Who would have thought that Cerebras—once the darling of investors and widely touted as a challenger to Nvidia—would watch its stock price plummet by nearly 80% in just half a year after going public?

On May 14, 2026, Cerebras' shares surged to as high as $386.34 on its first day of trading. By the close on November 6, however, the stock was trading at just $89.68—a staggering 76.8% drop from its peak.

What makes this even more puzzling is that Cerebras' financial performance has hardly been bad.

From the fourth quarter of 2025 through the second quarter of this year, the company's year-over-year revenue growth rates reached 110.3%, 94.4%, and 74%, respectively. Though growth has moderated somewhat, it remains robust by any reasonable standard.

Behind that growth, however, Cerebras' business model is undergoing a profound transformation.

In the second quarter of last year, cloud and other services accounted for just 32% of Cerebras' total revenue. By the second quarter of this year, that share had climbed to 70%, making it the company's primary revenue source. Meanwhile, hardware revenue declined quarter over quarter for two consecutive quarters.

In other words, Cerebras—a company that once sought to challenge Nvidia by selling chips—now increasingly resembles an AI cloud service provider.

This raises a question worth pondering: why is a chip company with globally leading inference performance becoming ever more dependent on cloud services to make money? The answer may lie precisely in the technological advantage Cerebras is most proud of.

Cerebras' near-extreme token generation speed has consistently failed to translate into commensurate commercial value. And while pivoting to cloud services lowers the barrier for customers, it also forces Cerebras to shoulder heavier capital expenditure.

When technological advantages prove difficult to convert into pricing power, and revenue growth depends on massive capital investment, investors naturally begin to reassess Cerebras' commercial value.

/ 01 / From Selling Chips to Selling Cloud Services

Rather than calling Cerebras a chip company, it would be more accurate to describe it as an AI cloud company.

In the second quarter of this year, Cerebras' cloud and other services revenue reached $126 million, up 281% year over year and accounting for as much as 70% of total revenue.

Notably, that proportion stood at just 32% in the same period last year. In the span of a single year, cloud services have become Cerebras' primary revenue driver.

By contrast, hardware sales—once the core business—have been declining.

From the fourth quarter of 2025 through the second quarter of 2026, Cerebras' core hardware revenue fell from $122 million to $112 million and then to $82.14 million, declining quarter over quarter for two consecutive quarters.

Why has this shift occurred? One important reason is that, compared with selling hardware, cloud services make it far easier to open up the market.

On the second-quarter earnings call, Cerebras Chief Financial Officer Bob Komin revealed that procurement and delivery for some hardware customers are being affected by the pace of data center construction.

Put simply, customers purchasing Cerebras' chips must still resolve a whole series of infrastructure issues—server rooms, power, cooling, and so on. Not only is the upfront investment substantial, but the construction cycle is also lengthy.

For a large number of AI application companies, however, what they want more than an expensive supercomputer is high-speed inference capability available on demand. Cerebras spotted this business opportunity as well.

In August this year, CEO Andrew Feldman revealed in an interview that the company initially built its cloud platform simply to let potential customers experience chip performance firsthand. Unexpectedly, as demand continued to grow, this service—originally intended as a product showcase—gradually evolved into a real business.

Now, Cerebras has even begun proactively shifting hardware capacity toward its cloud business.

On the first-quarter earnings call this year, management explicitly stated that hardware revenue is expected to decline in the coming quarters because most newly produced equipment will be deployed first to its own cloud platform to fulfill contracts with large customers.

In terms of customer expansion, this model has already begun to bear fruit.

In the second quarter, Cerebras signed new cloud capacity agreements with AI programming companies Cognition and Lovable, while financial information platform AlphaSense and pharmaceutical giant GSK also became customers.

This means Cerebras is moving away from its past model of relying on a small number of large hardware orders and instead covering a broader AI application market through cloud services.

But here lies the problem: although cloud services have opened up a larger market, they have also turned Cerebras into a heavier business.

In the past, when selling hardware, equipment procurement and data center construction were mainly borne by customers. Now, in operating a cloud platform, Cerebras must invest substantial funds upfront to purchase computing equipment and build computing infrastructure, then gradually recoup the investment through future service revenue.

This shift has already been reflected in the financial statements.

In the first half of this year, Cerebras' capital expenditure reached $549 million, nearly tripling year over year, most of which was used to purchase computing equipment required for cloud services.

As investment continued to expand, the company's free cash flow net outflow in the first half of the year reached $596 million, nearly double that of the same period last year. In the second quarter alone, free cash flow net outflow was $477 million—equivalent to 2.65 times revenue in the same period.

In other words, for every $1 of revenue Cerebras generates, roughly $2.65 of free cash flow flows out during the same period.

At the same time, the expansion of the cloud business has begun to squeeze profit margins.

In the second quarter, Cerebras' overall core gross margin fell from 46.5% in the previous quarter to 40.6%. To meet customer demand, the company even had to lease equipment back from customers, and the additional costs thereby incurred dragged down gross margin by about 5 percentage points.

This also exposes another side of Cerebras' cloud business expansion: revenue is growing quickly, but supporting that growth requires continued enormous capital investment, and profitability is under pressure.

Of course, heavy capital investment is not frightening in itself; the key lies in whether sufficient returns can be earned. But this is precisely the biggest problem Cerebras faces: leading chip performance cannot be converted into a comparable degree of pricing power.

