Cerebras: How Wafer-Scale AI Chips Challenge NVIDIA? Analysis of CBRS Financial Reports and Technical Barriers

On May 14, 2026, Cerebras Systems—an AI chip company still little known to the general public, yet whose technological roadmap is capable of overturning the entire AI chip industry landscape—completed the largest technology IPO of 2026 so far on the Nasdaq. Priced at $185 per share, it opened with a gap up to $350, then rose 68% on the first day. Dubbed “NVIDIA’s strongest challenger,” this AI chip newcomer launched its most direct technical challenge to GPU leader NVIDIA with a “plate-sized” wafer-level chip.

However, just a little over a month after listing, Cerebras’s first quarterly report triggered sharp disagreement in the market—revenue beat expectations and losses narrowed significantly, but guidance for gross margin dropped sharply, sending the stock down more than 10% at one point after hours. What exactly is the market worried about? Does the wafer-level chip’s independent route truly have the long-term logic to challenge NVIDIA? We will break down Cerebras systematically across four dimensions: its technological roadmap, financial performance, the industry competitive landscape, and trading channels.

The Value of Cerebras

The traditional logic of chip manufacturing is as follows: on a 12-inch silicon wafer, lithography is used to pattern hundreds of chips, which are then cut, packaged, and tested. The area of each chip is constrained by the lithography machine’s mask size, so it cannot be made large. Cerebras’s approach overturns this paradigm—instead of cutting and splitting the wafer, it completes full-area lithography wiring at once, turning an entire silicon wafer into a single super-large chip.

This is Cerebras’s Wafer-Scale Engine (WSE) wafer-scale engine. The latest WSE-3 is built on TSMC’s 5nm process, with a single-die area of 46,225mm², integrating 4 trillion transistors and 900,000 AI cores, and equipped with 44GB of on-chip SRAM, delivering 125 petaflops of AI compute power. For comparison, NVIDIA’s main AI data center chip, H100, contains about 80 billion transistors—meaning the transistor count of WSE-3 is 50 times that of the H100.

But the huge disparity in transistor counts is not Cerebras’s real moat. The bigger differentiation lies in memory architecture.

Conventional GPUs (such as H100) rely heavily on external HBM high-bandwidth memory, and the data transfer between the chip and memory is limited by physical bandwidth. In the industry, this is known as the “memory wall”—no matter how strong the compute units are, performance is futile if data cannot be moved. Cerebras integrates 44GB of SRAM directly on the chip, and its on-chip memory bandwidth reaches 21 PB/s. Some analysts point out that the memory bandwidth of WSE-3 is 2,625 times that of NVIDIA’s B200 packaged chip. In AI inference scenarios, this means model weights do not need to be frequently transferred from off-chip, significantly reducing inference latency.

Of course, a wafer-scale solution also comes with trade-offs. If a wafer-wide manufacturing process has even one fatal defect, it could render the entire chip unusable. Cerebras’s response is “redundant core repair”—by designing large numbers of spare compute cores, it can automatically bypass defective regions. This undeniably increases design complexity and cost. Using an entire wafer as a single chip means defect tolerance and yield management are fundamentally different from traditional chip manufacturing.

The essence of the technical-route difference: NVIDIA follows a “large-scale cluster + high-speed interconnect” route, using countless GPUs to build supercomputers; Cerebras follows a “maximally amplified single-chip” route, using one giant chip to replace hundreds or thousands of GPUs. The former has no equal in software compatibility due to decades of ecosystem accumulation; the latter may have theoretical efficiency advantages in specific inference scenarios, but its software ecosystem still needs to be built from scratch.

The First Month After Listing: From Frenzy to Pullback, CBRS’s Price Trajectory

On May 14, 2026, Cerebras listed on Nasdaq at an IPO price of $185 per share. It opened with a gap up to $350, surged more than 108% intraday, and triggered a circuit breaker. It ultimately closed at $311.07. Up 68% on the first day, it became the largest U.S. tech company IPO of 2026 to date.

In the following weeks, CBRS’s stock price saw extreme volatility. It once reached an all-time high of $386 and also fell to an all-time low around $197. As of the regular session close on Tuesday, June 23, CBRS was $226.72—still about 23% above the IPO price, but down more than 27% from the first-day closing price.

On the night of June 23, Cerebras released its first post-IPO quarterly financial report, which was followed immediately by a sharp after-hours plunge of more than 10% in the share price. In the overnight trading session on June 24, CBRS fell further by nearly 11%, to $201.8.

As of the time of writing, CBRS’s market capitalization is about $49.8 billion, and its trailing P/E ratio is about 527x. This valuation reflects the market’s expectations for rapid growth, but it also means that any signal that falls short of expectations could trigger sharp volatility.

