Wright's Law applied to 150 technologies, including the first systematic experience curves for carbon removal and AI compute. The headline numbers, where they came from, and what they mean for climate and capital.
Costs for most things fall by about 20% every time we double how much we have made. That is Wright's Law, and it has been true for a hundred years across 150 technologies. This paper updates the dataset to 2024 and adds the two domains everyone is currently arguing about — carbon removal and AI compute.
The headline number is GPU compute: 89.1% cost reduction per doubling, the highest learning rate on record. Nuclear, by contrast, has gotten more expensive with every reactor built — an anti-learning rate of −88.1%. Carbon removal is still early, but the better-positioned technologies are learning at 5–16%, comparable to where solar PV sat in the 1980s.
The practical takeaway: before you forecast cost decline for any new technology, ask which side of the modular-versus-bespoke line it sits on. That matters more than the technology itself.
This is broadly consistent with Nagy et al. (2013), the most-cited prior validation, which found a median in the 15–20% range across 62 technologies. The slight upward shift comes from updating fast-learning digital technologies and adding AI compute.
What it tells you: when someone forecasts the cost of a new technology, the prior should be a learning rate near 20%. The interesting question is always which direction the technology deviates from that prior — and why.
Three things compounded to produce this: transistor scaling (the classic Moore's Law mechanism), architectural parallelism (GPU core counts grew from ~100 to over 16,000), and yield learning at fabs. None of these would have produced 89.1% on its own. Stacked, they did.
The implication for AI economics: LLM inference costs are now declining at roughly 69% per year, about 3× faster than Moore's Law alone. Frontier inference reaches cents per million tokens by 2028 if current rates hold. The marginal cost of intelligence approaches zero well before any physical limit kicks in.
The DAC split is the interesting result. Liquid solvent (Carbon Engineering-style) is approaching solar-PV-grade cost decline. Solid sorbent (Climeworks-style) costs initially rose as plants scaled from 50 tCO&sub2;/yr pilots to the 36,000 tCO&sub2;/yr Mammoth facility — first-of-a-kind plants typically need a manufacturing revolution before learning takes hold.
If DAC liquid solvent's 16.4% rate persists for three to four doublings, costs reach the $50–70/tCO&sub2; policy-relevant zone before 2035 with aggressive deployment. That is a conditional projection — CDR data is thin (3–8 points per technology) and forecast error grows about 2.5% per year.
Why this happens: nuclear is bespoke and site-specific. Each plant is essentially a custom build, with escalating regulatory requirements, shifting safety standards, and institutional knowledge that has to be relearned every cycle because the gaps between projects are too long. The factory-style learning loop never closes.
The lesson generalises beyond nuclear. Modularity and manufacturability are prerequisites for strong learning. A technology can be physically promising and still fail to follow Wright's Law if it cannot be standardised, repeated, and improved on a short cadence. This is the test we should apply to every "promising technology" claim — including in CDR.
Individual experience curves measure cost decline along a single dimension. Φ-s captures what happens when four independent improvement vectors compound. Compute cost down 13×, energy efficiency up 21×, algorithmic efficiency up 342×, deployment reach up 1,000× — over the same twelve years.
The total improvement exceeds Moore's Law over the same period by more than six orders of magnitude. Solar PV gains by more than seven. This is not a single technology getting cheaper. It is a phase transition in the complexity of the symbolic layer.
If you are forecasting cost decline for a deep-tech investment, the prior is a 20% learning rate. The interesting question is what makes your bet faster or slower than that.
If a technology is bespoke and site-specific (nuclear, large hydro), don't assume it will follow Wright's Law. If it's modular and factory-built (solar, batteries, DAC liquid solvent), expect cost decline conditional on deployment.
Use the 7.5% mean CDR learning rate as a baseline. Allocate more to technologies that are already showing 12%+, less to those flatlining. Treat the early estimates as preliminary indicators, not stable parameters.
If frontier inference costs are halving annually, your three-year capacity plan is wrong. Model on a one-year halving, not three. The marginal cost of intelligence is going to zero before physical limits.