Etched, a four-year-old chipmaker built by a trio of Harvard dropouts, has raised $300 million in a Series C round that pushes its valuation to $10.3 billion — more than double the $5 billion mark it commanded in December 2025. Sequoia Capital led the round, joined by Andreessen Horowitz, memory giant SK Hynix, trading firm Jane Street, and Diffusion, with individual backers including Peter Thiel, Andrej Karpathy, Figma’s Dylan Field, and Replit’s Amjad Masad also participating. Sequoia has called it the highest valuation it has ever backed at the Series C stage.
The pitch is narrow and, on paper, almost heretical: rather than building a general-purpose GPU that can train and run any kind of AI model reasonably well, Etched builds chips that do one thing — running already-trained models, a process called inference — and does it aggressively better than anything general-purpose can. CEO Gavin Uberti and co-founder Chris Zhu started the company in 2022 on the bet that as AI shifts from a training-bound industry to an inference-bound one, the market would eventually reward specialization over flexibility. Three years and roughly $800 million in outside capital later, that bet looks less contrarian than it did.
Etched’s architecture splits the inference problem into two custom pieces. A “prefill” chip processes an incoming prompt at a much lower voltage than competing AI silicon, which means less heat and, in turn, room for more transistors on the same die. A separate memory and interconnect system handles “decode” — generating a model’s response token by token — by letting many chips share memory at very low latency, avoiding the bottleneck that usually slows a large model down as it writes out an answer. Chief operating officer Robert Wachen has described the approach simply as low-voltage inference, but the more important claim is architectural: the chips are not tied to one model family. Early skepticism accused Etched of building silicon that would only run a narrow set of large language models — a fair worry, since specialized chips built around one architecture obsolete themselves the moment that architecture falls out of favor. The company says it has since demonstrated the opposite: its stack runs standard transformer models, mixture-of-experts systems like those behind DeepSeek and Qwen, and newer non-transformer designs such as Mamba, without a redesign.
That flexibility is what turned a hardware story into a revenue story. Etched disclosed in June 2026 that it had booked more than $1 billion in customer orders, manufactured through TSMC, and is running a 2-megawatt data center at its San Jose headquarters while building a second, 10-megawatt, 80,000-square-foot facility in Milpitas to handle demand. Wachen, notably, hasn’t dressed the moment up as easy. “We had no idea how hard it was going to be,” he said of scaling the company from a chip design into a functioning data center operator. “I think we still have to be humbled by what it will take to actually get to scale.”
The bigger context is what makes this round worth more than a chip-nerd curiosity. Nvidia’s dominance in AI hardware was built on training — the extremely compute-heavy process of building a model in the first place — but the industry’s own cost structure is shifting under it. Once a model exists, it gets called billions of times a day by users and other software, and every one of those calls is an inference workload running, for most companies, on a general-purpose GPU doing a specialized job inefficiently. That gap is exactly where Etched, alongside a widening field of inference-focused challengers, is trying to wedge itself in. A $10.3 billion valuation for a company with no history of shipping consumer hardware is aggressive by any ordinary measure — but it is aggressive in proportion to how large investors think the inference market alone will become as AI usage, not just AI training, turns into the dominant cost line for every software company on earth.
It’s also a useful data point on where late-stage AI capital is actually flowing this year. The largest checks in AI right now are not going to yet another consumer chatbot or another foundation-model lab chasing the next benchmark — they are going to the unglamorous infrastructure layer underneath all of it: chips, power, and memory. Etched’s doubled valuation in seven months says investors increasingly believe that layer, not the application layer sitting on top of it, is where the most durable and defensible businesses in this AI cycle will actually get built.
What This Means for Philippine Founders
Etched itself is irrelevant to a Philippine startup’s roadmap — nobody here is designing inference silicon. What is relevant is the underlying signal: the most capital-efficient AI bets right now are not “we built a better chatbot,” they are “we found one narrow, expensive, unglamorous layer of the AI stack and made it dramatically cheaper or faster.” Philippine founders building on top of AI — in fintech, business process outsourcing tooling, agritech, or logistics — should read Etched’s raise as a reminder that global investors are rewarding infrastructure-level specificity over horizontal ambition. A Philippine startup that can credibly claim it makes one specific, expensive AI workload — translation across Philippine languages, voice inference for call center automation, or document processing for bank and government compliance — meaningfully cheaper or faster than a generic API call has a sharper pitch than one selling a broad “AI platform for X.” The valuation math obviously won’t translate; Philippine seed and Series A rounds are priced in millions, not hundreds of millions. But the underlying investor logic — that narrow and technically defensible beats broad and generic — travels well regardless of check size.
Share this article