Etched Weighs $40B to $50B Offers for Its Nvidia-Challenging AI Chips
Etched is reviewing investment offers that value the company at $40 billion to $50 billion, according to people familiar with the company. The talks were reported on October 5, 2026. Terms could still change. Etched declined to comment. TechCrunch
The $40 billion offers came from top-tier investors, while the $50 billion offers came from lesser-known backers, according to a person familiar with the talks. The discussions remain early. A price in that range would be roughly double the company's last priced round. Etched previously raised $700 million at a $21 billion valuation, led by Jane Street. TechCrunch
That round followed a $300 million round at a $10.3 billion valuation announced in July 2026 and led by Sequoia. The company states that round included Andreessen Horowitz. Etched Etched also states it has raised $800 million across four unannounced financings and received a strategic investment from VentureTech Alliance. Etched
Etched is building complete AI hardware systems powered by its own chips. Co-founders Gavin Uberti and Chris Zhu met in an advanced math course at Harvard. Uberti is co-founder and CEO. The company has 400 employees, with roughly 15% previously at Nvidia. Etched states its engineers also worked on Google TPUs and at Broadcom, SK Hynix and TSMC. Etched
Its approach is narrow by design. Etched claims its chips process more tokens, the units of text an AI model reads and generates, faster and at lower cost than Nvidia chips. Inference, which means running a trained model to answer requests, is the sole focus. The pitch is higher sustained throughput per watt and per dollar on transformer workloads, the architecture behind most large language models, in exchange for giving up generality.
To support that claim, Etched points to two design choices. It states its Low Voltage Inference architecture runs the chip's math blocks at under half the voltage of most AI chips, which lowers power use. It states its Cluster Scale Memory uses a proprietary interconnect, a fast link between chips, with very low delay and high bandwidth to create a shared memory pool across its scale-up domain, or group of linked chips. Etched Etched claims its systems can run trillion-parameter sparse MoE models, very large models that activate only part of their network per task, at over 80% of peak FLOPs, a measure of math operations per second, without thermal throttling, or slowing down from heat.
On manufacturing, Etched states its A0 silicon, the first version of its chip, came back from TSMC N4P, TSMC's 4-nanometer production process, earlier this year. It manufactured its test chip at a TSMC factory in summer 2026 and said in July 2026 it had secured $1 billion in customer orders. The company states it is now validating its first rack-scale product, a full rack of servers, with customers to fulfill $1 billion in demand, with first racks shipping this summer and production started to fulfill over $1 billion in customer contracts.
That customer testing is tied to its last lead investor. Jane Street, an Etched customer, took delivery of an early Etched system. The company states Jane Street led its $700 million funding round after testing Etched's hardware. Etched
Etched operates a 10-megawatt datacenter in Silicon Valley. It states it built a data center, test house and NPI prototyping lab, for testing and preparing new products for production, in its San Jose office. It established a facility in Taiwan to coordinate production near TSMC.
The broader context here is that delivery matters more than the funding figure. A chip designed for one model class lives or dies on utilization, steady performance under sustained load, compiler maturity, and how the interconnect behaves at rack scale. Test silicon plus $1 billion in orders shows demand for an alternative to general-purpose GPUs for dense inference. It does not yet answer questions about yield, bring-up, failure domains, or how quickly customer code reaches that stated 80% of peak without hand tuning.
In my view, the customer-investor overlap deserves attention. Jane Street took hardware before leading a round, which is a stronger signal than a valuation offer. It suggests the diligence included power, throughput and stability measurements on real traffic, not only roadmaps. The open issue is breadth. One trading workload proving out does not establish performance across long-context serving, large sparse experts, and multi-tenant SLOs, or strict reliability targets for shared customers.
Looking ahead, if Etched ships racks on schedule and holds its voltage, thermal and memory-pool claims in production, operators gain another lever for token economics alongside GPUs, TPUs and other inference chips. That would make capacity planning more mixed across hardware types, not simpler, but with lower cost per million tokens where the model fits the hardware.
