The transformer architecture is literally hardcoded into the silicon, fabricated on TSMC's 4nm process node. The company claims an eight-chip Sohu server hits over 500,000 tokens per second on Llama 70B. An equivalent eight-GPU Nvidia H100 setup produces around 23,000 to 25,000 tokens per second. That's roughly a 20x throughput advantage, and if it holds in production, it reshapes the cost calculus of AI inference entirely.\n\nThose figures haven't been independently verified. Etched hasn't published batch-size numbers at the scale production environments typically require, and no third-party benchmarks exist yet. But the company emerged from stealth on June 30 with working silicon, a rack-scale system already running models including DeepSeek, Qwen, and Llama for test customers, and more than $1 billion in signed customer contracts. Investors aren't buying a roadmap. They're buying early evidence.\n\nThe obvious criticism of transformer-only silicon is the architectural risk: what happens if the field moves past transformers? It's a real question. Etched's answer, essentially, is that the transition timelines are long enough and the economics compelling enough that being the best inference chip for the next several years is a defensible business. Frankly, given that transformers underpin virtually every major deployed AI system, from ChatGPT to Claude to Gemini, that's not a speculative position. It's an observation about where the compute demand actually is right now.\n\nThe broader shift matters here. AI infrastructure spending has moved through a training phase dominated by raw GPU clusters toward an inference phase where the question isn't \"how much compute can we throw at a model\" but \"how cheaply can we serve tokens at scale.\" That transition benefits purpose-built hardware dramatically. General-purpose GPUs are extraordinarily capable at training; they're also expensive and architecturally over-specified for inference workloads. Sohu exists to exploit that gap.\n\nAs TechCrunch noted in its coverage of the round, Etched has had to push back against years of skepticism from investors and engineers who doubted the company could produce working silicon at all. The June 30 stealth exit, with chips running and contracts signed, was a direct answer to that skepticism. Co-founder and CEO Gavin Uberti put it plainly: \"Our chips work, people want them, and it's time to ship.\"\n\nUberti dropped out of Harvard alongside co-founders Chris Zhu and Robert Wachen to start the company in 2022. That origin story matters less than the fact that the trio navigated one of the most capital-intensive industries imaginable - chip design - to functional silicon in roughly four years. Not bad going. Etched has also moved fast on manufacturing infrastructure: a Taiwan facility is already operational, and the company recently opened an 80,000 square-foot facility 15 minutes from its office in Milpitas, California, a 10-megawatt site built to house an NPI lab, an in-house SMT line, and expanded deployment capacity.\n\nFirst rack shipments are scheduled for summer 2026. That's the real test. Signed contracts and stealth demos are promising; production volume at scale is where the architecture either proves itself or reveals its limits.\n\n## What comes after $10.3B\n\nThe story didn't stop at the Series C close. According to reporting from the Wall Street Journal and market data tracked by PYMNTS and MarketScale, Etched is already in talks for a follow-on round targeting a $20 billion valuation, potentially structured as two concurrent transactions, one led by Sequoia, one by Jane Street. Neither has closed, and terms could change. But if the $10.3 billion round was a proof-of-concept moment, the $20 billion conversation is a signal that investors think the proof held.\n\nThat trajectory - $5B to $10.3B to a potential $20B in the span of roughly eight months - puts Etched in territory usually reserved for companies with several years of production revenue behind them. The difference is the size of the prize Etched is pointed at. Nvidia's dominance of AI inference infrastructure is enormous, and the inference layer is where AI compute spending is increasingly concentrated. You don't have to believe Etched will unseat Nvidia to see why betting on them makes sense. You just have to believe the inference economics shift enough to make transformer-specific silicon a viable, large, and durable market. On the current evidence, that's not a hard case to make.\n\n**Also read:** [Paper raised $34M on the exact thesis that wiped out half of Figma's market cap](https://startupfortune.com/paper-raised-34m-on-the-exact-thesis-that-wiped-out-half-of-figmas-market-cap/) • [Nubank is buying a $6 million bank so it can keep calling itself a bank](https://startupfortune.com/nubank-is-buying-a-6-million-bank-so-it-can-keep-calling-itself-a-bank/) • [How to Build a B2B SaaS Sales Funnel Before You Hire a Sales Team](https://startupfortune.com/how-to-build-a-b2b-saas-sales-funnel-before-you-hire-a-sales-team/)", "url": "https://wpnews.pro/news/etched-raises-300m-at-a-10-3b-valuation-on-a-bet-that-transformer-only-chips-at", "canonical_source": "https://startupfortune.com/etched-raises-300m-at-a-103b-valuation-on-a-bet-that-transformer-only-chips-will-displace-nvidia-at-the-inference-layer/", "published_at": "2026-07-25 10:49:08+00:00", "updated_at": "2026-07-25 10:52:45.633172+00:00", "lang": "en", "topics": ["ai-chips", "ai-infrastructure", "ai-startups", "artificial-intelligence"], "entities": ["Etched", "Sequoia", "Nvidia", "TSMC", "Gavin Uberti", "Sohu ASIC", "Llama 70B", "H100"], "alternates": {"html": "https://wpnews.pro/news/etched-raises-300m-at-a-10-3b-valuation-on-a-bet-that-transformer-only-chips-at", "markdown": "https://wpnews.pro/news/etched-raises-300m-at-a-10-3b-valuation-on-a-bet-that-transformer-only-chips-at.md", "text": "https://wpnews.pro/news/etched-raises-300m-at-a-10-3b-valuation-on-a-bet-that-transformer-only-chips-at.txt", "jsonld": "https://wpnews.pro/news/etched-raises-300m-at-a-10-3b-valuation-on-a-bet-that-transformer-only-chips-at.jsonld"}}