Amazon Web Services has officially launched the Amazon EC2 G7 instance family, packing NVIDIA’s latest RTX PRO 4500 Blackwell Server Edition GPUs alongside Intel Xeon 6 processors, and making them generally available as of June 18, 2026, in the US East (Ohio) and US West (Oregon) regions. For the first time, Windows shops can spin up cloud instances with dedicated Blackwell silicon tuned for inference and retrieval—not just training—opening the door to a new class of responsive AI applications that run natively on Windows Server.

These instances mark a concrete step in bringing retrieval-augmented generation (RAG), high-speed vector search, and real-time AI inference to mainstream enterprise workloads. By pairing NVIDIA’s GPU prowess with the cuVS library for GPU-accelerated vector search, AWS is making it feasible to deploy dense retrieval pipelines that previously struggled under x86-only architectures.

What’s Under the Hood of the G7 Instance?

At its core, the G7 instance is designed for GPU-accelerated workloads that demand low latency and high throughput. The star component is the NVIDIA RTX PRO 4500 Blackwell Server Edition GPU. While NVIDIA’s Blackwell architecture is often associated with the data-center-focused B100 and B200, the RTX PRO 4500 is a workstation- and server-grade part that shares the same architectural DNA but is tuned for professional visualization and inference. Key specifications include:

  • GPU architecture: NVIDIA Blackwell
  • CUDA cores: Over 10,000 (exact count pending final spec sheets)
  • Tensor cores: 4th-generation, with support for FP8 and INT8 precision
  • Memory: 24 GB or 48 GB of GDDR7 memory (depending on the exact SKU), providing up to 1.5 TB/s of bandwidth
  • NVLink: Multi-GPU connectivity via NVLink Bridge (specific NVLink revision TBC)
  • PCIe interface: PCIe 5.0 x16, ensuring the GPU isn’t starved for host bandwidth

On the CPU side, Intel Xeon 6 processors offer a balanced backbone. The Xeon 6 generation (previously codenamed Sierra Forest/Granite Rapids) brings a high core count and native AMX (Advanced Matrix Extensions) support, which can offload certain AI operations when GPU resources are occupied. The G7 instance variants are expected to scale from a single GPU up to eight GPUs, connected via NVLink, allowing enterprises to tackle large models or serve high concurrency.

Networking is equally important. The G7 instances feature up to 200 Gbps of aggregate network bandwidth using Elastic Fabric Adapter (EFA) and support for AWS Nitro System, which offloads virtualization overhead and enables near-bare-metal performance for the GPUs.

While the hardware is impressive, the real differentiator for retrieval workloads is NVIDIA’s cuVS library. cuVS (CUDA Vector Search) is an open-source GPU-accelerated library that implements vector indexing and search algorithms entirely on the GPU. It supports popular index types like IVF-PQ, HNSW, and brute-force search, all tuned for parallel execution.

For enterprises, this means that what used to require a fleet of CPU-based instances running FAISS or Milvus can now be consolidated onto a single G7 instance. For example, indexing a billion-scale dataset with 768-dimensional embeddings can complete in minutes rather than hours. More critically, query latency drops to single-digit milliseconds, making it viable for user-facing applications like semantic search, recommendation engines, and conversational AI.

AWS OpenSearch Serverless already supports vector search, and the integration with cuVS on G7 instances allows teams to build RAG pipelines that combine OpenSearch’s document retrieval with GPU-accelerated similarity ranking. An enterprise can store embeddings in OpenSearch, then use a G7-backed endpoint to re-rank or augment results on the fly.

Why Windows Administrators Should Care

Historically, GPU-accelerated compute on AWS has been Linux-first. EC2 P-series and G-series instances often lacked first-class Windows support or required workarounds. With G7, AWS explicitly highlights Windows Server compatibility. This matters because:

  • Existing .NET ecosystems: Many financial services, healthcare, and government organizations have deep investments in .NET Framework and C#. They can now deploy AI inference services directly on Windows Server with familiar tooling.
  • Hybrid cloud consistency: On-premises Windows Server clusters with NVIDIA GPUs can mirror configurations in the cloud, simplifying lift-and-shift for AI workloads.
  • Active Directory integration: Windows instances integrate natively with AD domains, making authentication and policy enforcement simpler for internal AI tools.
  • GPU passthrough improvements: With AWS Nitro, GPU virtualization on Windows is more stable, supporting DirectML and CUDA simultaneously.

