Nobody building a five-year AI strategy today can tell you with confidence which models will exist in 2029, what frontier capabilities will cost, or how capable open-weight alternatives will be. The landscape is too unstable, the economics too volatile, the competitive dynamics too unclear. That's a genuine planning problem with no clean answer.
There is, however, one thing you can be confident about: "democratization of AI" - the phrase in every major lab's mission statement - is not a safe planning assumption. Betting your strategy on vendor promises of broad, affordable access isn't just optimistic. It's structurally naive.
The companies making these promises have no economic or strategic incentive to fulfill them. When a model becomes capable enough to matter, restricting access becomes commercially essential. Preferred enterprise tiers, government partnerships, export controls - these aren't bugs in the democratization narrative. They're the product. Safety is the framing. Control is the mechanism. The labs that coined "democratizing AI" are now leading advocates for the frameworks that limit exactly that.
This dynamic won't stabilize. As capabilities increase, the strategic value of restricting access increases with them. What's happening now is not a transitional phase. It's the direction. The most concrete adjustment available to planners right now: assume AI costs more in three years than it does today, not less, and build that into long-term forecasts. "It'll get cheaper" is a vendor narrative. "It'll get more controlled" is what the structural incentives actually predict.
Thoughts? Find me on Bluesky.