Torsten Slok's note on AI's missing payoff, Venkatesh Rao's Cognitive Mercantilism and Anthropic's S-1, read together, settle an argument: the bound on AI's payoff is physical, and for now intelligence is a magic pudding.
The world is set to spend about $11 trillion on AI data centres between 2024 and 2029. That is so much money, so fast, that it is already pushing up memory prices, electricity demand and interest rates across the whole economy.
AI can grow only as fast as the supply of a few obscure materials such as helium, bromine and rare earths, which come from very few places, take years to scale up, and are at the mercy of wars and trade disputes.
Each AI answer now uses a tiny fraction of the power it once did, yet demand is rising so fast that power plants and grids have become the main brake on AI. Data centres could strengthen the grid instead of straining it if they were required to bring their own clean energy.
Nvidia looks like the king of AI, yet real control sits with the handful of firms that make chips and memory, chiefly TSMC, ASML and the memory makers. Their sold-out factories decide who gets computing power and are making phones and PCs dearer for everyone else.
AI models are the only part of the stack getting cheaper, because open and smaller models catch up to the frontier within months, and that makes it hard for the big closed labs to justify their huge valuations.
A raw AI model only predicts the next word. The real progress is now in the “harness,” the software wrapped around a model that gives it memory, tools and goals, and a good or bad harness can make the same model dramatically better or worse.
The whole AI buildout depends on applications people will pay for. Coding and finance already justify the spending, while AI agents threaten traditional software businesses and are starting to run whole companies for a fraction of the usual cost.
The interesting release this week isn't Alibaba's flagship — it's the flagship's child. If Qwen3.8-27B performs near Opus 4.5 on hardware people already own, the home watershed arrives less than a fortnight after the business watershed. Gradually, then suddenly.
Alphabet's overhaul of DeepMind is the moment Google stopped racing for the frontier and chose to be the landlord instead. A landlord never loses — and never wins. That leaves two American frontier labs, and one question about whether the lead was ever in talent, or only in time.
Apple could own SME agents outright — the hardware exists, the silicon is designed, the distribution is built. All it would take is an infinite supply of 256GB Mac Studios. But every Mac Studio eats the memory of sixteen iPhones, and that is not a close call.
Cognitive work collapsing to near-zero marginal cost looked like a one-off. It unfolds as a series: frontier, business, home, device. Each crossing is gated by the same constraint — memory.
At the recent Homebrew Agents Club, I had a conversation with a friend deep in a startup. "We're pay $100 a month for Xero - it's just not doing the job for us, and we're not getting the support we...