AWS to Deploy 2 Million More NVIDIA GPUs by 2028: Inside the Hyperscale AI Superclusters & Blackwell Ultra Expansion

In one of the largest infrastructure expansion announcements in cloud computing history, Amazon Web Services (AWS) and NVIDIA announced a massive expansion of their strategic partnership on August 27, 2026, with AWS committing to deploy 2 million additional NVIDIA GPUs across its global data center fleet throughout 2027 and 2028.

This monumental order builds upon an earlier commitment of 1 million GPUs announced in late 2025, pushing AWS's total announced multi-year NVIDIA accelerator rollout to over 3 million high-performance GPUs.

The expanded collaboration encompasses NVIDIA Blackwell Ultra, next-generation Rubin and Rubin Ultra architectures, full-stack NVIDIA Vera CPU deployments, and dedicated 100,000-GPU secure "AI Factories" for federal and defense workloads.

---

#1. The Global Compute Boom: Why 2 Million Additional GPUs?

As enterprises transition from simple conversational chatbots to autonomous agentic AI systems, multimodal foundation models, and real-time physical AI simulations (robotics and autonomous driving), hyperscalers face unprecedented compute deficits.

mermaidCopy
graph TD
    A[Surging Enterprise AI Workloads 2026-2028] --> B[Multi-Step Autonomous AI Agents]
    A --> C[Trillion-Parameter Mixture-of-Experts Training]
    A --> D[Real-Time Multimodal Video & Robotics Inference]
    B & C & D --> E[AWS Infrastructure Response: +2M NVIDIA GPUs by 2028]
    E --> F[Blackwell Ultra Superclusters - 2027]
    E --> G[Rubin & Rubin Ultra Superclusters - 2028]
    E --> H[100K GPU Sovereign AI Factories - U.S. Federal IL6+]

---

#2. Hardware Architecture Comparison: Blackwell Ultra vs. Rubin

The deployment will introduce NVIDIA's next two silicon generations across AWS EC2 instances, ultra-clusters, and SageMaker environments:

---

#3. Sovereign & Federal AI Factories: 100,000 GPUs on Secure AWS

A centerpiece of the new agreement is the joint construction of dedicated Government AI Factories:

• 100,000 Dedicated NVIDIA GPUs deployed in physically air-gapped AWS GovCloud and Secret/Top Secret cloud regions.
• Engineered to meet DoD Impact Level 6 (IL6) and Intelligence Community classification standards.
• Intended to accelerate defense simulation, cryptographic research, national intelligence analysis, and sovereign scientific discovery without data leaving government-controlled cryptographic perimeters.

---

#4. Full-Stack Co-Engineering: Annapurna Labs & NVIDIA NVHBM Integration

Beyond purchasing raw GPUs, Amazon's in-house chip design unit, Annapurna Labs, is collaborating directly with NVIDIA engineering teams:

mermaidCopy
graph LR
    A[Amazon Annapurna Labs] -->|Integrates NVHBM Memory & Spectrum-X| B[AWS Trainium 3 / Inferentia 3 Clusters]
    C[NVIDIA Hardware Stack] -->|Vera CPU + Blackwell Ultra| D[Amazon EC2 G7 & P6 Fleets]
    B & D -->|Unified Fabric: 800Gbps EFA & NVLink| E[Seamless Hybrid Inference & Training]

#Key Joint Engineering Initiatives:

1. NVHBM Integration with Trainium: Annapurna Labs will incorporate NVIDIA's high-bandwidth memory interface (NVHBM) into future iterations of AWS Trainium chips, boosting memory density by 40% while cutting per-token energy draw.
2. Spectrum-X Networking Fabrics: AWS will deploy NVIDIA Spectrum-X Ethernet switches alongside its proprietary Elastic Fabric Adapter (EFA), providing multi-tenant predictability for 100,000+ node AI clusters.
3. Launch of Amazon EC2 G7 Instances: Introducing the NVIDIA RTX PRO 4500 Blackwell Server Edition, optimized for generative 3D graphics, enterprise rendering, and mid-tier LLM fine-tuning at 35% lower cost than full FP8 training nodes.

---

#5. What This Means for Enterprises, Startups & Developers

1. Capacity Reliability: Alleviates severe GPU waitlists and compute quotas for foundation model labs and enterprise AI developers.
2. Lower Per-Token Inference Costs: Higher memory bandwidth on Blackwell Ultra and Rubin allows larger models (e.g., Llama 4, Claude 3.7, DeepSeek-V3) to run on fewer GPUs with lower operational overhead.
3. Hybrid Architecture Flexibility: Developers can seamlessly train on NVIDIA Blackwell clusters while running high-throughput production inference on AWS Trainium chips using unified AWS SageMaker orchestration.

---

#Building Scalable, Cloud-Native AI Applications with Brandomize

As global cloud infrastructure reaches unprecedented compute scale, building enterprise software that fully leverages distributed GPUs, optimized inference APIs, and low-latency pipelines is essential for competitive advantage.

At Brandomize, we engineer robust web architectures, scalable API microservices, and AI-powered digital products that scale seamlessly from prototype to millions of active users.

Ready to build enterprise-grade digital systems? Get in touch with the Brandomize team today.