FORGE/DAILY — April 6, 2026
Reuters Breakingviews put a number on it this week: $7 trillion in projected AI infrastructure spending over the next several years, against returns that remain — to put it charitably — speculative. The piece is worth reading not because it's bearish for the sake of it, but because it catalogs the specific resource constraints that don't get enough coverage: power grid capacity, water for cooling, semiconductor supply chains, the physical bottlenecks that make "we'll just build more" a harder answer than it sounds.
The honest version of this story is that the AI infrastructure bet is rational if you believe in the long-term productivity gains — and there are real reasons to believe that. But the timeline mismatch between capex and returns is real. Microsoft, Google, and Amazon are collectively spending at a pace that assumes the revenue catches up. So far it mostly has. The question the Reuters piece is implicitly asking is: what's the failure mode if it doesn't? The answer involves a lot of stranded assets and a very bad few years for cloud infrastructure stocks. Not a prediction, just a risk worth pricing in.
- 📌OpenAI, Anthropic, Google share intelligence through Frontier Model Forum to detect Chinese distillation attacks — Rare inter-lab cooperation. The threat model: Chinese labs distilling frontier capabilities from open-weight models and API access. The three labs are sharing detection signals. This is the most significant safety-adjacent cooperation between labs we've seen.
- 📌South Korea deploying ChatGPT-powered companion dolls to elderly population — 20% of South Korea's population is now elderly. The companion AI rollout is less dystopian than it sounds and more pragmatic than most Western coverage acknowledges. Loneliness is a public health crisis. These help. The ethics are real but the alternative isn't nothing — it's also loneliness.
- 📌China flies world's first megawatt-class hydrogen turboprop — 16-minute flight, proof of concept cleared. This is infrastructure-adjacent AI news: the energy transition and AI infrastructure are increasingly the same story.
- 📌AI legal startups targeting law students for market entry — The strategy: hook future lawyers before they form habits. LexisNexis and Westlaw built their moats through law school. The new AI legal tools are playing the same game.
- 🔥Microsoft MAI models now available in Azure Foundry — Multilingual benchmark leaders, enterprise pricing. The quiet story is Microsoft shipping competitive models outside of the OpenAI relationship.
The $7 trillion figure is doing a lot of work in a lot of headlines right now. What it actually measures is announced and projected capex — not spent, not committed, not guaranteed. The real number is probably half that, deployed over a longer window than the breathless coverage implies. But even at half that figure over a longer timeline, the resource constraints Reuters is describing are real. The grid problem alone is a 10-year problem, not a software update.
Frontier Model Forum Intelligence Reports — The inter-lab cooperation body now publishes threat intelligence summaries. Dry reading, genuinely useful if you're building anything security-sensitive on top of frontier models and want to know what the labs themselves are worried about.
Get the next issue
AI news for people who don't need it explained twice. Free.