The First Earnings Report: Revenue Beats Expectations—Why Doesn’t the Market Buy It?

Cerebras’s financial results for Q1 2026 (ended March 31) show a clear “two-sided” picture:

The “better-than-expected” side:

• Total revenue of $193.4 million, up 94% year over year, exceeding analysts’ expectation of $181.2 million
• Hardware revenue of $110.6 million, up 59%
• Cloud and other services revenue of $82.8 million, up 178%
• Net loss of $14.0 million, sharply narrower than the $23.9 million in the same period last year
• Loss per share of $0.22, better than analysts’ expected $0.25
• Full-year revenue guidance of $855 million to $865 million, above analysts’ expected $824.8 million

The “worrying” side:

• Q2 core gross margin guidance of 36%-38%, down sharply from 46.5% in Q1 by more than 10 percentage points
• Full-year core operating profit margin expected at -28% to -32%
• The company expects Q2 core revenue of approximately $194 million

Doubling revenue, narrowing losses, and raising guidance—this is a standout performance for any growth company. Yet the market’s reaction was a plunge after hours. The logic is not complicated: the market’s valuation of Cerebras is built on a double expectation of “high growth + high gross margin,” and the sharp drop in gross margin directly undermines the latter half of that foundation.

CFO Bob Komin explained the reason for the gross margin decline on the earnings call: a lack of data center space is forcing Cerebras to rent back some systems from customers, and the company is also “actively” expanding its own capacity. These costs are expected to reduce 2026 profit margins by roughly 10 to 15 percentage points. CEO Andrew Feldman was even more blunt: “Extremely ironic: after we—and NVIDIA—invented all these technologies, building the buildings themselves turns out to be the limiting factor.”

In other words, Cerebras’s current core bottleneck is not technology or demand, but the speed at which physical infrastructure can be supplied failing to keep up with the rate of order growth. This certainty places pressure on short-term profits, but from another angle, it also confirms the authenticity and urgency of demand.

OpenAI and AWS: A Structural Change in Customer Composition Behind $20 Billion Orders

Cerebras’s evolving customer structure is a key dimension for understanding its long-term value.

In the first half of 2024, the UAE AI company G42 accounted for 87% of Cerebras’s revenue. Such extreme customer concentration was one of the market’s biggest concerns. However, in January 2026, Cerebras announced a strategic cooperation agreement with OpenAI worth more than $20 billion—OpenAI will deploy 750 megawatts of Cerebras high-speed inference compute capacity by 2028. The two sides also jointly launched Codex-Spark, an AI model designed for near-instant coding, capable of generating more than 1,000 tokens per second.

At the same time, Cerebras has established a multi-year strategic partnership with Amazon AWS, planning to deploy the Cerebras CS-3 systems to AWS data centers. The two sides will introduce a “decoupled inference strategy”: AWS’s Trainium 3 chips handle the prefill computation, while Cerebras CS-3 is responsible for rapid inference decoding.

The significance of these collaborations goes far beyond the order amounts themselves. From a single customer, G42, to a dual-pillar setup with OpenAI and AWS, Cerebras’s customer concentration risk has been substantially optimized. More importantly, OpenAI and AWS represent two of the most core global AI inference demand scenarios—frontier model training and large-scale cloud service deployments. Being able to secure long-term orders from both of these giants is, in itself, a form of “market validation” of Cerebras’s technological route.

By the end of 2025, Cerebras had $24.6 billion in backlog of orders not yet delivered, and the company expects to convert $3.7 billion of that into recognized revenue by 2027. The ratio of unfulfilled orders to current revenue is about 48x—this figure reflects both the visibility of future revenue and that Cerebras is still in the early stage of large-scale delivery.

The Independent Wafer-Level Chip Route: The Backing and Limitations of Challenging NVIDIA’s Monopoly

Cerebras has chosen a technological path that is completely different from NVIDIA’s.

NVIDIA represents the industry mainstream route: the chiplet approach. Chips are split into separate chiplets for compute, cache, IO, and other functions, produced separately, and then assembled through advanced packaging. This route offers higher yields, controllable costs, and scalable mass production. NVIDIA’s B200 and Huawei’s Ascend both adopt this approach.

Cerebras’s wafer-level route is “casting the entire wafer”—without cutting or stitching, an entire wafer is used directly as a single chip. This route has theoretical efficiency advantages in inference scenarios, but it also faces challenges such as high manufacturing complexity, difficult yield management, and the need to build a software ecosystem from scratch.

The competition between the two is fundamentally a contest between two paradigms: “scaling effects” and “extreme efficiency.” NVIDIA’s advantage lies in decades of accumulated CUDA software ecosystem and large-scale mass production capability; Cerebras’s advantage lies in achieving speed advantages of more than 10x in specific inference scenarios.

