Amazon AWS Triples Nvidia Blackwell GPU Orders Amid Surging Enterprise Inference Demand

The Three-Question Framework

What Happened?

Amazon Web Services (AWS) confirmed a dramatic expansion of its semiconductor procurement contracts, tripling its purchase orders for Nvidia’s Blackwell GB200 and NVL72 rack-scale systems. The multi-billion dollar order comes despite Amazon’s heavy internal investments in its proprietary Trainium2 and Inferentia3 silicon chips.

Why It Matters

The move underscores that customer enterprise demand for cutting-edge generative AI and multi-step reasoning models is outpacing internal silicon development timelines. Enterprise developers are heavily locked into the CUDA software ecosystem; to prevent cloud customers from migrating to Microsoft Azure or Google Cloud, AWS must provide immediate, unconstrained access to Nvidia’s flagship architectures.

What Could Happen Next

AWS is commissioning liquid-cooled datacenter modules in Virginia, Ohio, and Oregon to host the new GB200 NVL72 racks, which consume up to 120 kilowatts per cabinet.

Executive Key Takeaways

AWS triples orders for Nvidia Blackwell GB200 chips to satisfy soaring cloud customer demand.

Liquid-cooled GB200 NVL72 racks deliver 30x faster inference for trillion-parameter reasoning models.

Demonstrates the enduring enterprise lock-in of Nvidia’s CUDA software ecosystem over custom cloud silicon.

Requires massive retrofit of datacenter cooling infrastructure to support 120kW per rack power densities.

In late August 2026, Amazon Web Services dramatically adjusted its datacenter capital expenditure projections, confirming it has tripled its hardware allocation orders for Nvidia’s Blackwell architecture.

For the past several quarters, Amazon leadership had emphasized its homegrown silicon roadmap, encouraging cloud customers to deploy workloads onto its proprietary Trainium2 accelerators to reduce dependence on Nvidia’s premium pricing. However, as enterprise customers began rolling out massive multi-turn reasoning models and agent swarms in production, the demand for Nvidia’s highest-end interconnect topologies proved irresistible.

At the center of the surge is Nvidia's GB200 NVL72 system. By connecting 36 Grace CPUs and 72 Blackwell GPUs via a massive copper-passive NVLink backplane, the rack operates essentially as a single giant GPU with 1.4 exaflops of FP4 inference compute and 30 terabytes of ultra-fast unified memory. For foundation models executing complex test-time reasoning chains, the architecture offers a 30-fold speedup over previous generation H100 clusters.

Deploying these systems at scale has required AWS to aggressively overhaul its facility designs. Standard datacenter racks historically operated at 10 to 15 kilowatts of power; the NVL72 systems consume over 120 kilowatts per enclosure, necessitating direct-to-chip liquid cooling manifolds and closed-loop heat exchangers.

The massive order reaffirms Nvidia's unchallenged position as the primary hardware supplier for the global AI revolution, with cloud hyperscalers continuing to commit immense balance-sheet reserves to ensure compute availability.