Cerebras: Can a New AI Architecture Challenge NVIDIA?
An AI IPO overshadowed by SpaceX may be one of the more interesting stories in the market.
Artificial intelligence has created one of the largest infrastructure booms in modern technology history. NVIDIA has become one of the most valuable companies in the world as demand for AI chips has exploded. But underneath the surface of the AI buildout, a growing group of companies is trying to build alternative AI hardware architectures that could compete in certain parts of the market.
One of the most interesting companies in this emerging field is Cerebras, which is not trying to build a slightly better GPU, but is instead attempting something much more radical: redesigning how AI chips themselves are built.
The company’s technology is so different from traditional AI hardware that understanding Cerebras requires understanding one of the biggest bottlenecks in artificial intelligence itself: moving data.
Modern AI systems perform enormous numbers of calculations simultaneously. Training and running large AI models requires constant movement of information between processors, memory, storage, and networking systems. The problem is that moving data often becomes slower than performing the calculations themselves.
This is one reason AI hardware has become so important. The industry is not simply trying to build faster processors. It is trying to build systems that can move enormous amounts of information with as little delay as possible.
Traditional GPUs, including those made by NVIDIA, are extremely good at this. But they still face a major limitation: chip size.
Modern chip manufacturing tools can only expose a certain amount of silicon at one time. This is called the “reticle limit.” As a result, companies usually manufacture many smaller chips and then connect them together into larger systems. That approach works extremely well, but it creates a tradeoff: communication between separate chips is slower than communication inside a single chip.
Cerebras Took A Different Approach
Instead of building many smaller chips, Cerebras developed a way to turn an entire silicon wafer into one giant processor. The company calls this a Wafer-Scale Engine, or WSE. Normally, a silicon wafer contains many separate chips that are cut apart after manufacturing. Cerebras instead leaves the entire wafer intact and connects everything together into one enormous processor.
The result is one of the largest chips ever built.
The advantage is speed. Because the entire wafer acts as a single system, information can move across the processor much faster than it can between many separate chips connected together externally. This is particularly important in AI because large models constantly move information between memory and processors.
Cerebras’ latest system, the WSE-3, contains roughly 900,000 AI compute cores, 44 GB of on-chip SRAM memory, and approximately 21 petabytes per second of memory bandwidth. By comparison, NVIDIA’s H100 GPU contains more total memory — about 80 GB of HBM memory — but far lower memory bandwidth at roughly 3.35 terabytes per second.
In simple terms, NVIDIA’s H100 can store more information overall, but Cerebras can move information between memory and processors dramatically faster. Cerebras’ memory bandwidth is measured in petabytes per second, while NVIDIA’s is measured in terabytes per second — a difference of roughly 6,000x. That difference matters because modern AI systems are often bottlenecked not by raw computing power alone, but by how quickly they can move data around inside the system itself.
That matters most in a part of AI called “inference.” Inference means actually using a trained AI model after training is complete. Examples include ChatGPT responding to users, AI coding assistants, voice assistants, AI search, and reasoning systems.
Modern AI systems generate responses one token at a time. The faster an AI system generates tokens, the faster it feels to humans. That matters enormously for real-time voice systems, AI wearables, reasoning models, interactive AI assistants, and coding copilots.
Cerebras is effectively betting that ultra-fast inference will become one of the most valuable parts of the AI economy.
At this point, many investors may ask: if Cerebras’ architecture is so fast, can the company compete with or supplement NVIDIA’s architecture?
The answer is complicated. In some specialized inference workloads, Cerebras may offer meaningful advantages. But NVIDIA’s dominance is not just about the chip itself. NVIDIA has spent years building software ecosystems, networking systems, developer tools, AI frameworks, and large-scale distributed computing infrastructure. Modern AI training often requires tens of thousands of GPUs working together simultaneously across enormous data centers, and NVIDIA’s ecosystem is optimized for exactly that type of distributed computing.
Cerebras’ architecture works extremely well inside one wafer, but large AI systems still require communication between separate systems once workloads become large enough. Cerebras can reduce communication overhead by making each compute unit much larger and faster internally, but it cannot eliminate networking entirely.
Cerebras also does not need to replace NVIDIA outright to succeed. The future AI market may ultimately involve multiple specialized architectures optimized for different tasks. NVIDIA may remain dominant for giant distributed training systems, while Cerebras could become useful for specific inference workloads where speed and low latency matter most.
A company could train large models on NVIDIA GPU clusters and later deploy those models on Cerebras systems for ultra-fast inference. In that sense, the two architectures may function as complementary parts of the same AI infrastructure stack rather than direct substitutes.
Manufacturing Tradeoffs
Traditional chipmakers manufacture many chips from one silicon wafer. If a defect appears on part of the wafer, manufacturers can discard only the defective chip. Cerebras is effectively treating the entire wafer as one processor. That makes manufacturing far more difficult and expensive because defects become much more costly. This issue is called “yield,” and yield problems can significantly increase production costs.
Like NVIDIA, Cerebras does not manufacture its own chips. The company designs the wafer-scale systems, but manufacturing is handled by Taiwan Semiconductor Manufacturing Company, or TSMC. That gives Cerebras access to some of the world’s most advanced semiconductor manufacturing capabilities, but it also creates supplier dependence because TSMC capacity is scarce and heavily demanded by much larger customers, including NVIDIA itself.
Cerebras also does not operate entirely outside the broader semiconductor ecosystem. Like many specialized AI accelerators, Cerebras systems use general-purpose processors and supporting components around the wafer-scale engine itself. Some of those functions may rely on technologies derived from Arm Holdings architectures, although the wafer-scale AI accelerator itself is Cerebras’ own proprietary design.
