“Our company only makes chips that run Transformers.” Two years ago, that pitch would have earned a polite smile and a shredded business card. This week it was worth five billion dollars.
On June 30, Silicon Valley AI chip startup Etched ended its stealth in one salvo: its Transformer-specialized chip Sohu has completed tape-out, the company has raised $800 million to date (including a fresh $500 million round), and customers have signed $1 billion in pre-orders. The investor and backer list reads like half the AI establishment — Andrej Karpathy, Geoffrey Hinton, Fei-Fei Li, Peter Thiel. But the money is not the story. The story is what the money is betting on: AI silicon is splitting from “one architecture serves all workloads” into a dedicated-hardware schism, and Etched’s ten-figure order book is the first billion-dollar validation of that split.
The Bet: A Die With the Generality Carved Out
Understanding Etched means understanding the fault line between an ASIC and a GPU. An NVIDIA GPU is a general-purpose machine: it runs Transformers, CNNs, mining workloads, and anything else CUDA can express. Generality has a price — on any specific task, a large share of transistors and memory bandwidth sit idle. During Transformer inference, a meaningful fraction of the GPU’s logic units spends its cycles on computation paths that have nothing to do with attention.
Sohu takes the opposite road: the Transformer’s attention computation is etched directly into silicon as fixed-function logic rather than executed as software on programmable units. No kernel launch latency. No memory allocation overhead. No scheduler decisions. The chip contains physical circuitry for attention and nothing else.
The cost of that purity is total: convolution will not run, state-space models will not run, diffusion U-Nets will not run. If a different architecture displaces the Transformer next year, Sohu becomes very expensive scrap. This is an extreme wager stated in one sentence: the Transformer architecture is now stable enough to be worth surrendering all programmability for.
The Numbers — With the Usual Caveat
Etched’s claimed figures deserve attention, filtered through the standard disclaimer that vendor benchmarks await independent verification. A server of eight Sohu chips supposedly delivers 500,000 tokens per second of throughput on Llama 70B inference, against roughly 700 tokens per second for a single H100 in the same scenario. That works out to one Sohu server substituting for approximately 160 H100s, with performance per dollar claimed at 140 times the GPU.
Whatever the exact multiple settles at once third parties measure, a gap of this order stops being “a better GPU” and starts being a different economic model. When inference costs drop by two orders of magnitude, entire categories that never penciled out — real-time agent collaboration, long-context reasoning at scale, full-corpus search — flip from theoretically possible to commercially runnable. That flip is the rational core of a $1 billion commitment signed before mass production: nobody pre-orders at that scale on faith. Customers have run real workloads on early silicon, and the results were good enough to sign for.
The Three-Way Fork in AI Silicon
Etched is not an isolated anomaly; it is the loudest data point in a six-month pattern. Cerebras completed a $5.5 billion IPO. Groq raised $650 million. OpenAI shipped its first custom chip with Broadcom, while Amazon, Google, and Meta all press their own silicon programs. The industry is converging on a fork with three roads:
Road one: general-purpose GPUs — NVIDIA’s shovel-selling model. NVIDIA sells universal shovels regardless of what you mine. The advantages remain real: a mature ecosystem, deployment flexibility, zero model-migration cost. The disadvantage is increasingly itemized on invoices: for customers whose workload is already known, the transistors and bandwidth wasted on generality are pure, measurable cost. When inference scales from thousands of cards to hundreds of thousands, the generality tax stops being negligible friction and becomes hundreds of millions of dollars a year in pure spend.
Road two: cloud-vendor silicon — dedicated but closed. Google’s TPU, AWS’s Trainium, Meta’s MTIA are fused to their home platforms and business requirements: costs tightly controlled, optimizations extreme, ecosystems sealed. You cannot buy a Trainium and run your own stack on it, and the innovations mostly stay in-house.
Road three: third-party specialized ASICs — the road Etched defines. Cerebras, Groq, and Etched build neither general chips nor captive ones; they build intensely optimized silicon for a specific computation pattern and sell it to everyone. Their shared thesis: when the software ecosystem around a workload is sufficiently standardized, a dedicated chip’s cost advantage can repay every concession generality gave up. Etched can claim “customers change zero code” — PyTorch-trained models run as-is — because the Transformer world already standardized model formats and inference interfaces. The ASIC only has to meet the ecosystem at that uniform layer. That is historically rare. ASICs have always meant new programming models, new software stacks, new migration costs. Sohu dodges all three not because Etched engineered them away, but because the Transformer ecosystem did the standardization work first.
The Zeroing Framework: Why the Generality Premium Shrinks
The three roads answer the same question differently: when a tool spends 99 percent of its life doing one thing, is the premium paid for the other 1 percent still worth it? The answer depends on two variables moving simultaneously.
Variable A: workload certainty. In 2018, nobody knew whether Transformers would still be the winning architecture in five years, so buyers purchased GPUs — the option value of switching tracks. In 2026, Transformers govern language, images, video, code, and music generation. That option has largely expired worthless.
Variable B: absolute cost at scale. With a few dozen cards, the GPU’s efficiency discount is tolerable. At hundreds of thousands of cards, the discount decides whether your product can be priced to reach its market.
Call it the zeroing framework: as workload certainty approaches 100 percent and scale approaches infinity, the premium for generality approaches zero. This does not dethrone NVIDIA. General GPUs retain irreplaceable value in research prototyping, architecture exploration, and multi-task settings. But inference — the largest and fastest-growing cost center in AI — is decoupling from “general” and entering the dedicated-optimization battlefield at full speed.
The same zeroing shows up wherever you look. On the model side, small specialized models are taking vertical-market share from general giants — not because the giants got worse, but because for a fixed task domain, a smaller specialized model wins on ROI. On the infrastructure side, dedicated inference providers price well under general-purpose cloud APIs for the same reason: they optimize every layer of the stack instead of renting general capacity. On the enterprise side, “one model for everything” is giving way to “a matched model per task.” That is not regression; it is the standard maturity path of any technology: first unify, then differentiate.
What Each Role Should Do About It
If you run a foundation-model company: the next order-of-magnitude drop in inference cost will not arrive inside a better GPU. If you are a serious player and have not started a POC with an Etched or a Cerebras, assume your inference cost curve will trail competitors by a generation within eighteen months.
If you are an application CTO: before 2027 your inference architecture options expand from “NVIDIA, by default” to at least three roads. Start testing through API access now, on your real workloads — migration costs and compatibility surprises are only discoverable by running, not by reading spec sheets.
If you invest: the winner-take-all narrative in AI silicon is loosening. NVIDIA will not fall — but it may travel from “the infrastructure” to “one of the infrastructures,” and that transition from one to many will mint new tens-of-billions companies while burying teams that sign orders they cannot deliver. Betting on whether Etched can ship on Sohu’s promises is a more interesting trade than betting on whether NVIDIA grows.
If you simply watch this industry: Etched does not prove “ASICs win.” It proves the 2023 consensus of GPU supremacy has lost explanatory power worth auditing. Run the zeroing framework across your own domain: wherever workloads have hardened and scale has exploded, look for the generality premium — the crack is already there.
Note: technical figures come from Etched’s announcements and third-party coverage (TechCrunch, Spheron analysis, QbitAI, Leiphone) and have not yet been independently verified.
