AWS EC2 G7 Instances Bring NVIDIA Blackwell at 4.6x AI Inference

AWS describes itself as the 'first major cloud provider' to bring NVIDIA's RTX PRO 4500 Blackwell Server Edition GPU to market, and the June 18 announcement signals a deliberate push into a tier that has attracted less coverage than the headline training-cluster race: professional visualization, lighter AI inference, and virtual desktop workloads.

The performance numbers are headline-grabbing. AWS claims up to 4.6x AI inference performance and 2.1x graphics performance over G6 instances, with GPU memory bandwidth running 2.45x higher. The networking story is arguably the more striking figure: configurations ship with 700 Gbps of EFA-enabled networking throughput, which AWS says is 7x the bandwidth of G6. Each GPU carries 32 GB of memory with 5th Gen Tensor Cores and 4th Gen RT Cores on board. The top configuration, the g7.48xlarge, stacks 8 GPUs, 192 vCPUs, 768 GiB system memory, and up to 7.6 TB of local NVMe SSD.

The use cases AWS highlights tell you who the intended buyer is: AI inference, graphics rendering, video transcoding, spatial computing, and virtual desktop infrastructure. The 1.5x concurrent video streams figure versus prior generation is specific enough to suggest media and broadcast operators are a genuine target. Spatial computing developers building for immersive environments get access to Blackwell RT Cores without committing to the larger, pricier training-class instance families.

The honest caveats: G7 launches only in US East (Ohio) and US West (Oregon), which creates an immediate gap for teams with latency requirements or data-residency obligations in Europe or Asia-Pacific. The 4.6x inference claim is a peak figure for an unspecified benchmark, not a per-workload guarantee, so testing before committing to On-Demand pricing matters. What the announcement does not give you is any specific pricing -- purchasing options are named but no dollar figures are disclosed, making TCO comparisons against competing offers difficult until you actually price out an instance.

For developers and infrastructure teams shopping for mid-tier GPU capacity, G7 is a useful marker: Blackwell GPU architecture is now available below the training-cluster tier, and the networking upgrade alone reshapes what multi-GPU inference serving looks like on AWS.