Cerebras targets Nvidia's GPU shortage with wafer-scale AI chips

The rising cost of Nvidia GPUs

The demand for Nvidia's H100 GPUs has created a supply crunch that is driving up costs for AI developers. According to data from SemiAnalysis, the rental price for an H100 has climbed roughly 40% since last October. These prices have moved from about $1.70 to $2.35 per GPU-hour based on a survey of more than 100 industry participants.

This price hike comes at a time when on-demand capacity is effectively sold out across major hyperscalers and neocloud providers. Many companies are struggling to find the hardware needed to run large-scale AI models. This scarcity is putting significant pressure on the budgets of firms trying to build out AI infrastructure.

Cerebras is positioning itself to fill this gap by offering an alternative to the standard GPU cluster. The company argues that the industry's heavy reliance on Nvidia is beginning to crack under its own weight. By providing a different hardware path, Cerebras hopes to capture the market as GPU costs become less sustainable.

How wafer-scale chips work

Cerebras uses a unique approach called wafer-scale architecture to build its processors. Instead of using many small chips connected together, they create a single piece of silicon the size of a dinner plate. Their current WSE-3 chip packs 4 trillion transistors and 900,000 cores onto this single wafer using a 5-nanometer TSMC process.

This design allows the company to skip certain manufacturing bottlenecks that affect traditional chips. Specifically, it avoids the sold-out CoWoS packaging step used by many other manufacturers. By making the chip as large as the wafer itself, they can achieve high performance without the complexity of interconnecting hundreds of smaller components.

The technical advantage of this design is reflected in how it handles large language models. Cerebras claims that its hardware requires 97% less code than standard GPU clusters to train the same models. This simplicity is intended to make the hardware more efficient for complex AI workloads that require massive amounts of compute power.

Performance and customer validation

The company is backing its performance claims with significant new contracts from major industry players. In January, OpenAI entered into a long-term agreement exceeding $20 billion to utilize 750 megawatts of Cerebras systems from 2026 through 2028. This deal serves as a major validation of the wafer-scale approach for large-scale AI deployment.

Cerebras also recently partnered with General Compute, a neocloud provider that specializes in alternative chips. General Compute committed to using Cerebras for inference tasks involving agentic coding, which require many sequential calls to plan and write code. This deal represents the largest single hardware commitment the neocloud provider has ever made.

Independent testing has supported these speed claims in several benchmarks. Artificial Analysis measured Cerebras running a trillion-parameter model at 981 output tokens per second. This result was roughly 6.7 times faster than the next-fastest GPU cloud provider and nearly 23 times faster than the median inference provider.

Market position and hurdles

Cerebras recently went public on the Nasdaq, with its initial share price of $185 pushing its market capitalization to roughly $95 billion. While the stock has cooled since its first-day trading pop, the public listing gives the company a direct line to capital markets. This distinguishes it from some GPU-renting neoclouds that are currently facing investor skepticism over debt and cash flow.

Despite its growth, Cerebras still faces significant hurdles in competing with Nvidia's dominance. It currently lacks the CUDA software ecosystem, which is used by over 4 million developers. It also lacks the same level of interconnect technology that Nvidia has spent years perfecting for its GPU clusters.

Because of these factors, most large AI labs are unlikely to replace their existing GPU training pipelines immediately. However, Cerebras is focusing on a narrower claim for inference tasks. As coding agents and chatbots multiply, the speed of wafer-scale silicon may make it an increasingly attractive option for companies looking to reduce costs.

What comes next for Cerebras

The company is scheduled to report its third-quarter earnings on November 12. This upcoming report will be a critical test for the company's new business model. Investors will be looking to see if the large deals with OpenAI and General Compute are successfully translating into actual revenue.

The market will also be watching to see if the stock's recent slide reflects broader fears about AI infrastructure spending. There is ongoing debate about whether the capital required for these systems will generate enough return to satisfy investors. The results will help clarify whether the demand for alternative chips is growing fast enough to offset the high costs of development.

If Cerebras can successfully convert its high-profile contracts into steady growth, it could establish a new standard for AI inference. For now, the company remains in a high-stakes period of proving its hardware can compete with the established GPU status quo. The next few months will determine if it can move from a niche alternative to a mainstream infrastructure pillar.