The AI Inference Chip Worth Ten Billion Dollars
The clearest threat yet to Nvidia's dominance is an AI inference chip that refuses to do anything except run transformers. This week Etched closed a $300 million funding round that values it at $10.3 billion, staking the entire company on exactly that idea. There is a version of the artificial intelligence story where Nvidia never loses, where every model, every chatbot and every agent runs on its graphics processors and the company collects a toll on all of it. A small firm in California has just made that version look less certain, and it is a narrow bet where the narrowness is precisely the point.
Etched was founded in 2022 by three Harvard dropouts, Gavin Uberti, Chris Zhu and Robert Wachen, who decided early that the future of computing would settle on one neural network design and built a processor for that world. The Series C was led by Sequoia Capital, with participation from Andreessen Horowitz, the trading firm Jane Street and the memory manufacturer SK Hynix. According to people involved in the deal, it is the highest valuation Sequoia has ever assigned to a Series C it has led, and it more than doubles the roughly $5 billion price the company carried earlier this year.
The valuation is startling because Etched sells a product most of the industry would call reckless. Its chip, called Sohu, cannot train models. It cannot run the diffusion systems behind image generators. It does one job, running transformer models for inference, and it does that job in hardware so fixed that changing the underlying math would mean designing a new chip. In a business that worships flexibility, Etched has wagered ten billion dollars that flexibility is overrated.
Why a Chip That Does Less Might Win
To understand the bet, it helps to understand what an AI inference chip actually spends its time doing. Inference is the moment a trained model answers a question, writes a paragraph or generates a line of code. As chatbots and agents reach hundreds of millions of users, inference, not training, has become the largest recurring cost in artificial intelligence. Nvidia's processors handle it well, but they were built to be general. They carry circuitry for a wide range of tasks, and when they run a transformer they leave much of their theoretical power idle.
Etched's argument is that generality has become a tax. By hardwiring the transformer's operations, the attention mechanism, the projections, the softmax step and the normalization layers, directly into silicon, its AI inference chip strips away the overhead a programmable processor drags around. Data flows through a fixed pipeline instead of waiting on an instruction scheduler. The company claims this keeps Sohu's arithmetic units busy more than 90 percent of the time, against roughly 30 percent for a general purpose GPU running the same attention heavy work.
That is a philosophical break from how the industry has thought about chips for a decade. The prevailing wisdom held that model architectures would keep changing, so the hardware beneath them had to stay adaptable. Etched looked at the same landscape and concluded the opposite, that the transformer has won so decisively that betting against it is the safer trade.
The Numbers Behind the Claim
The performance figures Etched publishes are, if accurate, hard to dismiss. Sohu is manufactured on TSMC's N4P process and paired with 144 gigabytes of HBM3E memory. The company says a single server holding eight of the chips can process more than 500,000 tokens per second running Llama 70B, a widely used open model. An equivalent server built from eight Nvidia H100 chips manages roughly 23,000 tokens per second on the same task, and even Nvidia's newer B200 systems land near 45,000. That gap is the basis for the headline claim that this AI inference chip runs transformers around twenty times faster than the hardware it hopes to replace.
Numbers like these arrive with an obvious caveat. They are the company's own, produced under conditions the company chose. As of the middle of this year, Sohu had not shipped in volume to outside customers and no independent laboratory had published its own benchmarks. The distance between a controlled demonstration and a data center absorbing real traffic is exactly where many promising chips have quietly failed.
Still, the direction of the claim matches where the money is going. Every large buyer of AI hardware is now hunting for ways to cut the cost of inference, and an AI inference chip that could do the job at a fraction of the power and price would reset the economics of running a model at scale.
The Investors Reading the Same Signals
What lifts this round above curiosity is who is paying for it. SK Hynix is not a passive check. It is one of the two companies that matter most in high bandwidth memory, the scarce component every AI accelerator now depends on, and its presence signals that a critical supplier takes Etched's roadmap seriously. Sequoia leading its most expensive Series C ever, alongside Jane Street, a firm that lives and dies by microscopic efficiencies, points to the same conviction from the financial side.
The logic is straightforward. The largest buyers of AI hardware, the cloud providers and frontier labs, want a second source with real urgency. Dependence on a single vendor inflates costs, weakens negotiating leverage and concentrates risk. Even a partial alternative that lowers the price of inference is worth enormous sums, which is how a company with roughly $1 billion in signed contracts and no volume shipments can command an eleven figure valuation. Blumefield has tracked the widening field of custom silicon challengers trying to pry open Nvidia's grip, and Etched is now the most richly valued of them.
The Risk Etched Cannot Design Away
Etched's greatest strength is also its deepest weakness. The company has bet everything on the transformer, the architecture behind almost every major model since 2017. If that dominance holds, an AI inference chip built exclusively for transformers could seize a meaningful share of the fastest growing cost in computing. If the field moves on, toward state space models, new attention alternatives or something not yet invented, a processor with the transformer literally etched into its circuits cannot follow. Nvidia's generality, the very thing Etched mocks, doubles as an insurance policy.
Then there is the ecosystem. Nvidia's dominance rests as much on its CUDA software as on its silicon, a mature stack developers know intimately. Etched must convince customers to rebuild parts of their workflow around unfamiliar hardware, a hurdle that has broken better funded rivals. The $10.3 billion valuation assumes the company clears all of it, ships at scale, and does so before the architectural ground shifts underfoot.
For now Etched has done something rare. It has made a credible, specific and testable claim that the most valuable company in the world can be beaten at one of its core tasks. Whether the bet pays off will be settled not in a funding announcement but in the data centers that either adopt this AI inference chip or leave it as the most expensive engineering demonstration of the decade. The rest of the industry, and Nvidia above all, will be watching to see which it becomes. Readers can follow the story as it develops at Blumefield.