Amazon Web Services (AWS) announced a multi-year partnership with Cerebras Systems to deploy the startup's wafer-scale AI training processors across its cloud infrastructure. The deal, valued at an estimated $4.5 billion, represents the largest commitment by a hyperscaler to a non-Nvidia AI chip platform and signals a shift in the semiconductor race that has powered the AI boom.
Cerebras builds the world's largest computer chip. Its WSE-3 processor contains 4 trillion transistors on a single wafer-scale die, compared to Nvidia's H100 GPU with 80 billion transistors. The architecture excels at training large language models where data movement between chips creates bottlenecks.
What This Deal Changes
- AWS becomes the first major cloud provider to offer Cerebras-powered AI training at scale
- The deal is worth an estimated $4.5 billion over five years
- Cerebras chips train large language models up to 10x faster than traditional GPU clusters for certain workloads
- Nvidia shares dipped 2.1% on the announcement before recovering by end of day
Why Amazon Is Diversifying Away from Nvidia
Amazon, Google, and Microsoft collectively spent over $200 billion on AI infrastructure in 2025. Nvidia captured roughly 80% of the AI training chip market, giving it enormous pricing power. H100 GPUs sell for $30,000-$40,000 each, and the newer B200 chips exceed $50,000.
By partnering with Cerebras, Amazon reduces its dependence on a single supplier and gains leverage in future procurement negotiations. AWS already designs its own Trainium and Inferentia chips for specific workloads. The Cerebras deal adds another option for customers who need maximum training performance.
"The AI chip market is too important to be a monopoly," said Andy Jassy, Amazon CEO, during the announcement. "Our customers deserve choice, competition, and innovation across the hardware stack."
How Cerebras Technology Works
Traditional AI training distributes a model across hundreds or thousands of GPUs connected by high-speed networking. Each GPU processes a portion of the data, then shares results with its neighbors. This communication overhead limits scaling efficiency.
Cerebras takes a different approach. Its wafer-scale engine places all compute on a single, massive chip. This eliminates inter-chip communication delays. For models with 100 billion to 1 trillion parameters, Cerebras claims training speedups of 5-10x compared to equivalent GPU clusters.
The Cost Advantage
Speed is only part of the equation. Cerebras systems consume 20-30% less electricity per unit of compute than GPU clusters, according to the company's published benchmarks. For hyperscalers running thousands of training jobs per month, the energy savings translate to tens of millions of dollars annually.
Nvidia's Response and Market Position
Nvidia remains the dominant force in AI hardware. Its CUDA software ecosystem, which allows developers to write code that runs efficiently on Nvidia GPUs, creates a switching cost that competitors struggle to overcome. CEO Jensen Huang has repeatedly said the company's software moat is deeper than its hardware lead.
Nvidia's stock dipped 2.1% on the Amazon-Cerebras news before recovering. Analysts at Bank of America maintained their "Buy" rating, noting that Nvidia's Q4 2025 revenue of $42 billion demonstrates demand that no single competitor can absorb.
The Broader AI Chip Market in 2026
The AI semiconductor market is projected to reach $200 billion by the end of 2026, up from $125 billion in 2025. Beyond Nvidia and Cerebras, several players are competing for share. Google deploys its TPU v5 chips internally. AMD's MI300 series has gained traction with cost-conscious buyers. Intel's Gaudi 3 targets the inference market.
For investors, the Amazon-Cerebras deal validates the thesis that the AI hardware market will support multiple winners. A diversified approach through the VanEck Semiconductor ETF (SMH) or individual positions in Nvidia, AMD, and Cerebras (when it goes public) captures the full breadth of the opportunity.
