Sacra estimates that Cerebras booked sales of $250M in 2023, representing 201% growth year-over-year from about $83M in 2022. The company recently announced that it was on track to double in 2024, with total customer commitments approaching $1B. This growth trajectory places Cerebras firmly in the realm of major players in the AI hardware and services market, and makes them the largest revenue-generating AI chip startup.

Cerebras has secured several high-profile customers and partnerships that contribute to its revenue growth. These include contracts with national laboratories like Argonne, Lawrence Livermore, and Sandia, as well as large commercial entities such as GlaxoSmithKline, TotalEnergies, and Mayo Clinic. The company has also entered into a strategic partnership with G42, an Abu Dhabi-based AI company, to build world-class supercomputers.

Founded in 2016, Cerebras Systems is an AI hardware manufacturer and computing services provider, specializing in wafer-scale technology for artificial intelligence workloads. The company's flagship product is the CS-3 system, powered by the Wafer Scale Engine 3 (WSE-3). The WSE-3 is the largest and fastest chip ever created, measuring 46,000 square millimeters (about the size of a dinner plate) and built on TSMC's 5nm process.

Traditional AI chips, like those from NVIDIA or Groq, are typically the size of a postage stamp and are clustered together to achieve high performance. In contrast, the WSE-3's design offers a few critical advantages for AI workloads:

Data Locality: By keeping all computations on a single, large chip, Cerebras dramatically reduces the need for data movement. In traditional GPU clusters, data must constantly move between chips, consuming time and energy. The WSE-3 keeps data local, resulting in significantly faster processing and lower power consumption.

Memory Bandwidth: The WSE-3 has vast on-chip memory with unprecedented bandwidth. This allows for faster data access and reduces bottlenecks commonly experienced in distributed GPU systems.

Simplified Scaling: While GPU clusters require complex software to distribute workloads across multiple chips, Cerebras's single-chip approach simplifies this process. This makes it easier to scale up to larger AI models without the overhead of managing distributed computing resources.

These advantages make Cerebras systems particularly well-suited for training large AI models and handling complex, data-intensive workloads. For instance, training a 175 billion parameter model on 4,000 GPUs might require 20,000 lines of code for distribution, while Cerebras can accomplish this with just 565 lines of code in one day.

Overall, Cerebras positions itself for customers who have outgrown smaller-scale AI solutions and need supercomputer-level performance. Their systems are designed for organizations training large models or working with vast datasets, typically those spending millions on AI compute annually. In addition to hardware, Cerebras offers professional services, including assistance with data preparation, model design, training oversight, and optimization. This service component accounts for between a quarter and a third of new customer engagements.

The company has found product-market fit across several sectors.

In the government and research space, Cerebras has run the table on supercomputer labs, including Argonne National Laboratory, Lawrence Livermore National Laboratory, and Sandia National Laboratory, among others. Governments worldwide are also investing in Cerebras technology for sovereign clouds to meet domestic AI requirements.

In the private sector, Cerebras serves customers in pharmaceuticals and life sciences, such as GlaxoSmithKline, which uses the technology for pioneering work in epigenomics, while Mayo Clinic is using Cerebras systems to enhance medical diagnostics and personalized medicine.

In the energy sector, TotalEnergies, a $100 billion French company, is a notable customer. While not specifically named, Cerebras also works with companies in the finance space on complex modeling and risk assessment.

Cerebras Systems has a few separate revenue streams:

1. Hardware Sales: The primary revenue source comes from selling CS-3 systems, powered by the Wafer Scale Engine 3 (WSE-3). These systems are priced in the millions of dollars, targeting high-end customers with substantial AI computing needs.

2. Professional Services: Accounting for 25-33% of new customer engagements, this includes data preparation, model design, training oversight, and optimization services.

3. Cloud Services: Cerebras offers cloud-based access to its systems through partnerships and its AI Model Studio, providing a recurring revenue stream.

4. Subscription Models: The company is exploring subscription-based offerings, placing hardware on customer sites under a license agreement.

Hardware sales likely constitute the majority of revenue, followed by professional services. Cloud and subscription services are likely growing segments.

While gross margin hasn't been publicly reported, AI chipmaker AMD reported gross margins of 47% last quaarter.

The AI accelerator hardware market is dominated by NVIDIA's GPUs, which hold approximately 85% market share. NVIDIA's H100 GPU is the current industry standard for AI training and inference, offering high performance and widespread software support through CUDA.

