The Engineer Behind Moore’s Law for RF Has a New Paper on How Artificial Intelligence Scales

Moore’s Law is one of the most useful ideas in the history of technology. Gordon Moore’s 1965 observation that the number of transistors on a chip would double roughly every two years did not just describe what was happening in the semiconductor industry. It gave engineers, investors, and policymakers a shared framework for planning around a future they could not yet see. Chip designers built roadmaps around it. Venture capitalists bet on it. Governments used it to guide industrial policy. For five decades it held.

Artificial intelligence has grown faster than almost any technology in history, yet until now it has had no equivalent framework. People in the field have known that AI systems are getting more powerful at a remarkable rate. What nobody had formally established is why that rate holds, what is driving it, and what will eventually stop it. A new preprint from Asif Alam, an RF engineer and researcher at Florida International University, and co-author Muhmmad Shah Alam of Imam University, proposes exactly that framework for the first time.

The paper is titled Scaling Laws for Artificial Intelligence. It is the first to propose a scaling law for AI, one that goes beyond documenting that growth is happening and answers the more fundamental questions: what is actually causing it, and which of those causes will run out first?

To understand why that question matters, it helps to understand what the paper is actually measuring. Training a large AI model requires an enormous amount of computation, the kind measured in numbers so large they are almost meaningless to describe in ordinary terms. The biggest models today require computational resources that would have been unthinkable just a few years ago, and the resources required for the biggest models have been growing at a pace that makes frontier AI from even two years ago look modest by comparison. The question the field has struggled to answer is whether this growth will continue, slow down, or suddenly stop, and what would cause each of those outcomes.

Alam’s framework breaks the growth of AI training compute into three separate components, each driven by a distinct and independently measurable mechanism. The first is hardware: the raw computational power of the chips used to train AI models. The second is infrastructure: how many of those chips have been deployed globally. The third is what the paper calls the efficiency absorption factor, which captures how improvements in algorithmic efficiency get reinvested into more ambitious training runs rather than returned as cost savings.

The insight that ties these three together is that they form a chain, and the chain is directional. Better chips make it worthwhile to deploy more of them. More deployed infrastructure makes it possible to train larger models. The economic returns from larger models create pressure to improve algorithms. And those algorithmic improvements get plowed back into even more ambitious training runs, sustaining the growth of the whole system. Each link in the chain has its own independent driver and its own independent ceiling. The paper derives both.

That ceiling analysis is where the work becomes practically significant. The three components do not all run out at the same time or for the same reason, and the sequence matters enormously for anyone trying to forecast what AI development will look like over the next several years.

The first constraint to bite is not the chips themselves. It is the electrical grid. The global infrastructure supporting AI compute has been growing at a rate that infrastructure planners who rely on historical data center growth projections will systematically underestimate. The paper shows that power capacity for AI infrastructure is the most consequential near-term constraint on frontier AI growth, and it is the first ceiling to bind in the cascade.

The second constraint is the energy required for a single training run. As training runs have grown in scale, the electricity consumed by each one has grown proportionally. The paper projects that if current trends continue, a single frontier training run will require an amount of energy comparable to the annual output of a large power plant. This energy ceiling on individual runs binds after the power ceiling on infrastructure but before the hardware ceiling.

The hardware ceiling, the point at which transistor physics fundamentally limits further improvement in chip performance, arrives last. This is the same physical limit that Alam’s RF scaling work identified as relevant to wireless transistors, where frequency scaling eventually runs into quantum mechanical constraints at the atomic scale. The paper draws the parallel explicitly and notes that the pattern of improvement is likely to follow the same step-plateau structure seen in RF: steady progress within each device paradigm, followed by a discrete transition to a new paradigm at a modified rate.

The framework also carries a direct implication for policy. Because power infrastructure is the first constraint to bind, energy policy is a more consequential lever for governing AI capability growth over the coming decade than semiconductor export controls. Controls on hardware affect the deployment and infrastructure component of the chain, but that effect is bounded by the energy ceiling that arrives first. This is a specific, testable, and practically actionable claim.

The paper places this work in the tradition of enduring scaling laws: Moore’s Law for semiconductors, Koomey’s Law for computational energy efficiency, Nielsen’s Law for internet bandwidth, Edholm’s Law for wireless data rates, Cooper’s Law for spectral efficiency, and Alam’s own RF Frequency Scaling Law, which unified transistor scaling across a century of development and is cited in the new paper as a structural precedent. What all of these laws share, the paper argues, is not just an observed doubling time but a mechanism. Laws that survive changes in growth rate survive because the mechanism was correctly identified. A curve fit that captures growth without explaining it becomes obsolete the moment the rate changes. A mechanistic framework remains structurally valid even as the numbers evolve.

That is the ambition of the AI scaling paper. Not to predict a specific number, but to identify the structure of the system, the components that drive it, the direction of causality between them, and the sequence in which their constraints arrive. Whether those specific growth rates hold or shift, the framework is designed to remain useful as a forecasting tool through the transitions.

For a researcher whose primary body of work involves designing antennas and waveguide transitions for next-generation wireless systems, publishing a mechanistic scaling law for artificial intelligence is an unusual move. The connection is not incidental. The same analytical approach that produced the RF Frequency Scaling Law, finding the mechanism behind an observed exponential trend and deriving the ceiling at which the mechanism fails, is what produces the AI scaling identity here. The field is different. The method is the same.