What happened
A new report highlights the growing strain on the US power grid from AI data centres. According to Ars Technica AI, the massive energy consumption required for training and inference is beginning to drive up electricity prices for all industrial users. This surge is creating new competition for energy resources across the country. The core issue is that demand from new data centres is outstripping the pace of new energy generation, putting pressure on existing infrastructure.
How the room's reading it
Infra teams on X are flagging this as a major, unpriced risk for scaling AI products. The consensus is that 'cheap' compute might not stay cheap if the underlying energy costs spike. Energy analysts see a gold rush for power generation, with major tech players now striking long-term deals directly with utility companies and even exploring small modular reactors. Some developers believe this pressure will force the industry to prioritise model and hardware efficiency — a move away from simply building bigger models. The debate is split between finding more power and using the power we have more wisely.
Sailfish's take
We think this is more than just an infrastructure headline — it's a direct threat to the unit economics of many AI startups. For years, builders could treat electricity as a cheap, abstracted-away part of their cloud bill. Those days are ending. We're advising our teams to start modelling energy consumption as a core operational cost, especially for inference-heavy applications. This isn't about saving the planet, it's about saving your margins. The smartest teams we see are already optimising for model efficiency, not just raw performance. If your scaling plan assumes energy costs will stay flat, it's time for a rewrite.