Key Takeaways
- Microsoft, Google, Meta, and Amazon are collectively spending over $300 billion on data centre infrastructure in 2025-2026.
- A single next-generation AI training cluster consumes 50-100 MW of power comparable to a small town.
- Grid interconnection queues in the U.S. now exceed 2,000 GW of requested capacity, with multi-year wait times.
- Natural gas-fired peakers and combined-cycle plants are the fastest way to add large-scale power capacity 18-24 month build times.
- Canadian gas producers with pipeline access to U.S. markets are experiencing structurally higher demand and price signals.
The capital expenditure plans announced by the four largest U.S. hyperscalers Microsoft, Google, Meta, and Amazon for 2025 and 2026 total over $300 billion. The majority of this spend is on data centre infrastructure: servers, networking, cooling systems, and critically power infrastructure. At this scale, the energy requirements are not a footnote; they are a primary constraint on AI development.
The Power Problem
A modern AI training cluster the type of facility used to train the frontier large language models that have driven the AI boom requires 50-100 megawatts of continuous power. A major hyperscale campus may host dozens of such clusters. The power requirements are continuous, not intermittent: AI training and inference workloads run 24 hours a day, 7 days a week. Solar and wind generation the cheapest sources of new electricity capacity cannot reliably serve this load profile without enormous battery storage capacity that does not yet exist at the required scale.
Grid Interconnection: The Hidden Bottleneck
The power grid itself is the binding constraint on U.S. data centre buildout. Grid interconnection queues the waiting list to connect new power generation or large industrial loads to the grid have exploded to over 2,000 gigawatts of requested capacity. The wait time to get a new large industrial load connected to the grid in high-demand regions has extended to 5-7 years in some cases.
Natural gas-fired generation is the fastest solution available. A gas-fired combined-cycle plant can be permitted, built, and commissioned in 18-24 months versus 5-7 years for a new nuclear reactor and 3-5 years for wind or solar with storage. For AI companies that need power now, gas is the only viable option.
“Every data centre CEO I speak to is focused on power. They are not negotiating with us on price they are negotiating on timing. Can you connect us in 18 months? 24 months? That is the conversation.” Senior executive at a major U.S. utility, 2026
The Canadian Natural Gas Opportunity
Canada’s natural gas producers are well-positioned to benefit from this structural demand shift. Several large data centre campuses are being developed in proximity to Canadian pipeline systems including facilities in Washington State, Montana, and the Dakotas that draw gas from the NOVA and Alliance pipeline systems. Canadian gas is competitive on price and reliable in delivery.
The structural implications for AECO pricing are positive. As U.S. gas demand grows structurally, the incentive to pull incremental Canadian gas south increases. AECO prices have historically been depressed by regional oversupply and pipeline constraints both of which are easing as LNG Canada diverts volumes and U.S. demand absorbs more Canadian production.