AWS makes Nvidia Blackwell instances generally available on EC2 to accelerate enterprise AI workloads
Enzo Reno ·
Amazon Web Services on Tuesday made Nvidia Blackwell based instances generally available on Amazon EC2, opening access to B200 and GB200 powered clusters in select US and EU regions. The launch targets customers training and serving large language models that have strained H100 capacity across clouds, and it lands as AI buyers set budgets for 2027 scale outs.
AWS said customers can access the new P6 family through EC2, managed services in Amazon SageMaker, and integrations with Amazon Bedrock for inference. The company framed the rollout as part of a broader push to expand accelerator choice alongside its in house Trainium and Inferentia chips.
Blackwell on EC2
Nvidia introduced the Blackwell platform at GTC 2024, listing the B200 GPU at up to 20 petaFLOPS of 4 bit AI performance and promoting the GB200 Grace Blackwell superchip for memory bound workloads. Nvidia said Blackwell enables up to four times faster training and as much as 25 times lower total cost of ownership for large language model inference compared with Hopper generation H100 systems, depending on model size and configuration.
AWS is positioning P6 clusters for high throughput training with Elastic Fabric Adapter networking and for low latency inference with multi node NVLink domains. The company said early access customers moved production inference for models in the 7 billion to 70 billion parameter range to Blackwell based nodes to reduce latency and energy use, and it signaled continued support for mixed precision formats including FP8 and FP4.
Enterprise adoption and competitive context
The timing matters because capacity is a differentiator among hyperscalers. Microsoft and Google said in 2024 that they would offer Blackwell in their clouds, while Azure expanded AMD Instinct MI300X capacity in 2024 and Google Cloud scaled H100 A3 and A3 Mega. Omdia estimated Nvidia held more than 80 percent share of data center AI accelerators in 2024, keeping vendor roadmaps anchored on its platform.
AWS linked the EC2 launch to managed AI services. Amazon Bedrock customers can target Blackwell for hosted model endpoints, and Amazon SageMaker supports training and batch transform on the new instances. Amazon also reiterated its commitment to Trainium for training and Inferentia for inference, noting a multi year partnership with Anthropic, which Amazon valued at up to 4 billion dollars in 2023, to use AWS silicon and Nvidia GPUs for future Claude models.
For enterprises, general availability of Blackwell on EC2 provides a clearer upgrade path off H100 era fleets, more headroom for long context and multi agent workloads, and a signal that GPU supply is improving. It also raises the bar for AMD and custom accelerators, steering near term AI platform choices toward clouds that can deliver Blackwell capacity at scale.