AWS and NVIDIA have confirmed plans to deploy 2 million additional GPUs across AWS's global infrastructure in 2027-2028, on top of an existing 2026 commitment of over 1 million units. For UK buyers, this changes the maths behind on-prem versus hyperscaler capacity planning for the next agentic AI investment cycle.
View the data behind this chart
| 2026 (GTC pledge) | 2027-2028 (expansion) | |
|---|---|---|
| GPUs pledged (millions) | million…1 | million…2 |
What AWS and NVIDIA actually announced
The two companies confirmed a major expansion of a partnership now in its 16th year, covering far more than raw compute. Alongside the 2 million additional NVIDIA Blackwell Ultra, Rubin and Rubin Ultra GPUs planned for AWS's global infrastructure in 2027-2028, the deal brings NVIDIA Vera CPU-based infrastructure to AWS, extends NVLink Fusion with custom high-bandwidth memory, and deepens integration with the AWS Nitro System and Elastic Fabric Adapter for security and reliability.
This builds directly on AWS's GTC 2026 pledge to add more than 1 million NVIDIA GPUs starting in 2026 — a figure the companies say demand has now exceeded. AWS is also expanding Blackwell capacity for Amazon EC2 G7 instances, including RTX PRO 4500 Blackwell Server Edition GPUs, which AWS says deliver 4.6x the AI inference performance of the previous G6 generation.
Why the 2027-2028 timeline matters for UK capacity planning
The headline number is not a 2026 shipment — it's a multi-year deployment window running into 2028. That gives UK IT leaders a rare planning advantage: visibility into hyperscaler GPU supply roughly two years ahead of workload need. Organisations building business cases for agentic AI rollouts in 2027 or 2028 now have a concrete signal that Blackwell Ultra, Rubin and Rubin Ultra capacity will be available at hyperscale rather than gated by chip scarcity.
That doesn't remove the need to model costs carefully. Buyers should still calculate your AI infrastructure TCO across the full 2027-2028 window rather than locking in assumptions based on today's pricing, since a 2-million-GPU supply expansion of this scale is likely to affect availability and commercial terms across the wider market.
Agentic AI workloads need more than GPU count
NVIDIA's framing is explicit: this capacity is meant for agentic AI, scientific discovery, enterprise automation and physical AI, not just large-scale model training. Vera CPU-based infrastructure and NVLink Fusion with custom high-bandwidth memory point to a shift toward tightly integrated CPU-GPU platforms designed for the orchestration and reasoning overhead that agentic systems generate, rather than standalone accelerator boxes.
For UK buyers, this reinforces a point already visible in NVIDIA's own enterprise reference designs: inference performance, large VRAM and high-bandwidth interconnects are becoming the real differentiators for agentic workloads, not headline GPU counts alone. Anyone specifying infrastructure for 2027-2028 should size your AI GPU requirements against inference and orchestration load, not training benchmarks.
On-prem still has a role — just a narrower one
Hyperscaler GPU supply expanding this dramatically doesn't make on-premises or colocation obsolete. Vendors are actively positioning smaller on-prem platforms for agentic AI: HPE's ProLiant DL394 Gen12, for example, lists agentic AI sandbox workloads as a target use case, aimed at financial services and reinforcement-learning environments where latency, data residency or regulatory control outweigh raw scale.
The practical decision for most UK organisations is now a segmentation exercise rather than a binary choice. Regulated or latency-sensitive workloads may still justify dedicated capacity — worth reviewing via our guide to the on-premise, colocation, or cloud debate for UK compute — while burst, experimental and large-scale inference workloads are the natural fit for the expanded AWS-NVIDIA capacity.
Government and regulated-sector implications
A notable strand of the announcement is the plan to build AI factories for the U.S. government, including 100,000 GPUs on secure AWS infrastructure for workloads classified at Impact Level 6 and above. This matters to UK public-sector and defence-adjacent buyers because it demonstrates that hyperscaler infrastructure — built on the AWS Nitro System and Elastic Fabric Adapter — is now considered viable for the most sensitive national-security workloads, not just commercial AI.
For UK procurement teams evaluating sovereign or high-assurance AI capacity, this sets a reference point for what integrated hyperscaler security and reliability commitments now look like, even where UK-specific frameworks and data residency rules will still apply separately.
What UK infrastructure buyers should do now
The scale of this expansion — moving from a 2026 pledge of over 1 million GPUs to an additional 2 million committed for 2027-2028 — suggests GPU supply constraints that have shaped procurement decisions since 2023 may ease at the hyperscale tier faster than many roadmaps assume. That has direct implications for build-versus-buy decisions and financing structures.
Before committing capital to new on-prem GPU estates, it's worth revisiting financing models: understand financing options for your AI build to see whether GPUaaS or hybrid commitments now offer better flexibility than a 2026-27 capex cycle. Teams still planning dedicated infrastructure should also browse our range of GPU accelerators and review server configuration options built around the same Blackwell-generation platforms AWS is now deploying at scale.
- 01NVIDIA Newsroom — AWS and NVIDIA to Deliver 2 Million Additional GPUs and Next-Generation Infrastructure for Agentic and Physical AI · 26 August 2026
- 02Data Center Dynamics — AWS to deploy more than one million NVIDIA GPUs over next 12 months · 1 July 2026
- 03NVIDIA — AI Factory Reference Design for Government White Paper · 1 January 2026
- 04HPE — ProLiant Compute DL394 Gen12 · 1 January 2026