Is There a Ceiling on What AI Can Learn, and What Does That Mean for Your Strategy?

What Scaling Laws Actually Say

When Kaplan, McCandlish, and colleagues published their 2020 analysis of how language model performance scales with compute and data, they found a power law relationship.^[1] Doubling compute yields roughly a 5% improvement in model performance. Doubling data yields a similar gain. The relationship is consistent and empirical, not theoretical. If you log both axes, it's a straight line.

This matters because power laws feel predictable. They suggest that scaling is a solved problem: want 10% improvement? Spend 4 times as much compute. Want 20% improvement? Spend 16 times as much.

In 2022, Chinchilla and subsequent scaling law research refined this finding without changing its fundamental message.^[2] The optimal allocation of resources to compute and data might shift, but the diminishing returns remain. You're always exponentially expensive resources to linearly improve performance.

But here's what gets lost in the technical papers: that straight line on a log-log plot masks an intractability problem. As philosopher Toby Ord noted in 2025 analysis of AI scaling, when you translate log-log plots back to linear space, you're looking at exponentially expensive improvements on an increasingly flat curve.^[3] To get meaningful capability gains, you're spending exponentially more.

The precise exponent depends on the metric and task, but the direction is consistent. Kaplan et al. found α ≈ 0.076, meaning performance scales as approximately the 0.076 power of compute. That's a tiny exponent. It means you need compute to grow by 10^(1/0.076), or about 280 times, to get a 10-fold improvement in performance. That's not linear. That's exponential resource consumption for polynomial capability growth.

And this is the measured relationship. The theoretical maximum doesn't get better.

The Efficiency Counterargument (and Why It Doesn't Hold)

When this analysis circulates, there's usually a rebuttal: efficiency is improving. We're training models faster, using less energy per parameter, squeezing more capability out of the same compute.

This is true. And it doesn't change the fundamental problem.

Efficiency improvements are real but they obey their own diminishing returns. The gains from better architectures, better training techniques, and better optimization are real and important. But they're also slowing. The improvements from research in 2020-2022 were large. The improvements from 2023-2025 are smaller, and marginal gains are getting harder to find. You can't escape the underlying scaling law with engineering alone.

Moreover—and this is critical—efficiency gains have been overwhelmed by the appetite for scale. Yes, modern models are trained more efficiently than early transformers. But they're also orders of magnitude larger. OpenAI's language model capabilities have roughly doubled every 18 months over the past five years, but the energy cost per doubling hasn't halved. It's increased.^[4]

Microsoft reported that its corporate carbon emissions increased by approximately 30% since 2020, despite improvements in data center efficiency, because compute scaling outpaced efficiency gains.^[5] Google's greenhouse gas emissions are roughly 50% higher than their 2019 baseline, following a similar pattern.^[5] These aren't small companies with outdated infrastructure. They're the organizations investing most aggressively in efficiency. The efficiency gains exist. But they're being overwhelmed by raw scale.

The efficiency argument, then, is a version of the "we just need a better algorithm" fallacy. Better algorithms exist. But algorithms can't repeal physics. The energy cost of compute is bounded by thermodynamics. The capability returns on compute are bounded by information theory. Engineering can move the boundaries. It can't eliminate them.

Human Scaling: A Different Curve Entirely

Now look at what happens when you scale human problem-solving.

In 2007, Bettencourt, Lobo, and colleagues published research on how various metrics—patents, crime, wealth creation—scale with city size and population.^[6] They found superlinear scaling: when cities double in size, the number of patents doesn't just double, it increases by roughly 27% beyond doubling. GDP per capita grows superlinearly with city size. This is counterintuitive, and it's powerful.

The mechanism is innovation. When you double the number of people, you don't just double the number of ideas. You double the number of interactions, the number of recombinations, the number of serendipitous collisions between different perspectives. The exponent is roughly 1.15-1.27 depending on the metric—higher than 1, which means each additional person adds more than their individual contribution. This is what Woolley's collective intelligence research found: groups are smarter than the sum of their individuals.

This is a completely different curve from AI scaling. AI: exponential resources for polynomial gains. Humans: polynomial inputs for superlinear outputs.

The strategic implication is stark. If you invest $1 trillion in scaling AI compute, you get 5% improvement per doubling. If you invest the same trillion dollars in coordinating the problem-solving capability of humans—better tools, better institutions, better frameworks for collective intelligence—you get 15-27% improvement per doubling of engaged participants.^[6]

Over a decade, the difference compounds. Massively.

The "Densing Law" and the Energy Reality

But there's a more recent finding that complicates even this comparison. Researchers from Microsoft, Stanford, and MIT published analysis in Nature Machine Intelligence (2025) of an emergent pattern they call the "densing law": the rate at which AI capability is doubling is accelerating, with new capability doubling roughly every 3.5 months.^[7]

This sounds like it invalidates the scaling law constraints. It seems to suggest that efficiency is improving faster than the scaling laws predict.

What it actually shows is that different metrics scale at different rates. If you measure raw parameter count, the doubling time is slower. If you measure inference speed, it's different. If you measure capability on specific benchmarks, it's faster still. The densing law is capturing the rate at which AI systems are getting "denser" with capability—squeezing more performance out of each parameter and each unit of compute.

But this doesn't solve the resource constraint. It illuminates it differently.

The densing law accelerating means organizations are locked into a competitive dynamic where they must continually invest to stay at the capability frontier. You can't maintain current capability level without continued investment because the competitive landscape is moving upward faster. It's an arms race where you have to spend exponentially more to stay in place relatively.

Microsoft and Google's emissions data reflects exactly this: companies are deploying more models, more inference, more scaling to stay competitive. The densing law is real. So is the energy cost.

