What is computational scale beyond Moore’s Law?

It is the shift from device-level transistor scaling to systems-level computational progress driven by specialized accelerators, datacenter parallelism, algorithmic efficiency, data, energy, and capital allocation.

Why does AI compute still grow if Moore’s Law has slowed?

AI compute can grow through larger clusters, better accelerators, improved networking, more efficient algorithms, larger budgets, and better data pipelines, even when transistor-density doubling slows.

What are the main constraints on frontier AI scaling?

The main constraints are electrical power, data availability, capital intensity, semiconductor supply chains, and interconnect latency across very large training systems.

Why does this matter for executives?

It matters because AI capability increasingly depends on infrastructure economics and operating model design. Strategy teams need to understand compute access, energy exposure, vendor dependence, data readiness, and where advanced AI can change unit economics.

Computational Scale Beyond Moore’s Law

By Juan Beltrán · 2026-04-23

Computational progress did not end with Moore's Law. It moved from transistor scaling to systems-level scaling across accelerators, datacenter networking, algorithms, energy, data and capital. For executives, AI capability is an infrastructure and operating-model question as well as a software one, with power, data and vendor concentration as the constraints to watch.

The shorthand that no longer fits

Many executives still plan technology as if Moore's Law were running the show: chips get faster and cheaper, and better software follows on its own. An AI strategy built on that assumption treats capability as something that simply arrives, and treats the compute behind it as a supplier's problem.

That model is now incomplete. The classical Moore-Dennard regime has slowed, while the computational capability deployed into frontier AI has kept compounding. The old law did not quietly survive. Progress moved from the individual transistor to the whole production system, which combines accelerators, datacenter networking, algorithms, energy, data, capital and organizational intent.

My argument is that compute now behaves more like energy, capacity and balance sheet than like an invisible input. It shapes what is strategically possible before a business case is written. Leaders who treat AI as a software feature will underestimate the infrastructure, economics and operating model behind it. Leaders who treat it as a systems-level production function can see more clearly where to rely on frontier platforms, where to build proprietary advantage, where energy and data constraints bite, and where algorithmic efficiency changes the business case.

From transistors to systems

For roughly five decades, computing improved because transistor density rose while power density stayed manageable. Moore's Law and Dennard scaling together gave the industry a reliable roadmap: more transistors, higher clock rates, better performance per watt and lower cost per operation.

Dennard scaling broke first, as leakage current and threshold-voltage limits made it impossible to keep raising clock speeds without unacceptable heat. Moore's Law then slowed as lithography, manufacturing complexity and cost made each new node harder to reach.

Frontier AI moved the other way. The paper's evidence review finds that training compute for notable AI models has grown by roughly four to five times a year since 2010, with frontier language models on steeper paths still. Algorithmic efficiency has improved independently, by roughly three times a year, so the same compute produces more capability over time. The new regime is led by systems rather than transistors.

The paper's main findings for strategy are these:

Classical Moore-Dennard scaling has weakened, while effective AI compute has accelerated.
Training compute for notable AI models has grown by about four to five times a year since 2010.
Algorithmic efficiency has improved independently by about three times a year.
Compute-intensive AI has already produced breakthroughs in structural biology, materials discovery, weather forecasting and formal mathematical reasoning.
The binding constraints are shifting from lithography to power, data, capital intensity and interconnect latency.
Through 2030, the most plausible limits are thermodynamic and economic rather than transistor density.

I read these as operating facts rather than optimism. Capability growth is real, and it depends more and more on scarce physical and economic inputs.

Scientific discovery as evidence

The strongest evidence for the new regime is a run of scientific results that would have been difficult or impossible under older computational assumptions, more than any benchmark chart.

AlphaFold 2 changed protein structure prediction by reaching accuracy comparable to experimental methods for many targets, and AlphaFold 3 extended the approach to interactions involving proteins, nucleic acids, ligands, ions and modified residues. GNoME expanded the known space of stable inorganic materials by orders of magnitude compared with earlier computational pipelines. GraphCast and GenCast showed that learned models can match or beat traditional weather-forecasting systems on key medium-range tasks while using far less compute per forecast.

