Etched exits stealth with $800M raised and a broader inference bet

Why it matters

Etched is trying to move the Nvidia challenge from chip benchmarks to delivered inference systems, where production, software and customer economics decide who gets bought.

Gavin Uberti (@UbertiGavin), Robert Wachen and Chris Zhu's Etched came out of stealth Tuesday with a working chip, first racks built, $800 million raised and what the company says are more than $1 billion in signed customer contracts, according to a 12-post thread on X and a company press release.

The announcement turns Etched from a two-year-old Nvidia challenger with an unusually narrow chip thesis into a systems company trying to sell rack-scale inference capacity before the market's biggest AI buyers finish locking in their next generation of compute supply. Etched said its first racks ship this summer, after what it described as a successful first-pass A0 tapeout, early customer testing and production ramp work in Taiwan and San Jose.

Etched said it has raised the $800 million across four previously unannounced financings. The latest was a $500 million financing in December 2025 at a $5 billion post-money valuation, according to the company. The cap table now includes Jane Street, Hudson River Trading, Jump Trading, Two Sigma, Stripes, Ribbit Capital, Radical Ventures, Primary VC, Positive Sum and a strategic investment from VentureTech Alliance, along with individuals including Peter Thiel (@peterthiel), Geoffrey Hinton (@geoffreyhinton), Andrej Karpathy (@karpathy), Tri Dao (@tri_dao), Noam Brown (@polynoamial), Scott Wu (@ScottWu46) and Arthur Mensch (@arthurmensch).

The strongest part of Tuesday's announcement is not the round size. It is the claim that Etched has moved from a chip story to hardware that customers are testing. The customers are unnamed, the performance numbers are not disclosed, and the contract structure is not described. Those omissions matter because hardware startups can book demand before they prove manufacturing yield, field reliability and service economics. But Etched is now making a more concrete claim than it made in 2024: it says A0 silicon came back from TSMC's N4P process, that it is validating a rack-scale product with customers, and that its systems are running DeepSeek, Qwen, Mamba and Llama workloads.

That last list is the tell. In 2024, Etched's public pitch centered on Sohu, a transformer-specialized ASIC. TechCrunch reported then that the company had been founded by Harvard dropouts Uberti and Zhu, later joined by Wachen and former Cypress Semiconductor CTO Mark Ross, to build a chip that did one thing: run AI models based on the transformer architecture. The bet was explicit: give up generality to gain speed and cost efficiency in the architecture behind large language models.

Tuesday's messaging is broader. Etched now describes itself as building "frontier inference clusters" for prefill and decode workloads, many-trillion-parameter mixture-of-experts models, long context and agentic workloads. Its website says the company has co-designed chips, packages, PCBs, cold plates, interconnects, racks, software and manufacturing methods around that goal. Etched is no longer asking customers to believe in a single chip spec. It is asking them to believe the company can deliver a full production system.

That repositioning is rational. The buyers Etched wants do not buy benchmark slides. They buy deployable capacity, thermals, uptime, software integration and supply. Nvidia's advantage is not only silicon performance. Nvidia reported record fiscal Q1 2027 revenue of $81.6 billion, with data center revenue of $75.2 billion, evidence that the market is still buying full-stack AI infrastructure at scale. To compete, Etched has to convince customers that specialization can beat Nvidia's general platform where inference cost and latency matter most.

Etched is also making a manufacturing claim. The company said it has opened a Taiwan factory for 24/7 engineering cycles and built a data center, test house and new product introduction prototyping lab at its San Jose headquarters. Its site says it built a 2 megawatt data center in the office, a detail that captures the operating premise Wachen put in the press release: "Production is the product." That is a more grounded slogan than most AI infrastructure marketing because the risk for a young chip company is almost always production, not ambition.

The founder story fits the product risk. Uberti's official Etched bio says he is a Harvard Thiel Fellow, Math 55 alumnus and AI compiler engineer who developed the Cortex-M backend for TVM. Wachen is listed as a Harvard Thiel Fellow who co-founded Prod and Mentor Labs, and Etched now lists Ross as CTO, with a company-supplied note that he previously served as Cypress CTO and shipped systems on A0 silicon. The new leadership page also names executives with Nvidia, Google TPU, Broadcom and supply-chain backgrounds, including Brian Loiler, who Etched says spent 22 years at Nvidia and helped build HGX and DGX systems, and David Munday, who Etched says built the TPU software team from v1 through v5.

The hiring disclosures matter because Etched's current promise depends on coordination across silicon, memory, networking, thermals and software. The company says it now has more than 400 engineers from Nvidia, Google TPUs, Broadcom, SK Hynix, TSMC and quantitative trading firms. That is a large headcount for a startup that only publicly announced a $120 million Series A in June 2024, when Reuters reported Etched was building a processor tuned to run transformer models and had been valued at $34 million at its 2023 seed round.

Etched's technical framing on Tuesday centered on two named approaches: Low-Voltage Inference and Cluster-Scale Memory. Low-Voltage Inference is Etched's answer to thermal throttling at high FLOPs utilization: the company says it runs math blocks at less than half the voltage of most AI chips and can run trillion-parameter sparse mixture-of-experts workloads at more than 80% of peak FLOPs without thermal throttling. Cluster-Scale Memory is its answer to decode latency: Etched says it built a lower-latency shared memory pool across its scale-up domain using a proprietary high-bandwidth interconnect and an HBM/SRAM hybrid design.

Those are company claims, not independent benchmarks. Etched did not publish rack pricing, named customer results, contract durations, gross margin targets, production capacity or comparative numbers against Blackwell or Rubin systems. It also did not disclose how much of the $1 billion in customer contracts is contingent on delivery, performance, acceptance testing or financing. For a company selling hardware into hyperscale infrastructure cycles, those details separate demand from durable revenue.

Still, the announcement marks a real escalation. Inference has become the cost center AI labs and cloud providers cannot avoid: every agent session, coding task, voice interaction, search result and long-context workflow consumes serving capacity after the model is trained. Etched's bet is that a narrower system, optimized around those workloads, can undercut general-purpose GPUs where the next wave of AI usage actually runs. The next proof point is not another financing. It is whether the racks that ship this summer perform in customer environments at the throughput, latency and power levels Etched is now putting at the center of the company.