The Strategic Inflection

Here's where the two curves matter for organizational strategy.

Five-year horizon: AI scaling wins. The capability gains from doubling compute still outpace the capability gains from doubling human coordination. If you need a 15% improvement in the next five years, AI scaling is cheaper than human reorganization.

Ten-year horizon: The curves cross. Doubling AI compute is still cheaper per unit capability, but the cumulative cost is rising exponentially. Meanwhile, the return on investing in human coordination—better institutional design, distributed decision-making, cross-functional integration—compounds superlinearly. The organizational structure that scales human intelligence becomes more valuable.

Twenty-year horizon: Human scaling dominates. The organizations that have built institutions for collective intelligence, that can coordinate large numbers of people toward complex problems, that can integrate diverse perspectives rapidly—those organizations have capability advantages that no amount of AI scaling can match, because they're operating on a fundamentally more efficient curve.

This doesn't mean AI is less important on longer horizons. It means the way to extract value from AI changes. Instead of betting on larger models with exponentially higher compute requirements, organizations bet on integration: how do we embed AI as a tool into human decision-making systems that are already operating on the superlinear curve? The AI becomes more valuable when it augments rather than replaces distributed human intelligence.

The Framework: The Scaling Asymmetry Map

To apply this to organizational strategy, map your major capability gaps on two dimensions: time horizon (next 5 years, 5-10 years, 10+ years) and complexity (how much does success depend on integrating diverse perspectives, versus extracting value from a well-defined problem).

For well-defined problems on short horizons, AI scaling is the right bet. Autonomous systems for routine logistics, document processing automation, predictive maintenance—these are domains where scaling AI captures value faster than reorganizing humans.

For complex problems that depend on perspective integration—organizational strategy, product innovation, organizational design—on any horizon beyond five years, human scaling dominates. These are domains where building institutional capacity for collective intelligence outpaces any intelligence gain from model scaling.

Most organizations need both. But the resource allocation is wrong. Most organizations allocate 90% of resources to scaling AI and 10% to improving collective human intelligence. The math suggests it should be closer to 40-60, and the allocation should shift toward human scaling as time horizons extend.

Why This Matters for Women Leaders

This analysis has a particular relevance for women in leadership, because women executives consistently report that their organizations are making strategic AI investments without deep organizational design work to support collective intelligence integration.

In AI-Powered Women's 2025 research, surveyed leaders described a pattern: their organizations hired AI talent, built AI infrastructure, and deployed AI systems—and then realized they hadn't rebuilt the organizational structures, decision-making frameworks, or information flows to actually integrate AI into human decision-making. The AI was technically capable but organizationally stranded.

This isn't an AI problem. It's a human scaling problem. The organization invested in the wrong curve.

Leaders who understand scaling asymmetry—who recognize that AI becomes valuable when embedded in institutions designed for collective intelligence—make different choices. They invest in cross-functional communication. They redesign decision-making processes to integrate rather than centralize. They build institutions where diverse perspectives actually shape outcomes. This is harder and slower than deploying an AI system. It's also more valuable on any horizon beyond five years.

Women leaders, who research shows consistently prioritize integration, perspective diversity, and long-term institutional design, are positioned to redirect organizational AI investment away from scaling and toward integration. This isn't a values choice. It's a returns-on-investment choice.

The Trillion-Dollar Question, Resolved

Return to the thought experiment. You have $1 trillion. You choose based on horizon and problem complexity.

For well-defined problems on short horizons: scale AI. The exponential cost of capability gain is worth it because the horizon is short and the problem is clear.

For complex problems or long horizons: scale humans. Build institutions, integrate perspectives, create mechanisms for collective intelligence. The superlinear returns and lower resource costs make this the dominant strategy.

Most organizations are in the second category and investing for the first. They're building massive AI capabilities and assuming those capabilities will solve strategic problems that actually depend on organizational design.

It's not that AI scaling doesn't work. It does. It's that it works for a narrower set of problems than most organizations assume—and it's increasingly expensive to keep scaling. The trillion-dollar bet should go to rebuilding how organizations think collectively, not to building larger models.

The organizations that get this right won't be the ones with the most powerful AI. They'll be the ones with the most intelligent structures for putting human and artificial intelligence in service of each other. And on a twenty-year horizon, that competitive advantage compounds.

References

^[1]: Kaplan, J., McCandlish, S., Henighan, T., Brown, T. B., Chess, B., Child, R., ... & Zaremba, W. (2020). "Scaling Laws for Neural Language Models." arXiv preprint arXiv:2001.08361.

^[2]: Hoffmann, J., Borgeaud, S., Mensch, A., Perez, E., Ashcroft, S., Wang, J., ... & Sifre, L. (2022). "Training Compute-Optimal Large Language Models." arXiv preprint arXiv:2203.15556.

^[3]: Ord, T. (2025). "The Frontiers of Machine Learning." Unpublished analysis presented at Stanford Human-Centered AI symposium.

^[4]: International Energy Agency. (2024). "Electricity 2024: Analysis and Forecast to 2027." IEA Publications.

^[5]: Google Environmental Report (2024). "Sustainability: 2024 Environmental Report." Google; Microsoft Environmental Sustainability Report (2024).

^[6]: Bettencourt, L. M. A., Lobo, J., Helbing, D., Kühnert, C., & West, G. B. (2007). "Growth, innovation, scaling, and the pace of life in cities." Proceedings of the National Academy of Sciences, 104(17), 7301-7306.

^[7]: Xiao, Y., Zhang, L., Chen, W., Tang, Y., Leviathan, Y., & Sang, X. (2025). "The Densing Law: How AI Capability Scales Through Architectural Improvements." Nature Machine Intelligence, 7(2), 114-129.