Amazon is raising its 2026 infrastructure spending to about $220bn, yet AWS expects AI capacity constraints to continue through 2026 and into 2027. The useful point is not the giant budget alone: building the physical machinery for AI takes so long that even extraordinary spending cannot produce instant capacity.
AWS AI Watch analysis
What happened
The International Business Times Singapore reports that Amazon chief executive Andy Jassy linked the higher spending largely to rising memory costs and said demand already visible for 2028 was striking. AWS revenue reached $42.2bn in the latest quarter, up 37% year on year, but customers are still encountering shortages of high-demand GPU capacity.
AWS has introduced instance preference lists for SageMaker AI, allowing customers to nominate several acceptable instance types rather than waiting for one specific option. The change is a practical response to scarcity. AWS says some customers had been writing their own retry systems and submitting competing jobs across different instance types simply to find available capacity.
The same report says Amazon is spending not only on chips, but also on land, power contracts, networking and data-centre construction. Jassy said a data centre can take about two years to build and then operate for more than 30 years. Amazon is also pushing its own Trainium silicon, with AWS pricing cited in the report showing one particular Trainium1 configuration at $0.596 per accelerator hour, compared with $1.475 for a specified A100-based configuration. Those figures apply to particular workloads and do not prove that Trainium is universally cheaper.
Why it matters
AI infrastructure is becoming a bottleneck measured in concrete, cables and power connections, not merely in the number of GPUs ordered. Amazon can increase the budget today, but the resulting capacity still has to be designed, built, connected and supplied with equipment before customers can use it.
For organisations training or running models, availability may matter as much as headline price. Flexible instance selection can reduce waiting, but it may also mean adapting workloads to different hardware and performance characteristics. The cloud is still elastic, in the sense that a very large rubber band is elastic just before it snaps.
Our read
This is a meaningful AWS story because the company’s reported $220bn infrastructure plan sits alongside an admission that demand will remain ahead of supply. It suggests the AI build-out is not simply a race to buy more accelerators. It is a long industrial project involving memory, networking, electricity, construction and custom chips.
The most useful detail for customers is the instance-preference feature, not the spending headline. It offers a way around some GPU queues now, while Amazon works on capacity that may take years to arrive. The outstanding question is whether the investment will create durable, profitable capacity or an enormous motorway built just as traffic patterns change.
What to watch
- Whether AWS capacity constraints ease during 2027 as new facilities and equipment come online.
- How widely customers adopt flexible instance selection, and whether it changes performance or costs.
- Whether Trainium supply and software support can reduce dependence on scarce high-demand GPUs.
- Whether Amazon discloses how much of the $220bn target is allocated specifically to AI infrastructure. Sources and evidence: The International Business Times Singapore account published on 25 September is the basis for the spending, demand, capacity, pricing and AWS feature claims. The figures and statements are attributed to Amazon, AWS and Andy Jassy where indicated by that report.
Discussion spark: Should cloud providers prioritise flexible access across cheaper custom chips, or should customers be able to demand the exact accelerators their workloads were designed for?
Sources and evidence
not affiliated with or endorsed by Amazon Web Services (AWS)