Amazon Deploys Cerebras Technology 25 Times Faster Than Nvidia GPUs
ავტორი: Wendy Frey
Andrew Feldman inked the deal of his life on March 13. Amazon Web Services, king of clouds, agreed to deploy Cerebras CS-3 chips directly inside its data centers. First hyperscaler partnership. Ever. The wafer-scale insurgents finally broke into the fortress.
The numbers explode. Cerebras claims its dinner-plate-sized chips decode AI responses 25 times faster than Nvidia GPUs. Trainium handles the prompt. CS-3 crushes the output. David Brown, AWS Compute VP, calls it an "order of magnitude" leap. OpenAI already bet $10 billion on Feldman in January. Now Amazon doubles down. Bedrock becomes the gateway to blistering inference.
Disaggregated Architecture Bleeds Nvidia's Crown
The technical guts reveal brutal cleverness. AWS splits inference in two. Trainium3 manages the prefill stage. Cerebras WSE-3 decodes the output. Elastic Fabric Adapter stitches them together. Result? Five times more high-speed token capacity in the same footprint.
Nvidia dominated both phases for years. Not anymore. Startups like Cognition and Mistral already fled to Cerebras for agentic coding. Now AWS offers it to the mainstream. The $23 billion startup (fresh off a $1 billion Tiger Global round in February) finally challenges GPU supremacy at scale. No more buying Cerebras boxes. Just rent through Bedrock.
World Cup 2026 Timing. Enterprise Inference War Explodes
Speed kills. Cerebras delivers thousands of tokens per second while GPUs choke on hundreds. Real-time coding assistance. Interactive agents. Reasoning models that "think" through problems, generating 15x more tokens. Feldman built his empire on these workloads.
AWS rolls it out "in coming months." The second half of 2026 sees full deployment. Amazon Nova and open-source LLMs run on wafer-scale silicon. The $10 billion OpenAI deal (750 megawatts through 2028) suddenly makes sense. Feldman proved his tech at scale. Now AWS brings it to millions of enterprises. The inference bottleneck dies. Price tags remain premium. But for "time is money" workloads, Nvidia finally faces real heat in the cloud.
