Nvidia, with a market capitalization of $5.5 trillion, has met a formidable competitor.

Article author and source: Pencil News

Nvidia, with a market capitalization of $5.5 trillion, has met a formidable competitor.

Soon, chip company Cerebras will IPO on Nasdaq. Its IPO price is set at $189, raising $5.55 billion and valuing the company at $56.4 billion (approximately 380 billion yuan).

This is the largest IPO globally this year, and it is also likely to be the largest IPO in the first half of the year.

Why is Cerebras so popular? Because it's trying to challenge Nvidia's dominance in computing chips.

Cerebras truly has a unique skill.

Nvidia and AMD's GPUs are made by cutting a whole wafer into many small chips and then packaging them. Cerebras does the opposite; it uses a whole 12-inch wafer as a giant chip for AI inference.

This is the largest chip in human history, and OpenAI placed a $10 billion order.

Come on, let's see how it works.

Brute force thinking to solve long-standing problems

Founded in 2016 and headquartered in Sunnyvale, California, Cerebras' core business is developing AI training and inference chips, as well as related servers and cloud services.

The company's core product is called Wafer-Scale Engine (WSE). It has now reached the third generation, WSE-3.

How incredible is this chip?

A traditional GPU is about the size of a postage stamp. Cerebras' WSE, on the other hand, is a single wafer, roughly the size of a dinner plate. It boasts approximately 4 trillion transistors and over 900,000 AI computing cores, operating as a single chip across the entire wafer.

It attempts to address a growing problem in the AI industry: GPU clusters are increasingly resembling building blocks.

Training large models now requires thousands or even tens of thousands of GPUs, connected via a high-speed network.

However, the more GPUs you have, the higher the communication latency. Often, large models aren't unable to be computed, but rather get stuck on communication issues.

Cerebras' approach is very radical: instead of building a lot of GPUs, they create a single, super-large chip. It integrates computing, storage, and network communication onto a single wafer. Data doesn't need to be frequently moved between many GPUs.

The advantages of doing this are obvious: lower latency, lower power consumption, faster training and inference speeds, and easier scaling of very large models.

It is considered the most unconventional approach outside of NVIDIA's GPU system.

Where did this mind-blowing idea come from? To answer that, we have to mention the founder.

Founder Andrew Feldman is a serial entrepreneur who co-founded SeaMicro, a company that makes low-power servers, before founding Cerebras.

In 2012, SeaMicro was acquired by AMD for approximately $334 million.

He later realized that the cost of data movement in deep learning was becoming increasingly high. If the AI era continues to use the chiplet-based approach of CPUs/GPUs, it will inevitably encounter bottlenecks sooner or later.

So he decided to do something that many in the industry thought was impossible: to make wafer-level chips.

Industry insiders say he's crazy. Because if you make it into a single chip, if even one part fails, the entire chip might be unusable. That's why wafers must be diced.

Feldman once again thought in reverse: While I cannot make the entire chip completely flawless, I can make the system see and avoid flaws while it is working, and only work in the good areas.

In 2019, Cerebras released its first-generation wafer-level chip. They actually managed to pull it off.

Cerebras has raised approximately $2.5 billion to $3 billion in total funding, which is relatively high for an AI chip startup. Its valuation after its last funding round was $23 billion. Just one day before its IPO, its valuation was adjusted to $56.4 billion.

Revenue soared as customers placed another $24.6 billion in orders.

Cerebras' rise to prominence is a recent phenomenon, occurring within the last two years: AI inference is becoming increasingly important.

Cerebras' biggest selling point is its AI inference chip. It is positioning itself as an alternative to GPUs for AI inference because of its ultra-large single chip, ultra-high memory bandwidth, low GPU communication overhead, and exceptionally low inference latency.

Cerebras' revenue surged from $24.6 million in 2022 to $510 million in 2025, a more than 19-fold increase in four years, with a 76% year-on-year increase in 2025; net profit turned from a net loss of $482 million in 2024 to a profit of $238 million in 2025, successfully achieving a turnaround.