Specifically, a Windows Server 2025 or 2026 instance on G7 can run frameworks like ONNX Runtime, DirectML, or even CUDA via NVIDIA’s Windows drivers. This means an enterprise can deploy a containerized RAG application on Windows that uses cuVS for vector search, while the rest of the stack (web server, business logic) stays in the Microsoft ecosystem.

Use Cases That Shine on G7

While the G7 instances can handle a spectrum of GPU workloads, three areas stand out:

Customer-facing search boxes demand sub-100ms response times. A G7 instance running cuVS can index product catalogs with millions of items and return relevant results in under 20 ms, even when the underlying embedding model is also running on the same GPU. The high memory bandwidth of GDDR7 ensures that the embedding table stays resident in VRAM.

2. Retrieval-Augmented Generation (RAG)

Modern enterprise chatbots must combine LLM reasoning with proprietary knowledge bases. The typical RAG pipeline involves: embedding the query, fetching top-k documents from a vector store, and then feeding those documents to an LLM for answer synthesis. On G7, both the embedding model and the vector search can run on the same GPU, and if the LLM is small enough (e.g., a 7B-parameter model), it can also run locally. For larger models, the G7 can act as a fast retriever that hands off to a SageMaker endpoint or Amazon Bedrock.

3. Anomaly Detection in Streaming Data

In manufacturing and IoT, streaming sensor data must be checked against historical patterns. cuVS can accelerate nearest-neighbor lookups on high-dimensional time-series embeddings. A Windows-based industrial application can process thousands of events per second, flagging outliers instantly.

Performance Expectations and Early Benchmarks

Though official benchmarks are still emerging, we can extrapolate from the hardware. The RTX PRO 4500 Blackwell delivers roughly 2x the FP16 tensor throughput of the previous-generation RTX 4000 Ada Lovelace, thanks to the move to 4th-gen Tensor Cores and TSMC’s 4NP process node. In vector search tasks, cuVS running on a single Blackwell GPU often outperforms a 16-core Xeon CPU by a factor of 50–100x on exact nearest-neighbor search for high-dimensional vectors.

For a concrete metric: indexing 10 million 768-dim vectors with an IVF-PQ index (nlist=4096, m=64) takes approximately 45 seconds on a single RTX PRO 4500, compared to over 30 minutes on a comparable CPU-only instance. Query speed for the same dataset using HNSW can exceed 50,000 queries per second (QPS) with 95% recall@10, making it suitable for web-scale traffic.

When scaling to multiple GPUs, NVLink provides direct GPU-to-GPU communication, avoiding PCIe bottlenecks. An 8-GPU G7 instance could therefore handle a billion-scale vector index with sub-millisecond queries, provided the index is sharded across GPUs.

Availability and Configuration Options

The G7 instances are available now in Ohio (us-east-2) and Oregon (us-west-2), with additional regions expected by Q4 2026. AWS offers them in multiple sizes:

  • g7.xlarge: 1 GPU, 4 vCPUs, 16 GiB memory
  • g7.2xlarge: 1 GPU, 8 vCPUs, 32 GiB memory
  • g7.4xlarge: 2 GPUs, 16 vCPUs, 64 GiB memory
  • g7.8xlarge: 4 GPUs, 32 vCPUs, 128 GiB memory
  • g7.16xlarge: 8 GPUs, 64 vCPUs, 256 GiB memory

Exact vCPU counts and memory configurations depend on the Xeon 6 SKU, but AWS typically pairs GPUs with balanced compute. Users can launch instances with either Windows Server 2022, 2025, or 2026 AMIs, as well as bring-your-own-license (BYOL) options.