The pattern is consistent. When a scientific domain has reliable training data, a computable ground truth and a large search space, learned models can compress the search.

Constraints through 2030

Four constraints will shape the next phase of frontier scaling.

Electrical power comes first. The largest training systems are moving toward power requirements that make energy availability a strategic input and decide where, how and by whom frontier systems can be trained.

Data is the second. Public high-quality text, code, scientific data and multimodal corpora are finite. Synthetic data, better curation, domain-specific datasets and interaction data may extend the runway without removing the limit.

Capital is the third. Frontier training runs are becoming expensive enough that only a few firms, states and partnerships can fund them directly, which increases platform concentration and makes access strategy more important for everyone else.

Interconnect latency is the fourth. Scaling beyond a single accelerator requires fast communication across clusters, and at very large scale the network that connects compute matters as much as the chips.

Four scenarios

The paper sets out four scenarios for frontier compute through 2030.

In baseline continuation, physical training compute keeps growing quickly on the back of better accelerators, larger clusters, power buildout and continued capital commitment, implying several more generations of scientific systems comparable to AlphaFold, GNoME and GraphCast.

In efficiency-led scaling, physical compute growth slows earlier while algorithmic efficiency keeps improving. Capability still advances, and advantage shifts toward research talent, architecture, data quality and training methods.

In a constrained plateau, power, data, capital and supply chains bind at the same time. Progress continues more slowly, and frontier capability in 2030 looks like an extension of the 2025 to 2026 regime rather than a step change.

In a paradigm shift, a new architecture, training method or computing substrate changes the compute equation. It is less likely on a short timeline and carries the highest upside.

The strongest objection

Most executives will reasonably say that none of this is their problem. Their company will never train a frontier model, compute is the vendors' business, and API prices keep falling.

The first point is right, and for almost every organization training frontier models would be a mistake. The rest of the objection holds up less well. Concentration, energy constraints and capital intensity decide which capabilities are available, in which regions, at what price and with what latency. A company that depends on one provider in one region for a critical workflow is exposed to those constraints whether or not it owns a single GPU.

What leaders should do

AI transformation is becoming an infrastructure-aware discipline. Executives do not need to manage chip roadmaps, but they do need to understand how compute economics affect capability access, vendor concentration, model choice, data strategy and operating-model design. For most organizations the useful question is where frontier capability changes business economics, where proprietary context matters, and where advantage can be built without owning the compute stack.

That leads to a practical agenda: treat compute access as part of AI strategy, evaluate models against workflow value, latency, cost, security and governance, build proprietary advantage around data, process knowledge, distribution, integration and change execution, track energy exposure and platform concentration as strategic risks, and plan with algorithmic efficiency in mind as well as raw model size.

Monday move

List the AI capabilities your most important workflows depend on, with the provider and hosting region for each. Mark any workflow that depends on a single provider or region with no tested alternative, and ask the owner what it would take to switch within a quarter.

Read the full working paper

The web version is intentionally concise. The full paper includes the detailed empirical review, scientific examples, constraints analysis, scenario logic and references, and is available to download below.

Research summary

Abstract

For five decades, computing progress was governed by Moore’s Law and Dennard scaling. Both regimes have now attenuated, yet effective AI compute has accelerated through specialized hardware, datacenter parallelism, algorithmic efficiency, and unprecedented capital allocation. This working paper reviews empirical evidence on training compute growth, examines compute-enabled scientific breakthroughs, and evaluates the constraints likely to shape frontier AI scaling through 2030.

Research question

If classical Moore-Dennard scaling has slowed, what now governs the growth of effective computational capability, and what does that imply for scientific discovery and AI strategy through 2030?

Methodology

Desk-based synthesis of public empirical datasets, peer-reviewed AI and scientific-discovery literature, industry roadmaps, and quantitative forecasts from Epoch AI and adjacent research. The analysis connects historical device-level scaling with contemporary systems-level scaling across compute, data, algorithms, energy, and capital.