/ 02 / Technological Leadership Without Pricing Power

Cerebras' greatest commercial dilemma is that its enormous advantage in chip performance is exceedingly difficult to convert into commensurate commercial value.

To put it bluntly, even if Cerebras' chips are more than ten times faster than Nvidia's, customers may not be willing to pay a higher price for that speed.

Cerebras' core advantage stems from its wafer-scale chip architecture. By integrating a large number of computing cores and high-speed memory on a single wafer, it greatly reduces data movement and thereby achieves extremely high token generation speed.

According to test data from April 2026 disclosed in the prospectus, when running the Llama 3.3 70B model under specific conditions, Cerebras can generate 2,457 tokens per second, while a comparison cloud service using Nvidia GPUs generates only 164—meaning the former is nearly 15 times the latter.

The problem is that in AI inference scenarios, fast token generation alone is far from sufficient.

For scenarios such as batch data processing and offline content generation, enterprises care more about overall throughput and unit cost than about raw generation speed.

Even in today's hottest AI programming, faster token generation does not necessarily mean tasks are completed faster.

A programming agent, in addition to generating code, must also read files, call tools, run tests, and repeatedly revise based on results. These steps also consume a great deal of time.

This means that even if Cerebras increases token generation speed by more than tenfold, overall task completion efficiency may not improve by the same magnitude.

This brings up an awkward problem: although Cerebras leads technologically by more than ten times, the actual value it can create for customers may be far less than that.

This can also be seen from Cerebras' public pricing.

Taking the GPT-OSS-120B model as an example, Cerebras' published output price is $0.75 per million tokens, with a nominal generation speed of about 3,000 tokens per second.

By comparison, GroqCloud, which also focuses on high-speed inference, charges $0.60 per million tokens for output, with a nominal speed of about 500 tokens per second.

Although the two platforms were not tested under strictly identical conditions, Cerebras' nominal speed is several times higher, yet its output token unit price is only 25% higher.

This at least shows that an extreme speed advantage cannot be exchanged for a comparable price premium.

In addition to pricing power, Cerebras faces another challenge: storage capacity.

To pursue extreme speed, Cerebras uses a large amount of on-chip SRAM. This type of memory is extremely fast but occupies a relatively large chip area. Taking the third-generation wafer-scale chip WSE-3 as an example, compared with the previous generation, its transistor count increased by more than 50%, but on-chip storage capacity only rose from 40GB to 44GB.

At the same time, the parameter scale and context length of large models continue to grow.

When a single wafer cannot accommodate the entire model, multiple wafers must work together. Although Cerebras can solve this problem through model layering and pipeline technology, it also means more equipment must be invested in, along with more complex system deployment and scheduling costs.

In other words, although Cerebras has achieved a huge breakthrough in single-chip performance, turning that advantage into a cost advantage for an entire computing system still poses no small challenge.

Beyond the challenges facing its own technological path, Cerebras must also contend with a more formidable opponent: Nvidia.

In January this year, Nvidia disclosed that through software optimization, when running the DeepSeek-R1 model on the same Blackwell GPU platform, single-card throughput increased by up to 2.8 times within three months.

This means that even without replacing hardware, Nvidia can continuously improve GPU inference efficiency and narrow the performance gap with Cerebras.

More importantly, compared with Cerebras' specialized architecture, GPUs have another obvious advantage: generality.

For enterprises that have already invested heavily in purchasing Nvidia GPUs, the same equipment can be used to train models, perform inference, and adapt to different model architectures.

By contrast, Cerebras is mainly optimized for specific computing scenarios, so customers not only need to consider additional deployment and adaptation costs, but also bear risks such as future model changes and insufficient equipment utilization.

This puts Cerebras in an awkward position: it needs to rely on extreme performance to prove its value, but as long as Nvidia continues to improve GPU inference efficiency, it may continuously compress the premium customers are willing to pay for speed.

Shifting to cloud services now lowers the barrier for customers, but it has not fundamentally changed this commercial calculation.

Admittedly, Cerebras has built one of the world's fastest AI inference chips, but it still has a long way to go before proving that this is a sufficiently profitable business.

/ 03 / Conclusion

It used to be said that one trick can win the world.

But in today's AI computing market, that logic seems increasingly difficult to sustain.

The reason is simple: AI models need both training and inference; some applications pursue extreme speed, while others value throughput and cost more. Demand is becoming increasingly diverse, and no matter how outstanding a single performance metric is, it is hard to dominate the entire market.

In fact, specialized AI chips have long experienced similar difficulties.

The UK's Graphcore once challenged Nvidia with its unique chip architecture, and its valuation once reached $2.77 billion. But by 2022, the company's annual revenue was only $2.7 million, and it was ultimately acquired by SoftBank.

Another AI chip company, Wave Computing, also won capital favor with its specialized dataflow architecture, but filed for bankruptcy reorganization in 2020.

The experiences of these companies show that technological leadership does not guarantee commercial success. The real challenge is how to convert performance advantages into sufficiently broad customer demand and make money from it.

Now, Cerebras stands at the same crossroads.

It has proven that it has the ability to build one of the world's fastest AI chips, but whether it can convert this technological advantage into sustained revenue growth and stable profits remains unknown.

And this question may be even harder than building the world's fastest AI chip.

By Yuanyuan