Very few AI hardware companies can achieve the growth rate that Cerebras has.

However, its revenue is highly concentrated. In 2025, 86% of its revenue will come from two Middle Eastern clients, G42 and MBZUAI.

G42 is a technology company controlled by the UAE sovereign wealth fund, while MBZUAI is the Mohammed bin Zayed University of Artificial Intelligence in the UAE. MBZUAI alone accounts for 62% of the company.

This means that Cerebras is not currently expanding its market reach, but rather is being fed by a small number of super customers.

Why this revenue structure? Because Cerebras doesn't sell standard chips. It sells complete AI supercomputing systems, including: WSE chips, servers, networks, software, data center deployments, and AI inference facilities.

It is naturally suited for mega-projects such as national AI projects, supercomputing centers, and large-scale modeling companies. Coincidentally, the UAE wants to become a sovereign AI nation with independent AI infrastructure and its own AI cluster.

Cerebras' revenue growth wasn't the most shocking thing. What truly shocked the capital markets was Cerebras' $24.6 billion order backlog. A company with $510 million in revenue was claiming that $24.6 billion in revenue awaited it.

This means the AI industry is shifting from selling chips to pre-ordering token throughput capacity.

Who placed these orders?

The bulk of the funding came from OpenAI, and OpenAI CEO Altman himself was an early investor in Cerebras.

OpenAI and Cerebras signed a 750MW AI computing power agreement, conservatively estimated at over $10 billion, with the actual commitment potentially exceeding $20 billion. The agreement runs until 2028.

OpenAI may also provide Cerebras with approximately $1 billion in funding to support data center construction. It may also acquire up to approximately 10% of Cerebras' shares in the future. The relationship between the two companies has evolved beyond simply customer-supplier; it's more like a joint effort to build AI infrastructure.

Previously, many American media outlets, while appreciating the coolness of Cerebras' technology, would always add the comment: "It lacks truly large clients." But after OpenAI signed the contract, everything changed.

Another customer is G42, mentioned above. Of the $24.6 billion order, G42 likely still accounts for several billion dollars. However, the public documents do not specify the exact amount.

Challenging Nvidia? Hold on.

Can Cerebras truly challenge Nvidia? This is a question that investors and the AI industry are very concerned about.

To put it simply: it is still a long way from truly challenging Nvidia.

The real issue isn't whether Cerebras' chips are powerful enough, but rather that Nvidia is no longer a chip company.

Nvidia's biggest competitive advantage isn't just its GPUs, but its CUDA software ecosystem. Today, the vast majority of AI frameworks, training systems, inference tools, and engineering libraries worldwide are built around CUDA. Many companies don't just buy Nvidia products; they can't function without CUDA. Cerebras' software ecosystem, on the other hand, is currently far less mature than CUDA's.

More importantly, there's the customer structure.

Cerebras' revenue is currently highly dependent on a few large clients, which means it is more like a super-project company than a platform company that has built a broad ecosystem.

Furthermore, there's a real issue in the industry: there's no single right path for AI hardware. A study on AI accelerators published this year by Harvard and other institutions points out that the best hardware platform varies depending on the AI workload. Cerebras, Groq, TPU, Gaudi, and GPUs each have their suitable scenarios. In other words, Cerebras may significantly outperform GPUs in certain inference scenarios, but this doesn't mean it can completely replace Nvidia.

Nvidia has established extremely strong control over its supply chain. What the AI industry truly lacks now isn't just GPUs, but also CoWoS advanced packaging, HBM memory, power, data centers, and network equipment. Nvidia is almost entirely tied to TSMC's advanced packaging capacity, Micron's memory, and SK Hynix's HBM. This means that even if the Cerebras architecture is established, it will still need to rely on the same supply chain system.

Furthermore, Cerebras' business model itself makes it difficult for it to scale like Nvidia. Nvidia sells standardized products. But Cerebras is more like a highly customized project, which is inherently more resource-intensive, slower, and more dependent on large clients.

This article does not constitute any investment advice.