Amazon Web Services AWS opens EC2 G7 with NVIDIA RTX PRO 4500 Blackwell, claiming 4.6x AI inference over G6
Amazon EC2 G7 reached general availability on 2026-06-18 with NVIDIA RTX PRO 4500 Blackwell Server Edition GPUs, seven sizes, and AWS-quoted gains of 4.6x AI inference and 2.1x graphics over G6, extending the Blackwell push that started with G7e in January.
AWS pushed Amazon EC2 G7 to general availability on 2026-06-18, the third Blackwell-class GPU instance line it has shipped this year and the one positioned at mid-range inference, graphics, and VDI rather than the top of the rack. AWS named it the first major cloud to ship NVIDIA RTX PRO 4500 Blackwell Server Edition silicon, extending a Blackwell lead it has been building since the G7e launch in January and the Bedrock-heavy refresh at Summit New York two days earlier.
What shipped
EC2 G7 supports up to 8 NVIDIA RTX PRO 4500 Blackwell GPUs per instance with 32 GB of memory per GPU, for 256 GB total GPU memory on the largest size. AWS pairs the GPU package with custom sixth-generation Intel Xeon Scalable processors, up to 192 vCPUs, up to 768 GiB system memory, up to 7.6 TB of local NVMe SSD, and up to 700 Gbps of EFA-enabled networking. Seven instance sizes are available at launch from g7.2xlarge (1 GPU, 8 vCPUs) through g7.48xlarge (8 GPUs, 192 vCPUs), with g7.metal on the roadmap and US East (Ohio) plus US West (Oregon) as the launch regions. On-Demand, Savings Plans, Spot, and Dedicated Instances are all available on day one, with Dedicated reserved for 12xlarge and larger.
The performance claims AWS published put G7 at 4.6x AI inference and 2.1x graphics over G6, with 1.33x GPU memory capacity and 2.45x GPU memory bandwidth from the Blackwell move. AWS also calls out 1.5x concurrent video streams over G6 for transcoding lanes. G7 supports NVIDIA GPUDirect P2P across multi-GPU sizes, GPUDirect RDMA over EFA, and GPUDirect RDMA with EFA against Amazon FSx for Lustre, the same low-latency multi-node fabric AWS has been threading through Trainium and the higher-end G7e sizes for training and large-scale inference.
Where this lands in the market
The G7 line sits below the G7e family AWS shipped on 2026-01-20 with NVIDIA RTX PRO 6000 Blackwell silicon. That earlier launch absorbed the high-end inference and rendering workloads with the bigger GPU; G7 reads as the volume tier that gets Blackwell into more general workloads without committing teams to the 6000-class price point. The split keeps AWS aligned with NVIDIA’s RTX PRO Server Edition stack rather than the H-series, which matters for studios, ISVs, and platform teams that have been holding lines on RTX-compatible drivers and tooling rather than migrating to data-center-only architectures.
Google Cloud and Azure have not yet shipped a comparable Blackwell-accelerated managed instance at this tier. AWS first named RTX PRO 4500 support at GTC 2026 in March, and the gap between announcement and GA is shorter than the G6e or P5 cycles, which suggests AWS is treating Blackwell availability as a competitive lever rather than a capacity-constrained ship. The operational read is that AWS gets to bundle Blackwell into the Bedrock and AgentCore stories it told at Summit New York on 2026-06-17 without renting the silicon from a third party, which narrows what hyperscaler peers can claim on agent-runtime cost performance through the second half of 2026.
What’s worth watching
- Region expansion cadence. GA in Ohio and Oregon only is a sign of Blackwell allocation rather than a soft launch. Whether AWS adds Frankfurt, Tokyo, or Sydney during Q3 will set expectations for teams that need Blackwell-class inference under EU or APAC data residency.
- G7.metal availability. Spatial computing and VDI shops that need passthrough access have been waiting for the metal SKU to evaluate against Azure’s NVads-A10 v5 and bare-metal Equinix options. The metal SKU is the deciding piece for studios that prefer raw GPU access over a hypervisor layer.
- Bedrock and AgentCore steering. AWS has been quietly moving first-party inference traffic onto Blackwell-class instances for the Bedrock Managed Knowledge Base and AgentCore Web Search launches. G7 is the tier most likely to host the customer-facing side of that traffic. If Bedrock pricing on Anthropic or OpenAI models notably tightens in the next quarter, G7 is the infrastructure step that made it possible.