Sustained 10×/yr effective compute requires sustained 2.5×/yr investment growth, plus matching hardware production, plus matching energy supply, plus matching rare earth and water inputs.
At some point one of these breaks. The question: how soon, and how much warning.
| Input | Constraint |
|---|---|
| Capital | Investment can't compound at 2.5×/yr forever absent AI generating proportional economic returns. |
| Energy | Datacenter footprints already pressuring local grids. |
| Land + water | Cooling water; siting datacenters where power is cheap. |
| Rare earths / chips | Supply-chain bottlenecks. |
| Hardware physics | Memory bandwidth + interconnect limits — moving data between chips may dominate compute as models grow. |
"If continued AI progress relies mostly on scaling compute, data, and inputs like energy and computer hardware, then a main question is whether the economic cost of continued scaling (over orders of magnitude) can be sustained and for how long. To complicate matters, the point where increased investments into AI infrastructure and the cost of running large AI systems are no longer economically viable depends on the economic returns produced through AI."
Translation: AI must pay for its own scaling. If AI generates proportionally more economic value, the investment growth sustains itself. If not, scaling hits an economic wall.
The paper points at several recent works modeling whether AI could trigger explosive economic growth that sustains the loop: - Erdil & Besiroglu 2023 - Vollrath 2023 - Erdil et al. 2025 (GATE — integrated assessment model for AI automation) - Davidson et al. 2026 (when does AI R&D automation produce explosive growth) - Whitfill & Wu 2025 (will compute bottlenecks prevent intelligence explosion)
These are the modeling efforts to watch if you want to forecast which scenario plays out.
"Note that if the pathway from AGI to ASI is less dependent on scaling and is driven by algorithmic innovation, self-improvement, or paradigm shifts, then the required economic inputs can be scaled more slowly compared to the gains in capability and economic returns enabled through AI. In this case sustaining AI progress economically may be a marginal friction, whereas in the case where AI progress solely relies on scaling, this factor may become a major bottleneck."
Economics is fatal for 06 - Pathway 1 — Scaling, optional for 07 - Pathway 2 — Paradigm Shifts and 08 - Pathway 3 — Recursive Self-Improvement.
The paper notes: - Energy + land + water consumption - Rare-earth sourcing - Orbital datacenters as proposed escape valve → but introduce ozone-layer risk, re-entry pollution, orbital congestion → not actually a clean fix