Etched

Etched made a wager most chip companies would never dare: that the transformer, the architecture behind every major AI model, would win so completely that you could burn it straight into silicon and never look back. Its Sohu chip does exactly that. Instead of the flexible general-purpose logic a GPU carries, Sohu hardwires transformer inference into the metal, trading versatility for raw speed and efficiency.

The bet nearly killed the company first. Founders Gavin Uberti and Chris Zhu pitched a detailed memo on chip specialization in 2023 and watched every major investor pass, running month to month on fumes. Then the AI market consolidated around transformers exactly as they predicted. By mid 2026 Etched had A0 silicon back from TSMC, more than $1 billion in booked customer contracts, and an $800 million war chest. If transformers stay on top, this is one of the most valuable hardware companies in the world. If the field moves on, Sohu is an expensive paperweight, and Etched has said as much.

Etched does not just sell a chip. It sells the whole rack. The company co-designs chips, racks, software, and even manufacturing methods so frontier models can run inference with better throughput, latency, cost, and power than general-purpose hardware, across both the prefill and decode halves of a request.

The pitch is vertical integration taken to an extreme. Math block designers sit next to inference engineers, thermal experts next to supply-chain managers, all aimed at one goal the company repeats like a mantra: get to gigawatt scale as fast as possible. Its systems target the hardest workloads out there, many-trillion-parameter mixture-of-experts models, long context, and agentic runs, the exact places where GPUs choke.

The core problem Etched attacks is waste. On a general-purpose GPU, transformer inference typically uses only 30 to 40 percent of peak FLOPs, because the chip carries logic for workloads it isn't running and throttles its clock as power and heat climb. Etched calls out two specific bottlenecks and claims to have engineered around both.

Low Voltage Inference runs the chip's math blocks at under half the voltage of typical AI chips, which the company says lets it sustain 80 percent or more of peak FLOPs on trillion-parameter sparse models without thermal throttling. Cluster Scale Memory tackles the other wall, memory latency, with a proprietary low-latency interconnect and an HBM/SRAM hybrid that aims to hit SRAM-class decode speeds without giving up capacity. Both are claims, not yet independently benchmarked, but they name the real pain precisely.

Etched is not selling to hobbyists or small teams. Its buyers are the frontier AI labs, cloud providers, and hyperscalers running transformer inference at a scale where a few percent of efficiency translates into millions of dollars and megawatts. The company says it made co-design decisions hand in hand with leading AI companies and cloud providers, and ran terabytes of production traffic through its simulator before committing to silicon.

That focus is also the constraint. A turnkey inference cluster that ships by the rack and books contracts in the hundreds of millions is a product for a very short list of customers, the ones already spending at data-center scale on AI.

Everything Etched builds points at a single idea: run transformers, run them fast, skip everything else. The Sohu chip is the wager made physical, a transformer-only ASIC fabbed on TSMC's N4P process. Around it the company wraps a full rack-scale system, custom PCBs, cold plates, interconnects, and power delivery, plus the compiler and runtime software to drive it.

The throughput claims are eye-watering. Etched says an eight-chip Sohu server can push more than 500,000 tokens per second on Llama 70B, against roughly 23,000 on eight H100s. No third party has verified those numbers yet, but the roadmap is public: first racks shipping in summer 2026, with performance updates promised alongside.

Etched makes money the old-fashioned hardware way: it sells complete inference systems under large forward contracts, not chips off a shelf. Customers order full clusters, and the company has already booked more than $1 billion in those orders, which it is now racing to fulfill as its first racks ship.

There is no public pricing and no self-serve tier, which fits the buyer. This is enterprise capital equipment sold direct, with a Taiwan factory, a data center, and an in-house prototyping lab all stood up to turn that order book into shipped silicon.

Etched is a direct shot at NVIDIA's grip on AI compute, but it competes by inverting NVIDIA's strategy. Where a GPU is a Swiss Army knife, Sohu is a single blade, useless for anything but transformers and, Etched argues, far better at that one thing. Its rack-scale system is pitched as a turnkey alternative to NVIDIA's DGX and HGX.

The edge and the risk are the same fact. A transformer-only chip can devote all its transistors to the workload that matters today, but if the industry pivots to a new architecture, the company estimates it would take about three years to respond. Etched is betting the leadership pedigree, engineers from NVIDIA, Google's TPU program, Broadcom, and TSMC, plus a head start on transformer-specific design, buys enough time for the bet to pay off.

