Fusion research and AI data centres collide over the same electrons

On 29 September, CBC News reported that companies looking to build AI data centres have approached the Newfoundland and Labrador government. Energy and Mines Minister Lloyd Parrott said during a mid-September debate on the Churchill Falls agreement that "our door is open for business." The province would not name the companies, citing commercially sensitive talks.

The timing matters for anyone tracking fusion research money. The same dossier that carries the CBC report carries two research items, both posted 29 September, and they pull in opposite directions on how much electricity a decarbonised world actually needs.

Hannah Ritchie's Substack post, published 29 September, walks through numbers attributed to Oxford professor Nick Eyre. In that model, global final energy demand falls from 416 exajoules today to 247 EJ in a post-transition system, while electricity demand rises from 110 EJ to 189 EJ. Conversion efficiency at the point of use explains the gap. A petrol car turns roughly 20% of its energy into motion. An electric car manages around 80%. Ritchie is explicit about the limits. The model assumes no efficiency gains beyond electrification and a shift to hydrogen, assumes all non-electrified sectors run on hydrogen, and does not account for energy growth as countries develop. She calls it "a fairly simplistic model" and says it likely underestimates the reduction in demand.

The counterweight comes from an NBER working paper dated September 2026, by Matthew Schwartz, Isaiah Andrews and Jesse M. Shapiro. The authors ran an open-source LLM workflow across 4,452 published replication packages from five economics journals. It flagged discrepancies in 3,460 articles or their appendices. In 496 articles it cut computation time by more than a factor of 10, and in 923 cases it produced an extension not present in the original paper.

That is a result about research automation, not energy. But it lands in the same week as a funding announcement built entirely on the premise that compute efficiency is the binding constraint. Efficient Computer said on 29 September that it raised a $97M Series B led by TQ Ventures, taking its total raise to $173 million at a $650 million valuation. CEO Brandon Lucia writes that the company's Electron E1 chip, now launching at volume, claims 10-100x energy-efficiency improvement over traditional CPU architectures.

Lucia's argument leans on Amdahl's Law. If only 90% of a computation fits a specialised accelerator, the maximum possible gain is 10x, because the remaining 10% stays inefficient. Anyone who has watched AI hardware startups over the past three years has heard the pitch before, and the valuation is the company's own claim, not an audited figure.

Oracle, meanwhile, used the same day to announce Oracle Fusion Claw, a governed agentic execution runtime, along with 25 new Claw-powered applications. The press release quotes CEO Mike Sicilia saying the product moves customers "from AI assistance to execution." Oracle says the runtime runs on Oracle Cloud Infrastructure and is powered by frontier models including Gemini and OpenAI.

None of this is fusion. That is the point. The dossier's fusion-adjacent material is a set of recent headlines, not full text, and the working sources available here describe the demand side of the electricity equation rather than any new reactor result. The honest read is that the newest, best-documented developments on 29 September were about who gets the electrons and how efficiently they are used, not about a breakthrough in plasma confinement.

Stephen Wolfram's essay, dated 28 September, makes a related argument about AI and research fields generally. He writes that modern AI is "first and foremost, a way of leveraging the existing corpus of human knowledge," while computation is "an open-ended way to generate things that are fundamentally and irreducibly new." Mathematician Dan Romik, in a post published 29 September, puts it more sharply. AI models generate theorems "distance one away" from their training corpus, he writes, but cannot reflect on, simplify or generalise that output, so an autonomous loop "will grind to a halt shortly after finishing the distance one theorems."

CBC's report notes the same tension on the ground in Newfoundland and Labrador. TechNL CEO Andrea King told CBC that Labrador meets the criteria data centre proponents want, namely low-carbon hydro power, cold weather and land, but that this does not settle the question. "You have to look at economic costs and opportunity costs of what else you could do with that electricity," she said. Tech journalist Paris Marx, author of Hyperscale, made a similar point to CBC, asking whether it makes sense to hand that electricity to data centres when better uses might exist. The province has not named any proponent or project.