Think about the size of a single semiconductor chip.
Usually, it's about the size of a fingernail. Billions of circuits are packed into a square smaller than a postage stamp.
But one company asked:
"Why do we have to cut it small?"
And then they used a whole round silicon disk (called a wafer) as a single chip without cutting it.
This company is Cerebras Systems.
This idea of making a chip the size of a dinner tray seemed reckless at first. But now, OpenAI and Amazon are lining up to use their chips, and the company has surpassed a market capitalization of 50 trillion won on the Nasdaq. How did recklessness turn into 50 trillion won? And is that value justified?
This article follows those two questions.
Why Use the Whole Chip Now?
First, let's look at the background.
The chatbots and AI assistants we use today generate answers by having a massive "model" perform calculations. This process is called inference. When you ask a question, the AI spits out words one by one, and every single word is the result of a calculation.
The problem is that these calculations are slower than you'd think.
Until now, this has been handled by grouping multiple GPU chips made by Nvidia. But having multiple chips means data must constantly move between them. Imagine storing parts in a large warehouse (memory) on the outskirts of a city and having a factory (calculation unit) transport them by truck whenever needed. No matter how fast the factory is, if the trucks are stuck in traffic, the overall speed slows down.
This "truck congestion" is the biggest bottleneck in AI inference.
Cerebras's answer is simple yet bold.
It's putting the warehouse inside the factory, right on the workbench, instead of next to it. By using the entire wafer as one chip without cutting it, the calculation unit and memory are stuck together on the same plane. The roads for trucks are no longer needed. The parts are within arm's reach.
The proof that this idea is starting to work in the market is the revenue.
Cerebras's annual revenue jumped from about 32 billion won in 2022 to about 660 billion won in 2025. It has grown more than twenty-fold in four years.
It's a signal that the company is moving beyond a small laboratory level and becoming a real business.
However, just because revenue grew doesn't mean they are making a profit. We'll touch on this point again later.
Numbers Contained in a Single Wafer
Cerebras's chip is named WSE-3.
The name is dry, but the numbers inside are overwhelming.
It contains 4 trillion transistors (tiny switches that turn electrical signals on and off) and 900,000 AI cores responsible for calculations. The high-speed memory attached directly to the chip is 44 gigabytes, and the speed of reading and writing that memory reaches 21 petabytes per second.
A petabyte is about a million times a gigabyte. If the numbers are too large to grasp, just understand it as "a scale unimaginable for ordinary chips."
Here, a doubt arises for everyone.
"Won't such a large chip have defects?" That's a valid concern. In semiconductor manufacturing, a single speck of dust or a minute flaw can ruin a section. The larger the chip, the higher the probability of encountering a defect.
This is where Cerebras's solution is clever.
Instead of trying to "eliminate defects completely," they chose to "minimize the loss even if defects occur." WSE-3 has its cores divided very finely. So, even if a defect occurs in one core, they just turn off that one tiny core and use the rest normally. In fact, when shipping, they only activate 900,000 out of the 970,000 cores.
They designed it with plenty of redundancy from the start. It's the same principle as not discarding an entire building just because one light bulb goes out; you just replace that bulb.
What This Company Really Sells is 'Speed'
The most accurate sentence to understand Cerebras is this:
This company doesn't sell the "most powerful chip," but the "chip that reduces waiting time."
Why is this important?
Anyone who has asked an AI assistant to write code or summarize a long text knows. Those few seconds waiting for the answer to come out feel quite long. Especially in the so-called 'agent' method, where the AI processes tasks through multiple steps on its own, you have to wait dozens of times, not just once. In this case, speed is user experience, and user experience is money.
Actual numbers support this.
Mistral, a French AI company, stated that when they ran their chatbot on Cerebras, it poured out over 1,100 words per second. Cognition, which makes coding tools, said it was about 5 times faster than using GPUs. Cerebras claims inference up to 21 times faster than top-tier GPU systems. Of course, these numbers vary depending on the model and task, so you don't have to believe them literally. However, the big picture that "Cerebras is clearly fast in areas where waiting equals loss" remains unshaken.
Looking at who this strength has attracted makes the story even clearer.
- OpenAI announced it would gradually introduce 750 megawatts of low-latency computing capacity by 2028. A megawatt is a unit of power, and 750 megawatts is a massive scale equivalent to the electricity used by a typical city.
