AI Data Centres Meet Canada's Hydro: Newfoundland Opens the Door to Proposals

On 29 September, CBC News reported that companies have approached the Newfoundland and Labrador government about developing AI data centres. Energy and Mines Minister Lloyd Parrott told the House of Assembly, during a special session on the new Churchill Falls agreement, that "our door is open for business." The province's Department of Energy and Mines confirmed the approaches but declined to name proponents or projects, citing commercially sensitive discussions.

The proposal matters for two reasons. AI data centres have become one of the largest new sources of electricity demand in North America. And the province's main advantage is the same one that makes fusion research attractive: a large, low-carbon supply of power that is already built or already planned. TechNL CEO Andrea King told CBC that Labrador meets several criteria data centre proponents look for, including low-carbon hydro electricity, cold weather and a lot of land. She also said that does not make a project a good idea by itself. The province has to weigh the economic costs and the opportunity costs of what else it could do with that electricity.

That caution is the analytical core of the story. A data centre is a facility that houses and powers computer servers, networking equipment and other technology, and demand for that computing has spiked with the AI industry. CBC quotes tech journalist Paris Marx, author of Hyperscale, saying the products are computationally intensive and that tech companies found they needed a lot more computation to power them. Marx also questioned the long-term value of a project, asking whether it makes sense to give electricity to data centres when there might be better uses.

The province has not published a framework, a list of conditions or a named proponent. Nova Scotia Premier Tim Houston has outlined five starting conditions any data centre proponent must meet for consideration, CBC reported, and EverWind Fuels has expressed interest in building a data centre in that province. In Alberta, the Meta data centre planned for Sturgeon County is expected to cost $13 billion and come online in two to three years, according to the same CBC report, which also notes the project has generated division in the province.

No fusion result appears in the dossier. That is worth stating plainly. The recent headlines circulating around fusion, including UK commitments and a Chinese hydrogen-boron device, are not part of the source material supplied for this analysis and are not cited here. What the dossier does contain is a set of documents about how computation, energy and research are converging. Those documents are more useful for understanding why a province with hydropower would receive data centre proposals than any single laboratory announcement.

A 29 September post by Hannah Ritchie on her Substack, By the Numbers, works through numbers from Oxford professor Nick Eyre on what electrification does to final energy demand. The post states that electricity demand rises from 110 to 189 exajoules in the post-transition scenario, while total final energy demand falls from 416 to 247 exajoules. That is a roughly 40% reduction, achieved without assuming any efficiency gain other than electrification or a move to hydrogen.

Ritchie's post is explicit about the simplifications. All non-electrified sectors are assumed to be powered by hydrogen. The model does not account for energy growth as countries develop, though she argues that should not affect the ratio between the two scenarios. It is a comparison of final, not primary, energy, so it excludes the waste heat that primary energy accounting would include. She also says the estimate likely understates the reduction because it assumes no other efficiency measures.

The sector detail is where the argument becomes concrete. Road transport and buildings drop significantly, because cars and vans can be electrified and gas boilers can be swapped for heat pumps. Ritchie writes that an electric vehicle converts around 80% of its energy to motion, against around 20% for a petrol car, and that post-transition energy demand for cars and vans is about one-quarter of current demand. High-temperature industrial processes are harder. Some can move to hydrogen, and the steel sector is assumed to raise electric arc furnace use from around 25% to 50%, the current OECD mix.

For a province weighing a data centre, the relevant lesson is not the exact exajoule count. It is that electricity demand is expected to rise even as total energy demand falls, which means competition for clean electricity becomes sharper, not looser. The same post notes that electricity supplies three-quarters of final energy demand in the post-transition system, against one-quarter today. Every additional large load is a claim on that electricity.

