So there’s interesting stuff in here about neocloud AI pressures (writes Juan Pedro), industrial AI dynamics (me), suburban AI planning (says Susana), and international AI deliveries (Sulagna) – plus more, below the line, about challenges to make modern AI software work on old old systems (says IBM and Arm), make private 5G less complex and more scalable (XCOM RAN), and make LLMs work on laptops and local servers (Google). Plus-plus, copper thefts are a big deal (warns AT&T), and Wi-Fi is undergoing a renaissance (suggests Sean, at RCR). But, back from Easter, it is maybe easier to join the dots in the mainstream press, about the mainstream story: that the biggest constraint on AI is no longer models or money, but infrastructure physics. 

We know this already, of course, but there are a couple of pinchpoints here that ram it home (report Reuters and Bloomberg). Investors are pressing the likes of Amazon, Microsoft, and Google over their water and power usage, while the industry is running into bottlenecks in the most unglamorous layer of all: transformers, switchgear, and grid connections. In case there was any doubt: the story has vaulted from how quickly AI developers can build to how quickly they can find power – to whether they even should. The supply of electricity infrastructure is entrenched and expensive; the demand for AI is global and speculative. That is a killer mismatch. And there is probably only one answer: money. Glib, but true. Or maybe two: money and more tech.

Which loops back to everything else above. A lot of the innovation now looks like adaptation: how to do more with less. Less power, less complexity, less waiting around. There’s plenty of money in the system, as yet. But the grid, not the GPU, is dictating the pace.