Your newsletter digest — May 28, 2026

AI infrastructure

1. Space racks are more feasible than they look. A single Starlink-sized satellite already has the rough physical dimensions of an NVL72 GPU rack. Power and heat dissipation remain the engineering hurdles, but neither is categorically unsolvable.
2. The right workload is agentic inference, not training. Agents running overnight jobs without a human in the loop don't need low latency — they need cheap, abundant compute. Slower chips in orbit would be tolerable; the economics could work.
3. Zoning is the stick. Thompson notes that communities now have veto power over terrestrial data centers in a way they never had over factory closures. As opposition grows, the cost advantage of earth-based compute shrinks — creating a genuine opening for the alternative.

1. Agentic workloads tolerate latency. When an agent runs a multi-hour task with no human waiting, latency stops being the constraint. What matters is capacity — large context, persistent state, cheap DRAM — not peak token speed.
2. AWS's decade of "inferior" chip investment is paying off. Trainium was mediocre for years. It's now decent, and the trajectory is favorable. AWS's Bedrock runs on it silently, just as Graviton powers RDS without customers choosing it — the same playbook, replayed for AI.
3. Cerebras has a real but narrow window. Its wafer-scale chip has 6,000× the memory bandwidth of an H100, which makes it fast for answer inference. But the moment a workload needs more memory than fits on-chip — a larger model, a growing KV cache — the economics collapse.

Product and growth

1. Gate intensity, not intelligence. Rather than paywalling smarter models, gate volume. Google's AI tiers give each level higher usage limits and larger context windows — prepaid pricing that maps costs to actual compute usage.
2. Sell outcomes, not answers. Intercom Fin charges per resolved support ticket, not per API call. Google's autonomous browsing agent that completes web tasks is paywalled to higher tiers — customers are buying eliminated labor, not model access.
3. Reserve the compute-heavy modalities. Cinematic video generation, real-time simulation, persistent 3D environments — these naturally belong behind a premium wall. Consumers already expect them to cost more; the paywall doesn't feel punitive.

1. Communication first. Nobody Wants to Read Your Sht*, On Writing Well, and Storyworthy ranked ahead of any strategy or PM frameworks. The argument: as AI lowers the cost of building, distribution becomes the scarce resource — and distribution starts with writing.
2. Strategy books that actually teach strategy. Good Strategy/Bad Strategy and Playing to Win made the list; most other strategy books didn't. Rumelt's framework for diagnosing bad strategy (fluff, failure to face the problem, mistaking goals for strategy) remains the clearest test.
3. Management classics over management fads. High Output Management (1983), Radical Candor, The Making of a Manager — nothing on AI management or remote work hacks. The message: timeless > timely.

One thread to watch