AMD and Intel are NVIDIA's primary competitors in this space, with AMD's MI300 series and Intel's Gaudi2 chips aiming to capture market share.

The key characteristic that Cerebras is differentiating itself around in this category is simplicity of scaling.

While GPU-based systems require complex networking and software to distribute workloads across multiple chips, Cerebras' single-chip approach eliminates much of this complexity. This can potentially reduce the time and expertise required to set up and manage large-scale AI training infrastructure.

Major cloud providers like Google, Amazon, and Microsoft are developing their own specialized AI chips to reduce dependence on NVIDIA and offer differentiated AI services. Google's Tensor Processing Units (TPUs), Amazon's Trainium and Inferentia, and Microsoft's Azure Maia AI Accelerator represent significant investments in this area.

Cerebras competes in this category by offering its technology as a cloud service, allowing customers to access WSE-powered systems without the need for on-premises hardware. The company has partnered with Cirrascale Cloud Services to make its CS-1 and CS-2 systems available in the cloud.

This approach allows Cerebras to target customers who need high-performance AI training capabilities but may not have the resources or desire to manage their own hardware.

Cerebras' advantage here lies in its ability to offer a unique hardware architecture that isn't available from major cloud providers. The company claims its systems can handle larger models and provide faster training times than cloud-based GPU clusters, potentially attracting customers with demanding AI workloads.

A number of well-funded startups are developing novel AI chip architectures, including Graphcore, SambaNova Systems, and Groq.

These companies are exploring various approaches to AI acceleration, such as Intelligence Processing Units (IPUs) and tensor streaming processors.

In this category, Cerebras stands out for its wafer-scale approach and focus on very large language models. While other startups are developing chips that compete more directly with traditional GPUs, Cerebras has carved out a niche for training massive AI models that may be impractical on other hardware.

Cerebras' recent partnership with Qualcomm for AI inference represents a strategic move to address the full AI workflow.

By optimizing its training output for Qualcomm's Cloud AI 100 inference processor, Cerebras aims to offer a more complete solution that spans from model development to deployment. This collaboration could give Cerebras an edge over other AI chip startups that focus primarily on training or inference.

The AI hardware market is projected to reach nearly $250B by 2030, and companies are showing willingness to spend hundreds of millions or even billions on supercomputer-scale AI clusters.

With its unique wafer-scale technology providing substantial speedups over alternatives, Cerebras's upside hinges on becoming the architecture of choice for the high end of the market: multi-billion dollar AI supercomputers used by hyperscalers, governments, and the largest enterprises.

In addition to this core tailwind from the growth of the AI market, Cerebras has several opportunities to expand its overall TAM.

While Cerebras initially focused on training, where its huge memory capacity and bandwidth shine, the company is expanding to also accelerate inference. This opens up the broader market of deploying trained models into production to generate text, images, and other AI-powered experiences.

As generative AI takes off, demand for cost-effective, high-performance inference will surge. Cerebras' wafer-scale engine could become a popular choice for deploying the largest models where real-time responsiveness matters, such as robotics, autonomous systems, and interactive applications. The inference market is projected to be even larger than the training market long-term.

Cerebras has initially focused on the high-end AI research and development market, with some expansion into healthcare and life sciences, but there is substantial potential for expansion into other verticals that require massive computational power:

1. Financial Services: High-frequency trading, risk modeling, and fraud detection systems could benefit from the WSE's computational capabilities.

2. Climate Modeling: The company's systems could enhance climate simulations, improving our understanding of global weather patterns and climate change.

3. National Security and Defense: Government agencies could leverage Cerebras' technology for complex simulations, cryptography, and intelligence analysis.

Software compatibility: The AI ecosystem has standardized around Nvidia's CUDA platform. Cerebras' bespoke software stack may struggle to gain traction with developers accustomed to CUDA's mature tooling and libraries. Without a strong software moat, Cerebras risks being a marginal player.

Long-term business model: In the near-term, Cerebras can grow by selling high-margin hardware to enterprises for whom AI is mission-critical. But in the longer-term, it's unclear if being a specialized chip vendor is the best position in the value chain. The industry may re-centralize around a few hyperscale AI platforms who view chips as a cost center. Cerebras may need to forward-integrate into cloud services to capture sufficient value.

Niche market fit: Cerebras is building extremely large, expensive systems that only make sense for the most compute-intensive AI workloads. This limits their total addressable market to a small number of hyperscalers and deep-pocketed enterprises. If AI model growth slows or shifts towards efficiency over scale, Cerebras may be overly specialized for a shrinking niche.