The counterargument to building data centres on scarce clean power is being made with hardware. On 29 September, Efficient Computer announced a $97 million Series B round led by TQ Ventures, bringing its total raise to $173 million at a $650 million valuation, according to the company's own blog post. CEO and co-founder Brandon Lucia wrote that the company's technology brings 10-100x energy-efficiency improvement compared to traditional CPU architectures, and that the Electron E1 is launching at volume for embedded physical AI systems including robots, drones, infrastructure and wearables.

Lucia's post is a vendor argument and should be read as one, but the architectural point it makes is checkable in principle. He invokes Amdahl's Law to argue that the unaccelerated part of a computation caps the overall efficiency of a system, and that narrowly specialized fixed-function AI accelerators create obsolescence risk when algorithms change. Whether or not Efficient's numbers hold outside its own benchmarks, the claim that efficiency, not raw supply, determines how much computation a given amount of electricity can buy is the same claim that makes data centre proposals politically contested.

Two of the newest items in the dossier describe AI systems that consume compute to do research itself. On 29 September, CoreWeave announced ARIA, a coding agent built into Weights & Biases that reads experiments, builds live visualizations, forms hypotheses and launches runs through W&B Launch. The company's blog post, originally published on the Weights & Biases blog on 29 July 2026 according to the page, says ARIA can span projects and access training code, experiment logs, loss curves, metrics, artifacts and checkpoints. CoreWeave says the agent is trained on the platform's nuances so it can scope queries efficiently across projects with 20 or 20,000 runs.

A separate 29 September item, an NBER working paper by Matthew Schwartz, Isaiah Andrews and Jesse M. Shapiro, applies an LLM workflow to published economics research at a scale that implies real compute. The paper's abstract states that across 4,452 published replication packages for five economics journals, the workflow flags discrepancies in 3,460 articles or their appendices. It says that in 496 articles the workflow reduced a calculation's computation time by more than a factor of 10 at similar or greater accuracy, and that in 923 articles it developed an extension not in the original article. The paper discloses that LLMs were used in the analysis and writing, that Schwartz worked as a contractor for Anthropic, and that the results and views are not endorsed by Anthropic.

Those numbers are large enough to be a data centre demand signal on their own, but the dossier does not connect them to any specific facility or province, and this analysis does not either. What the two items show is that the research workflow itself is becoming an electricity consumer. That is precisely the kind of load Newfoundland and Labrador is being asked to host.

Two 29 September posts on mathematics research make the same point from the other direction: AI can automate more of the work, but the work does not disappear. On his blog, Dan Romik argues that mathematics research is a feedback loop, that current AI models generate new theorems that are "distance one away" from human-generated knowledge, and that without humans to reflect, simplify and generalize, the loop stalls under what he calls mountains of self-generated slop. He also quotes Scott Aaronson's reaction to OpenAI's claimed Navier-Stokes solution as a string of A's, and states that he does not believe we are yet in a world where AI puts the final QED on every theorem.

Stephen Wolfram's 28 September essay makes a compatible argument with a different emphasis. He writes that modern AI is useful, sometimes very useful, and that its greatest use in his own mathematical pursuits has been to mine the knowledge base of human mathematics thematically. But he distinguishes AI, which uses the existing corpus of human knowledge, from computation, which he says can generate things that are fundamentally and irreducibly new. He also argues that great mathematics is defined by the questions it asks, and that the human imagination guides which questions get asked.

Neither essay mentions data centres, and neither is a fusion result. They matter to this story because they describe the demand side of the electricity question. If AI systems are to be used for research, for economics replication at the scale of thousands of papers, or for agentic loops that launch their own experiments, then the compute has to sit somewhere. Newfoundland and Labrador is one of the places now being asked whether it should sit there.

The province has not said yes. CBC reported that a department spokesperson said data centres may be considered where proposals align with government policy objectives, including environmental requirements, and deliver clear benefits to Newfoundlanders and Labradorians. No proponent, no capacity figure, no power purchase agreement and no timeline has been made public. Until one is, the honest description of the state of play is that a door has been described as open, and nothing has walked through it yet.