In 1926, the smartest forecasters alive were predicting the future. They called radio empires, air travel, skyscraper cities. They missed the computer entirely. Not “underestimated it.” Missed it. The technology that would define the century wasn’t on the list.
Keep that failure in mind, because I’m about to make the same kind of forecast, and one of these paragraphs is probably wrong in a way neither of us can see.
The method
Here’s the approach I trust most, having just spent time mapping how 54 domains of life changed over the past thousand years: at short horizons, bet on institutional inertia, because institutions barely move in a decade. At long horizons, bet on compounding cost curves, because a cost curve that holds for 40 years survives every prediction of its death, and a century is 15 to 25 doublings. Most bad forecasting comes from mixing up the horizons: expecting institutions to transform in 5 years (they won’t) or expecting cost curves to stall over 50 (they usually don’t).
Right now three curves are compounding, and between them they decide most of what the next century looks like.
Energy
Solar module prices have fallen roughly 99.6 percent since 1976, and the learning rate (about 20 percent cost decline for every doubling of installed capacity) has held for four decades through every predicted plateau. Lithium battery packs cost about $1,400 per kilowatt-hour in 2010; BloombergNEF’s 2024 survey put them at $115, a decline of over 90 percent. In most of the world, solar plus storage is already the cheapest new electricity in history.
Follow the curve out and energy stops being the thing that constrains what humanity can do; the constraints move to grids, land, materials, and permitting. Those are real, but they are speed bumps, not walls. This is the safest bet of the three because it’s pure physics and manufacturing scale, and both are well understood.
Intelligence
The cost of running an AI model at a given capability level is collapsing faster than any technology cost in recorded history. Stanford’s AI Index measured the price of GPT-3.5-level performance falling from $20 per million tokens in late 2022 to $0.07 by late 2024, a 280-fold drop in two years. Depending on the task, inference prices have been falling anywhere from 9-fold to 900-fold per year.
Here’s the part people miss: this bet does not require AI to keep getting smarter. Even if capability froze at today’s level, two decades of deployment alone would absorb an enormous share of coordination work: scheduling, matching, screening, paperwork, reporting, the entire connective tissue of organizations. Diffusion is slow, boring, and nearly certain. That’s what makes it a curve you can build on rather than a hype cycle you have to time.
Biology
Sequencing a human genome cost about $100 million in 2001. It costs a few hundred dollars now, a decline that outran Moore’s law. The first CRISPR gene-editing therapy was approved in 2023. GLP-1 drugs demonstrated that a single molecule can move average body weight by around 15 percent and, within a few years, reach one in eight American adults. The direction is unmistakable: biology is becoming something we engineer rather than something we accept.
And yet this is the curve I trust least, and it’s worth being precise about why. For 60 years, the cost of developing a new drug roughly doubled every nine years, the opposite of a learning curve; researchers call it Eroom’s law, Moore spelled backward, on purpose. Reading DNA got cheap. Changing outcomes in living humans is still slow, expensive, and gated by clinical trials and regulators, which run on calendar time, not curve time. If my 100-year picture disappoints anywhere, it’s here.
Three disciplines
So: energy, very high confidence. Intelligence, high confidence for coordination and knowledge work, with the caveat that the social response is the unpredictable variable. Biology, medium confidence, where the constraint is institutional rather than technical.
First, when someone forecasts dramatic change in 10 years, ask which institution has to move, because institutions don’t move that fast. Second, when someone forecasts stability over 50 years, ask which cost curve has to break, because curves this durable rarely do. Third, hold a permanent line item for the missing curve. The 1926 crowd had radio and aviation on their list and no computer. Something is compounding right now, in a lab or a garage or a spreadsheet, that isn’t on mine.
The next post is about where these curves land first, and it starts with the biggest labor shift in human history, because we’ve run this experiment before.
Sources: solar and learning rates via Our World in Data; battery prices via BloombergNEF (2024); AI inference costs via Stanford AI Index 2025; genome costs via NHGRI; Eroom’s law via Scannell et al., Nature Reviews Drug Discovery (2012); GLP-1 uptake via KFF polling.