For investors, the key question is not whether “Cerebras can beat NVIDIA”—in the foreseeable future, that is almost impossible. The real question is: is the AI inference market large enough to support a technology route independent of GPUs? If the answer is yes, then Cerebras, as the only commercial player on this route, has a valuation logic rooted in its scarcity.

Risk Factors: Four Major Challenges Not to Be Ignored

Uncertainty in gross margin and the path to profitability. The Q2 gross margin guidance dropped from 46.5% to 36%-38%. The full-year core operating margin remains deeply negative. The company still has a considerable distance to go to achieve sustainable profitability. Morgan Stanley believes the margin compression is temporary and expects gross margins to return to the target level of 60% as Cerebras gradually phases out leasing infrastructure, but this judgment has not yet been validated by the market.

Structural changes in customer concentration still need time to be tested. Although the inclusion of OpenAI and AWS has greatly improved the customer mix, OpenAI’s $20 billion order still dominates the overall backlog (out of $24.6 billion backlog, OpenAI accounts for $20 billion). Any changes in deployment timing from OpenAI could have a significant impact on revenue.

Supply pressure caused by the expiration of lock-up periods. This Thursday (June 25), the lock-up period will expire, and about 13% of IPO shares will become available for sale by early supporters and insiders. An increase in float may create short-term downward pressure on the stock price.

Matching between valuation and growth rate. CBRS’s current price-to-sales ratio is about 91x, far higher than NVIDIA’s roughly 23x. While valuation premiums are reasonable for high-growth companies, once growth slows or gross margin improvements fall short of expectations, the risk of valuation compression cannot be ignored.

Conclusion

Cerebras’s rise is, at its core, a snapshot of the shift in AI compute demand from “training” to “inference.” As large-model training becomes more standardized and scaled, the extreme pursuit of minimal latency, cost, and energy efficiency in inference is opening a commercialization window for “non-mainstream” technology routes such as wafer-level chips.

From its first earnings report, Cerebras delivered a strong performance with both revenue and order figures beating expectations. However, the sharp drop in gross margin also reveals the growing pains of an early-stage company: the expansion speed of physical data center infrastructure cannot keep up with the pace of explosive compute demand. This is a “sweet problem,” but it is also a tangible erosion of profits.

For investors, Cerebras is not a “replacement for NVIDIA,” but rather “another possibility for AI inference.” The ultimate success or failure of this independent route depends on the evolution of two core variables: first, whether the AI inference market can continue to expand at high speed enough to support multiple technology routes coexisting; second, whether Cerebras can efficiently turn its $24.6 billion backlog orders into a virtuous cycle of revenue and cash flow in 2026-2027.

Can Cerebras’s wafer-level chips truly shake NVIDIA’s GPU empire? The answer may not be found today, but rather in the next 12-24 months at each key checkpoint of order conversion rate, gross margin recovery, and AWS partnership implementation.

FAQ

Q1: What are the core differences between Cerebras’s wafer-level WSE-3 and NVIDIA’s H100?

WSE-3 is a full, uncut 12-inch wafer that integrates 4 trillion transistors and 900k cores; H100 is a traditional cut-and-packaged chip. The core difference is in memory architecture: WSE-3 has 44GB of on-chip SRAM with bandwidth of 21 PB/s; H100 relies on external HBM with bandwidth of only 3.35 TB/s. WSE-3 has significant speed advantages in inference scenarios, but it also has higher manufacturing complexity and cost.

Q2: What are the core data points from Cerebras’s 2026 Q1 earnings report?

Q1 revenue was $193.4 million, up 94% year over year, exceeding the $181.2 million expectation; net loss was $14 million, down sharply from the $23.9 million in the prior-year period; hardware revenue was $110.6 million, and cloud services revenue was $82.8 million. Full-year revenue guidance is $855 million to $865 million.

Q3: Why did Cerebras’s stock price fall sharply after releasing its earnings report?

Although both revenue and losses beat expectations, the Q2 gross margin guidance dropped from 46.5% in Q1 to 36%-38%. The main reason is that there is insufficient data center space: the company is forced to rent systems back from customers and actively expand production, and the related costs will lower profit margins by about 10-15 percentage points. The market worries about the visibility of the profitability path.

Q4: What are the main risks Cerebras faces?

Four major risks: gross margin compression and uncertainty in the profitability path; customer concentration risk—OpenAI still accounts for a large share of the backlog (OpenAI’s $20 billion order is most of the $24.6 billion backlog); the lock-up period expires on June 25, and about 13% of IPO shares can be sold; the price-to-sales ratio is about 91x, far higher than NVIDIA’s 23x, making valuation compression risk significant.