- Amazon said it would create a structure where its AI service (Bedrock) handles the initial calculations with its own chips and delegates the fast answer generation at the end to Cerebras.
The fact that the world's most advanced AI companies chose Cerebras for the single task of "fast answers" is the company's most powerful business card.
But There Is Almost Only One Customer
Up to this point, it's a flawless growth story.
But good investment judgment comes from looking into the shadows when the light is strong.
Cerebras's biggest shadow is that it has 'too few customers.'
86 percent of 2025 revenue came from G42 and one entity close to it. G42 is a giant AI company in the United Arab Emirates. Simply put, nine out of ten dollars of the store's sales are filled by a single regular customer. If that regular's mood sours, their situation worsens, or trade is blocked due to political issues between the two countries, the store will falter overnight. In fact, Cerebras had to postpone its listing schedule once due to US government security review issues related to G42.
Of course, the aforementioned OpenAI and Amazon are the hopes to clear this shadow.
If their contracts turn into actual revenue, the dependence on one regular customer will decrease rapidly. However, there is always a time lag between an announcement and actual revenue. Until promises turn into invoices, it's honest to say that Cerebras's 'quality of revenue' has not yet been fully proven.
Being Fast Doesn't Mean Winning
Another shadow to note is competition.
Just because Cerebras leads in speed doesn't mean it takes the entire market.
Nvidia still has the broadest ecosystem. An ecosystem refers to the thickness of developers working on that chip, the tools used together, and the accumulated know-how. It's like a highway that is well-paved and has repair shops everywhere. Cerebras's road is fast, but it's still narrow, and the destinations it can reach are limited. Furthermore, giants like Amazon, Google, and Microsoft have started making their own AI chips. Google's inference-only chips and Amazon's proprietary chips are examples.
So, Cerebras's real opponent is no longer just Nvidia, but the entire range of chips being nurtured by giant corporations.
Cerebras also has a weakness in that it lags behind GPUs in memory capacity or the ability to hold very large models at once.
So, the realistic picture is this:
Rather than becoming a 'champion who takes over the entire market,' Cerebras is better suited to being a 'specialist sold at a high price in premium areas where one second of waiting equals money.'
Future Outlook: The Price Has Gone Too Far Ahead
Now, back to the second question from the beginning.
Is the value of 50 trillion won justified?
There is a simple yardstick to gauge whether a company is expensive or cheap in investment.
It's dividing the market capitalization by the annual revenue (P/S ratio). The larger this number, the more "expensive it is evaluated compared to revenue."
For Cerebras, this value is about 104 times. This is astronomically high compared to the average of 3.6 times for the top 500 US companies and 19.3 times for the semiconductor equipment industry. It means the market has already priced Cerebras not as a 'small company just starting out,' but as a 'winner that will almost monopolize the next era of AI infrastructure.'
The excitement on the first day of listing was great.
Buy orders exceeded sell orders by more than twenty times, and the stock price once soared to $386. However, investors who felt the burden later exited to take profits, and it has now come down to around $240.
So, what is a reasonable value?
We can draw the future in three parts.
- If everything goes well and large contracts quickly turn into revenue, the stock price could reach $320 (Optimistic).
- Conversely, if contracts keep getting delayed and the technical edge narrows, it could fall to $120 (Pessimistic).
- And right in the middle, $205 is seen as the most realistic value (Neutral).
The $240 currently being traded already exceeds this neutral value.
Being a good company is different from being a good stock.
Cerebras is clearly the former, but to be the latter, the price is a step ahead of expectations.
Conclusion: Questions Left by Recklessness
Cerebras's story began with the quirky question, 'Why do we have to cut chips small?' That question became a chip the size of a dinner tray, became the choice of the world's top AI companies, and finally returned as a value of 50 trillion won. The technology is undoubtedly top-tier.
However, the question remaining for investors is not the excellence of the technology, but the next step.
Cerebras has already proven 'how fast it is.'
Now, what it needs to prove is 'how consistently it can turn that speed into money.'
Expanding customers beyond one regular and turning announced promises into actual invoices. When it passes through that ordinary and arduous process, the current value will finally be justified.
Therefore, the appropriate stance now is 'neutral.'
Rather than chasing and buying at a high price, it's better to approach in stages when the market cools down and the price drops. Asking the price even in front of good technologyโthat is ultimately the attitude of an investor who survives for a long time.