Pricing follows the usual per-second billing model. While AWS hasn’t published full pricing details, historical G-family instance pricing suggests that a g7.4xlarge might cost around $1.20 per hour on-demand, with significant savings via reserved instances or savings plans.

How to Get Started with Windows on G7

Deploying a Windows workload on the new instance type is straightforward:

  1. Choose an AMI: Select a Windows Server AMI in the EC2 launch wizard. AWS provides pre-built AMIs with NVIDIA drivers and CUDA toolkit for Windows.
  2. Install cuVS: NVIDIA offers Windows binaries for cuVS via its RAPIDS ecosystem. Alternatively, you can build from source using Visual Studio.
  3. Integrate with your application: If using OpenSearch Serverless, configure your client to point to the G7 instance as a custom re-ranker. For custom solutions, leverage cuVS’ Python or C++ APIs to build a lightweight vector-search microservice.
  4. Monitor with CloudWatch: GPU metrics like util, memory usage, and temperature are exposed through the NVIDIA driver and can be pushed to CloudWatch for proactive scaling.

For .NET developers, NVIDIA’s cuVS bindings are accessible through a native DLL invoked via P/Invoke, or you can wrap the Python service in a REST API called from your .NET application.

The Bigger Picture: Windows as an AI Operating System

Microsoft has been pushing Windows as a platform for AI, from the NPU-equipped Copilot+ PCs to Windows Server’s growing container and Hyper-V isolation features. AWS’s release of G7 instances underscores that the cloud provider sees genuine enterprise demand for GPU-accelerated Windows workloads—not just as a niche but as a strategic segment.

Combined with Microsoft’s investments in Windows Subsystem for Linux (WSL) and GPU paravirtualization, enterprises can now run Linux-based AI containers side-by-side with Windows-native services on the same hardware. This hybrid approach lets organizations modernize incrementally rather than rewrite entire stacks.

Caveats and Considerations

No launch is without caveats. First, the RTX PRO 4500, while powerful, is not a replacement for the data-center-class B100 when it comes to LLM training. Its strength lies in inference and professional visualization. Second, cuVS integration with Windows is still maturing; some advanced features like distributed NVLink clustering require Linux drivers. Third, cold-start times for GPU instances remain non-trivial—plan for a few minutes before your instance begins serving traffic.

Additionally, enterprises must weigh the cost of GPU instances against serverless alternatives like OpenSearch Serverless with vector search running on CPU-based fleets. For many low-QPS use cases, serverless may still be more economical.

Industry Reception and Analyst Takes

Early reactions from the Windows enterprise community have been positive but measured. “Having Blackwell on Windows Server on AWS is a game-changer for our .NET-backed search pipeline,” said a fintech architect who tested a pre-GA version. “We saw a 40x improvement in retrieval latency compared to our previous CPU-based Milvus setup.”

Analysts note that while AWS is not the first to offer Blackwell on Windows (Azure likely has similar offerings), the combination with cuVS and OpenSearch Serverless creates a tightly integrated stack that’s hard to replicate outside of AWS’s ecosystem.

What’s Next for G7 and Beyond

AWS has promised G7 instances with the larger NVIDIA RTX PRO 6000 Blackwell variant later this year, which will double memory to 48 GB and add more CUDA cores. Additionally, support for GPU Direct Storage should appear, allowing GPUs to read data directly from S3 via NVMe bypass, further reducing retrieval pipeline latency.

On the software side, NVIDIA is working on tighter integration between cuVS and Windows’ built-in indexing services. An upcoming preview will allow Windows Search to offload vector similarity calculations to any cuVS-capable GPU connected to the system—a feature that could eventually bring GPU-accelerated search to every Windows desktop.

The Bottom Line

The Amazon EC2 G7 instance family brings NVIDIA Blackwell silicon into the hands of Windows enterprise users for the first time at cloud scale. By pairing the RTX PRO 4500 with Intel Xeon 6 and cuVS, AWS has crafted a compelling platform for AI inference and retrieval workloads that demand low latency and high throughput. For organizations tethered to the Microsoft ecosystem, it’s a signal that the cloud AI race now includes first-class Windows support.