Cerebras positions itself not just as a chip company, but as an AI infrastructure company. The company sells wafer-scale AI systems, cloud AI services, and large-scale inference infrastructure to organizations involved in AI model development, enterprise AI, government computing, and large-scale cloud infrastructure projects.
Investment Risks
Narrow Customer Base
Its customer base is one of the major risks investors need to understand.
A large percentage of Cerebras’ historical revenue has come from a relatively small number of customers, particularly UAE-linked entities. The most important names are G42, an Abu Dhabi-based AI company, and MBZUAI, the Mohamed bin Zayed University of Artificial Intelligence. Together, those two organizations reportedly accounted for the overwhelming majority of Cerebras’ 2025 revenue.
Cerebras has since announced relationships involving OpenAI, Amazon Web Services, and Mistral AI, including large inference infrastructure agreements extending through 2028. The bullish view is that these newer relationships could gradually diversify the company’s revenue base.
The risk is that Cerebras remains highly dependent on a relatively small number of major customers. If one or two customers reduce spending, delay deployments, move workloads back to GPUs, or build more infrastructure internally, Cerebras’ revenue growth could change very quickly.
Valuation
The technology story around Cerebras is compelling. The valuation story is far more controversial.
Cerebras reportedly generated roughly $510 million in 2025 revenue, up sharply from prior years as AI infrastructure demand exploded. Revenue growth has been extremely rapid, but profitability remains much less clear. While the company has reported accounting gains tied partly to financing structures, operating losses remain significant, and the business still consumes enormous amounts of capital.
At various points following its IPO, Cerebras’ valuation has ranged from roughly $35 billion to well over $70 billion depending on share count assumptions and trading levels. That creates an enormous valuation multiple relative to current revenue and suggests investors are pricing the company largely on expectations about the future growth of AI infrastructure rather than current financial performance.
The bullish argument is that Cerebras has potentially built a genuinely differentiated AI architecture at exactly the moment the industry needs alternatives to NVIDIA. If AI increasingly shifts toward inference, reasoning, voice interaction, AI agents, and low-latency applications, then ultra-fast token generation speed could become extremely valuable. At the same time, many large technology companies do not want the entire AI economy dependent on a single supplier, which could create opportunities for companies like Cerebras if hyperscalers and enterprises increasingly seek alternatives to NVIDIA’s ecosystem.
The bearish case is also substantial.
Cerebras remains highly dependent on future growth assumptions while operating in one of the most competitive and capital-intensive industries in the world. NVIDIA’s ecosystem advantages are enormous. Many technically impressive semiconductor companies have historically struggled because software ecosystems, manufacturing scale, developer adoption, and customer inertia matter just as much as raw hardware performance.
There is also the possibility that investor enthusiasm around AI infrastructure has pushed valuations far ahead of underlying business fundamentals. Cerebras may ultimately become an important AI infrastructure company while still producing disappointing stock returns if the market has already priced in too much future success.
Final Thoughts
Cerebras appears to have genuinely differentiated technology. Its wafer-scale architecture may solve real problems involving AI inference speed and communication bottlenecks, particularly in workloads where latency and rapid token generation matter most.
At the same time, the stock’s valuation depends heavily on future expectations, continued AI infrastructure spending, successful execution, and the company’s ability to scale economically against much larger competitors. Cerebras could potentially establish itself as an important inference platform and become extremely valuable over time, while still experiencing significant volatility if AI spending slows, valuations compress, customer concentration remains high, or the company struggles to scale economically.
Like much of the current AI infrastructure boom, Cerebras may ultimately represent both a genuine technological breakthrough and a market pricing in enormous future expectations at the same time.
Sources and Further Reading
Ben Thompson, Stratechery — The Inference Shift
https://stratechery.com/2026/the-inference-shift/
Cerebras Systems Official Website
https://www.cerebras.ai/
Cerebras Inference Platform
https://www.cerebras.ai/inference
Cerebras WSE-3 Technical Information
https://www.cerebras.ai/product-chip
Reuters — AI Chip Firm Cerebras Raises $1.1 Billion, Adds Trump-Linked 1789 Capital as Investor
https://www.reuters.com/business/ai-chip-firm-cerebras-raises-11-billion-adds-trump-linked-1789-capital-investor-2025-09-30/
Reuters — AI Chip Maker Cerebras Systems Raises $1 Billion in Late-Stage Funding
https://www.reuters.com/business/ai-chip-maker-cerebras-systems-raises-1-billion-late-stage-funding-2026-02-04/
Futurum Group — Cerebras S-1 Teardown
https://futurumgroup.com/insights/cerebras-s-1-teardown-is-the-23b-wafer-scale-ipo-the-end-of-gpu-homogeneity/
Mostly Metrics — Cerebras IPO S-1 Breakdown
https://www.mostlymetrics.com/p/cerebras-ipo-s1-breakdown
Investing in AI — Cerebras Systems and the Speed Bet
https://investinginai.substack.com/p/cerebras-systems-and-the-speed-bet
Yahoo Finance — Cerebras Coverage
https://finance.yahoo.com/
TSMC Official Website
https://www.tsmc.com/
ARM Holdings Official Website
https://www.arm.com/
NVIDIA Official Website
https://www.nvidia.com/
OpenAI Official Website
https://openai.com/
Amazon Web Services Official Website
https://aws.amazon.com/
Mistral AI Official Website
https://mistral.ai/
G42 Official Website
https://www.g42.ai/
Mohamed bin Zayed University of Artificial Intelligence (MBZUAI)
https://mbzuai.ac.ae/
