Amazon Plans $220 Billion AI Capital Expenditure by 2026, AWS Sold Out Until 2028
Amazon plans a capital expenditure of approximately $220 billion by 2026, mainly for AI infrastructure. CEO Andy Jassy stated that even with this investment, it will not meet all demand.
Jassy mentioned in the a16z podcast that AWS capacity will be sold out by 2026, largely sold out by 2027, and a significant portion will also be reserved for 2028, with the company only retaining some on-demand, Bedrock, and EC2 capacity.
He pointed out that AI business annualized revenue will exceed $25 billion within three years, growing much faster than early AWS, and that "investing more capital would not be a bad use".
Navin Chaddha previously analyzed that the current training infrastructure is not fully built, and inference workloads will be ten times that of training in the future, but currently account for less than 10% of usage, leading to continued growth in hardware capital expenditure.
He warned that if the speed of AI adoption does not keep pace with the initial training infrastructure build-out, there will be a significant market correction in the semiconductor, hardware, and model sectors; currently, only search and coding have substantial traction, while other areas are still in the early stages.
Cloud vendors and AI labs are "buying" long-term computing capacity, benefiting chip and data center suppliers from locked-in orders, but if inference adoption lags, hardware suppliers and model companies will face pressure.
Source: Public Information
ABAB AI Insight
Amazon previously raised its 2026 capital expenditure to approximately $220 billion, with AWS second-quarter revenue growing 37% and backlog orders reaching $496 billion. Jassy has repeatedly emphasized that even with large-scale investments, unmet demand remains, and has pushed for the self-developed Trainium chip to reduce inference costs.
In terms of capital flow, funds are mainly directed towards data centers, GPU and custom chip procurement, and power facilities, motivated by locking in long-term computing needs of AI labs and enterprise customers, moving AWS towards a trillion-dollar annual revenue target while pushing free cash flow into negative territory.
This is similar to previous cloud infrastructure expansions and semiconductor cycle paths, currently in the early stage of training infrastructure build-out, with inference scale yet to fully explode; coding applications (like Cursor, Claude Code) have shown scale, while other verticals are still at the starting point.
Essentially, it is a restructuring of the industry chain: computing power is shifting from on-demand consumption to multi-year reservations, forcing capital concentration on cloud vendors and chip supply chains. If inference adoption does not keep pace with training investment rhythm, it will trigger a redistribution of pricing power at the hardware and model layers.
ABAB News · Cognitive Laws
Training builds the road, inference runs the car; if there aren't enough cars, the road is empty.
Computing power reservations lock in future demand rather than current usage.
If adoption speed does not keep up with capital expenditure, hardware corrections will occur.