A BC law professor made the rounds this month with an argument that sounds reasonable on its surface. Small AI models can now run locally on a laptop with the wifi switched off, and one recent release from June handles text, images, and audio at roughly the capability paid cloud services were charging for two years ago. His conclusion was that some of Alberta’s gigawatt-scale data centre buildout could end up overbuilt, maybe even outdated, before the doors on some of these facilities open.
He’s Not Wrong About the Trend
Give credit where it’s due. Local and edge inference is real, it’s growing, and running a capable model on consumer hardware without a cloud connection was science fiction a few years ago. The professor himself acknowledges the limits too, small models still fall apart on multi-step reasoning and anything beyond fairly trivial coding tasks. Training frontier models, and running inference on the biggest ones, still needs large data halls with serious power behind them. That part of his argument isn’t in dispute.
Where the argument runs into trouble is the leap from “smaller models are getting good” to “so Alberta’s buildout will end up stranded.” That leap has been made before, and it was wrong the last time too.
This Isn’t the First Time
In January 2025, a cheap model release out of China wiped billions off tech stocks in a single day. Commentators lined up to call the data centre buildout finished, a wasted bet on infrastructure nobody would need. Compute demand kept climbing anyway. The cheaper model didn’t replace the need for large-scale training and inference capacity, it just added another workload to a demand curve that was already pointed up. Cheaper computing, as far as the historical record goes, has never once meant less total computing. It’s meant more people building more things on top of it.
That pattern isn’t unique to AI either. Every time computing got cheaper per unit, going back decades, the response wasn’t less demand for computing capacity, it was new uses nobody had bothered building because the older price point made them impractical. There’s no obvious reason this round breaks that pattern.
Alberta’s Actual Hedge
What actually protects Alberta here isn’t betting against edge computing, it’s the physical advantages baked into the province’s approach to hosting these facilities in the first place. Abundant natural gas means on-site generation is a realistic option rather than a multi-year wait for grid interconnection. Cold winter air cuts cooling costs in a way warmer jurisdictions can’t match. Flat, available land keeps construction straightforward. And Alberta’s regulatory framework already requires large facilities to bring their own generation to the table, which means the infrastructure isn’t purely a bet on one narrow category of AI workload.
If the tenant mix shifts, if more inference moves to the edge and data halls end up doing more training and less everyday chatbot traffic, the physical infrastructure doesn’t become worthless. Power generation, cooling capacity, and land don’t care what kind of AI workload is running through them. That’s a very different risk profile than a facility built purely around a single software trend.
Count What’s Actually Running
The more useful number to track isn’t how many data centres get announced in a press release. It’s how many are actually energized and drawing power, because those are the only ones paying anything into the local economy or the grid. Announcements are cheap. Energized megawatts are the number that tells you whether Alberta’s bet is paying off.
Should Alberta start publishing energized megawatt counts for approved data centres instead of just approval totals?
