SAN FRANCISCO — The chip is the size of a dinner plate, and that is the point. While the rest of the industry slices silicon wafers into small rectangles, Cerebras Systems builds one giant processor out of an entire wafer, packing computing power that would normally fill a rack of chips onto a single piece of silicon.
That bet is getting new money. Eclipse, a venture firm known for investing in hardware and industrial technology, has taken a stake in Cerebras in a deal that values the company at about $2.5 billion, according to a report from TechCrunch. The size of Eclipse’s investment was not disclosed.
Cerebras has spent years arguing that its approach is right for AI. Its latest chip, the WSE-3, packs roughly 4 trillion transistors and 900,000 computing cores onto a single wafer, a scale that lets it move data between processors at speeds conventional chips cannot match. The company says that design makes it unusually fast at running AI models, particularly for inference — the step where a trained model actually answers questions.
The pitch has found customers in places where speed matters more than price. Cerebras machines run high-performance computing centers, scientific research workloads, and large language models at data centers operated with partners such as G42, the Abu Dhabi technology group. The company says its systems have run some of the largest models in the industry and publishes benchmarks showing its chips beating Nvidia’s on certain workloads.
The market, however, is dominated by a different story. Nvidia’s graphics processors have become the default choice for AI computing, and its software ecosystem, built around the CUDA programming language, makes it hard for rivals to win customers even with faster hardware. Cerebras has positioned itself as the alternative for customers who need raw speed and are willing to work with a smaller software stack.
Eclipse’s investment is notable for what it says about the firm’s thesis. Eclipse has built its name on what it calls the physical world — investments in industrial automation, manufacturing, energy, and robotics — and its partners have argued that the next wave of technology value will come from hardware rather than software. A stake in Cerebras fits that worldview: a chip company that builds its own hardware, operates its own data centers, and sells to institutions that buy physical machines.
The valuation is a talking point in itself. Cerebras was valued at more than $4 billion in an earlier round, and a $2.5 billion valuation now would mark a decline — a sign of how much the AI chip market has changed, or of how much Cerebras needed the money. People familiar with the matter said the deal reflects current market conditions, where private valuations have fallen from their 2021 peaks and investors are demanding more discipline.
Cerebras has other paths to capital. The company has filed to go public, and its earlier plans to list were complicated by questions about its relationship with G42 and the scrutiny that U.S. regulators have applied to exports of advanced chips. A private investment from Eclipse does not resolve those questions, but it gives Cerebras time and cash while it works through them.
The strategic question for Cerebras is whether it can grow beyond its niche. The company has been winning orders for AI inference and for government and research customers, and it has said it is building a dedicated AI data center. Analysts said the path forward depends on whether customers who standardize on Nvidia’s software will be willing to run a second, smaller platform for the workloads where Cerebras is faster.
“There is room for more than one supplier if the performance difference is real,” one analyst said. “The hard part is getting customers to change their habits.”
Eclipse’s other investments suggest it is patient. The firm has backed companies that take years to mature — industrial software, robotics, manufacturing technology — and its partners expect returns from durable positions rather than quick exits. A stake in Cerebras gives Eclipse a seat in the AI hardware race without the risk of building the chips itself.
The deal also reflects a broader shift in how AI investors think. The first wave of AI funding went to software companies building on top of Nvidia’s stack. The second wave, increasingly, is going to companies that own the hardware, the energy, or the physical infrastructure underneath. Cerebras owns its silicon, its systems, and its data centers.
For Cerebras, the money comes at a useful moment. The company has been expanding its sales team, signing up customers beyond its early adopters, and preparing for the day when AI inference demand exceeds what GPU-based systems can deliver. Its wafer-scale design, once dismissed as a curiosity, has become a talking point in a market hungry for alternatives.
The risks are as large as the chips. Manufacturing a processor the size of a dinner plate means every flaw in the wafer can ruin the whole product, and yields are harder to control than with small chips. The company says it has solved those problems, and customers have validated the claim by buying its systems, but the economics of wafer-scale production remain less proven than conventional chipmaking.
Eclipse’s investment is a vote for the long game. The firm’s partners have said they believe AI computing will eventually need more than one architecture, and that specialized hardware built for specific workloads will take share from general-purpose chips. Cerebras, in their view, is the most serious attempt at that thesis in the market today.
Whether the bet pays off will depend on questions that no one can answer yet: how fast AI demand grows and whether Cerebras can turn its technical advantages into durable profits. Eclipse accepted the valuation because it believes the upside is worth the risk. For a company whose chips look like dinner plates, the prize is a place in the future of AI computing.
