AWS partners with Cerebras to deploy wafer-scale AI inference chips in global data centers

AWS and Cerebras are breaking Nvidia’s grip on AI hardware, reshaping how fast and cost-effectively global businesses can deploy advanced AI.

• AWS and Cerebras announced a multiyear partnership to integrate Cerebras’ WSE-3 chips with AWS Trainium processors for AI inference.
• The new setup splits AI workloads—Trainium handles prompt prefill, Cerebras CS-3 manages token decode—connected by ultra-fast networking.
• Initial rollout is imminent on Amazon Bedrock, promising up to 10x faster AI inference for customers worldwide.

• Cerebras’ WSE-3 chip is 56 times larger than the biggest GPU, offering thousands of times more memory bandwidth and challenging Nvidia’s dominance.
• AWS has invested billions in custom silicon like Trainium to lower AI costs and boost speed, especially as generative AI demand explodes.
• Dubai and the Middle East get access via AWS’s ME-CENTRAL-1 region, potentially accelerating regional AI projects despite recent data center disruptions.

Amazon Web Services (AWS) and Cerebras Systems are redrawing the AI hardware map. For years, Nvidia’s GPUs have powered most of the world’s AI models, but their high cost and supply chain bottlenecks have frustrated cloud providers and enterprise customers alike. Cerebras, a Silicon Valley startup now valued at $23.1 billion, has engineered a radical alternative: the Wafer Scale Engine 3 (WSE-3), a single chip 56 times larger than any GPU, with thousands of times the memory bandwidth.

AWS, meanwhile, has been quietly building its own AI chips—Trainium—for both training and inference, aiming to cut costs and reduce reliance on outside suppliers. The new AWS–Cerebras partnership is about combining these strengths: AWS’s custom silicon for parallel “prefill” (the initial stage of AI model inference) and Cerebras’ CS-3 systems for the serial “token decode” phase, where the model generates outputs word by word. This division of labor, called inference disaggregation, is connected by AWS’s ultra-fast Elastic Fabric Adapter networking.

Why does this matter? Inference—the process of running AI models to generate results—has become the biggest cost and performance bottleneck for companies deploying generative AI at scale. Traditional GPUs are expensive, energy-hungry, and often in short supply. By integrating Cerebras’ wafer-scale chips, AWS claims it can deliver inference speeds up to 10 times faster than current solutions, slashing both latency and cost.

For customers, this means AI-powered services—like chatbots, code generators, and language models—will respond faster and cost less to run. For AWS, it’s a strategic play to attract enterprise clients who are frustrated by Nvidia’s pricing and supply constraints. For Cerebras, it’s a validation milestone ahead of its anticipated IPO, signaling to investors and the market that its chips are ready for prime time at hyperscale.

The partnership also signals a broader industry shift: cloud providers are diversifying their hardware stacks, moving away from single-vendor reliance, and betting on new architectures to keep up with surging AI demand. The initial deployment will focus on open-source large language models and Amazon’s own Nova models, with global rollout on Amazon Bedrock—AWS’s managed AI service—expected within months. Middle East customers, including those in Dubai, will access the new capabilities via AWS’s regional data centers, potentially leapfrogging local infrastructure challenges.

The bottom line: AWS and Cerebras are betting that bigger, more specialized chips—and smarter workload division—will set the pace for the next wave of AI, making advanced inference accessible, fast, and affordable for businesses everywhere.

• Enterprise AI teams and SaaS providers: Faster, cheaper inference unlocks new product features and cost savings.
• Startups building on Amazon Bedrock: Lower barriers to entry for deploying large language models at scale.
• Dubai and Middle East tech firms: Access to cutting-edge AI performance despite regional data center risks.
• Nvidia-dependent cloud customers: New leverage and alternatives for negotiating hardware costs and availability.

• Cerebras IPO timing and investor appetite: The AWS deal is a major credibility boost; watch for market moves as Cerebras goes public.
• Adoption rates on Amazon Bedrock: Uptake by major AI customers will signal whether the new architecture delivers on speed and cost promises.
• Competitor responses (Nvidia, Groq, Google Cloud): Expect new hardware partnerships and pricing shifts as hyperscalers race to diversify.