The concrete milestones are real even where the benchmarks aren't. A0 silicon came back from TSMC's N4P process earlier in 2026, the company is validating its first rack-scale product with customers, and it has booked over $1 billion in contracts against those systems. To run engineering around the clock, it opened a Taiwan factory and built a data center, test house, and prototyping lab inside its San Jose office.

The caution worth stating plainly: the headline performance figures come from Etched, not from independent testing, and its first racks are only shipping now. The order book and the working silicon are the strongest proof on the table.

Etched draws a rare mix of excitement and open skepticism. On Hacker News and in the ML community, people credit the clarity of the thesis and the sheer speed advantage on paper, the idea that stripping out general-purpose overhead could free 2 to 3x more useful compute from the same transistors.

The doubts are just as consistent. Critics note that none of the performance claims have been independently verified, that early materials leaned on renders rather than shipped chips, and that betting the company on transformers is a single point of failure if the architecture ever falls out of fashion. It is admired as a gutsy call and distrusted as an unproven one, often in the same breath.

Etched has raised $800 million in total, and reached a $5 billion post-money valuation on a $500 million round led by Stripes that closed in late 2025 and surfaced publicly in January 2026, with Peter Thiel and Ribbit Capital among the participants. The company also touts a strategic investment from VentureTech Alliance, tied to its TSMC relationship.

What makes the number striking is where it started. In 2023 the founders could not get a single major investor to bite. The cap table now reads like a who's who of AI, angels and backers including Andrej Karpathy, Geoffrey Hinton, Fei-Fei Li, Arthur Mensch, and Stanley Druckenmiller, plus quant firms Jane Street, Hudson River Trading, and Two Sigma.

Etched has done the hard part of a hardware startup, working silicon, a real order book, and a factory to build against it. What it hasn't done is prove the performance claims in the open or ship at volume, and both come due now that first racks are leaving the dock in summer 2026.

The next year is the whole game. Independent benchmarks that back the token-per-second numbers would validate the entire thesis. A transformer-specific chip that lands as promised turns $1 billion in contracts into a durable business. The one risk it cannot engineer away is the one it chose on purpose: if the models the world runs stop being transformers, none of the rest matters.

Etched was founded in 2022 by Gavin Uberti and Chris Zhu, math students and Thiel Fellows who walked away from Harvard, with Robert Wachen as co-founder and president. Around them sits an unusually senior bench: a CTO who ran Cypress, a VP of Platform who spent 22 years at NVIDIA building HGX and DGX, a VP of Software who built Google's TPU software team across five generations, and engineers pulled from NVIDIA, Google, Broadcom, SK Hynix, and TSMC.

The culture is deliberate and demanding. Every technical hire is expected to work across engineering and research with no boundary between them, and the team is fully in person in San Jose at Santana Row, with a second hub in Taipei next to its manufacturing partners. The company's own values are blunt: own outcomes end to end with no training wheels, and treat production, not the prototype, as the real product.

Disclosed pay clusters where you'd expect for a well-funded Bay Area hardware startup. Across the many engineering roles that list a band, base salaries generally run from roughly $130k to $275k, with a handful of specialized and leadership roles reaching to about $300k. Comp is quoted in US dollars, matching Etched's San Jose center of gravity.

Equity is a real part of the package, with roles citing generous equity or option grants on top of base. The benefits lean toward keeping an in-person team fed and close to the office: full medical, dental, and vision, a $2,000 per month housing subsidy for people within walking distance, relocation help to Santana Row, and daily lunch and dinner in the office.

The registered entity is Etched, Inc., a Delaware corporation, operating from 3155 Olsen Dr, San Jose, CA 95117. Its published terms lean enterprise-standard and cautious, with a mandatory arbitration clause, a class-action and jury-trial waiver, and website access granted only for internal evaluation.

Security is treated as existential rather than boilerplate, which fits a company whose crown jewels are chip design files. Job postings describe zero-trust network architecture, segmentation that isolates ASIC development from the rest of the network, and DLP controls built to keep design IP from leaking, with SOC 2 and ISO 27001 named as the frameworks it works toward. Etched does not publish a subprocessors page, so its vendor list isn't public.

Etched spent years heads-down and then arrived loudly. The story that carried in mid 2026 was the same set of facts told two ways: a company emerging from stealth with working silicon and a billion-dollar order book, and an NVIDIA challenger reaching a $5 billion valuation. The $500 million Stripes round was reported earlier in the year and set up the moment.