Key findings

Classical Moore-Dennard scaling has attenuated, but effective AI compute has accelerated through system-level scaling.
Training compute for notable AI models has grown by approximately 4 to 5x per year since 2010.
Algorithmic efficiency has improved independently by approximately 3x per year, compounding the effect of physical compute growth.
Scientific breakthroughs in protein structure prediction, materials discovery, weather forecasting, and formal reasoning show how compute-intensive AI is changing discovery.
The binding constraints are shifting from lithography to power, data, capital intensity, and interconnect latency.
Through 2030, the most plausible constraint is thermodynamic and economic, not transistor density.

Executive implications

AI strategy can no longer be separated from compute economics, energy access, and infrastructure planning.
Executives should treat frontier AI capability as a systems-level production function, not a software feature.
The next competitive divide will come from organizations that understand how compute, data, algorithms, capital, and operating models compound together.
AI transformation roadmaps should include infrastructure realism, not only use case prioritization.

Limitations

This is a working paper based on public evidence available as of April 2026. The 2030 scenarios are directional, not deterministic forecasts, and should be updated as new empirical data on compute supply, algorithmic efficiency, energy availability, and model capability becomes available.

References

Moore, G. E. (1965). Cramming more components onto integrated circuits. Electronics.
Dennard, R. H. et al. (1974). Design of ion-implanted MOSFETs with very small physical dimensions. IEEE Journal of Solid-State Circuits.
Sevilla, J. et al. (2022). Compute trends across three eras of machine learning. International Joint Conference on Neural Networks.
Epoch AI. (2026). Trends in AI compute and hardware performance.
Jumper, J. et al. (2021). Highly accurate protein structure prediction with AlphaFold. Nature.
Merchant, A. et al. (2023). Scaling deep learning for materials discovery. Nature.
Lam, R. et al. (2023). Learning skillful medium-range global weather forecasting. Science.

Key takeaways

The post-Moore era is a shift from transistor scaling to systems-level scaling rather than a slowdown in computational capability.
Training compute for notable AI models has grown about four to five times a year since 2010, and algorithmic efficiency about three times a year.
The next bottlenecks are power, data, capital intensity and interconnect latency rather than lithography alone.
Compute-intensive AI has already produced results in structural biology, materials discovery, weather forecasting and formal mathematical reasoning.
Executive AI strategy needs infrastructure realism: compute access, energy exposure, vendor concentration, data strategy and operating-model design.

Decision brief

More compute does not guarantee proportional business value. Leaders should translate infrastructure progress into task economics: what becomes reliable, affordable, fast, and governable enough to change a workflow that matters.

Decision criteria

The performance gain changes an economically important task threshold.
Energy, latency, data movement, and integration costs are included.
The organization can absorb the capability without creating control debt.

Evidence to check

Cost and quality per completed task at production volume.
Energy, latency, and infrastructure utilization under realistic load.
Sensitivity of the business case to model and hardware price changes.

What I would do Monday

For one compute-heavy AI proposal, remove model parameters and benchmark headlines from the first page. Replace them with cost per useful task, acceptable error rate, response time, and business owner.

Questions executives ask

What is computational scale beyond Moore’s Law?: It is the shift from device-level transistor scaling to systems-level computational progress driven by specialized accelerators, datacenter parallelism, algorithmic efficiency, data, energy, and capital allocation.
Why does AI compute still grow if Moore’s Law has slowed?: AI compute can grow through larger clusters, better accelerators, improved networking, more efficient algorithms, larger budgets, and better data pipelines, even when transistor-density doubling slows.
What are the main constraints on frontier AI scaling?: The main constraints are electrical power, data availability, capital intensity, semiconductor supply chains, and interconnect latency across very large training systems.
Why does this matter for executives?: It matters because AI capability increasingly depends on infrastructure economics and operating model design. Strategy teams need to understand compute access, energy exposure, vendor dependence, data readiness, and where advanced AI can change unit economics.

Canonical URL: https://juanbeltran.ch/blog/computational-scale-beyond-moores-law