What Is Moore’s Law and Why Does It Matter for AI Progress?

What Moore’s Law Actually Says

Moore’s Law is named after Gordon Moore, co-founder of Intel, who in a 1965 paper observed that the number of transistors that could be packed onto an integrated circuit was doubling approximately every year, and he predicted that trend would continue for at least a decade. In 1975, Moore revised his estimate to a doubling roughly every two years, a version of the observation that became the widely cited form of the law. Importantly, Moore’s Law was never a law of physics in the way gravity is. It was an empirical observation and an industry expectation, one that chip manufacturers like Intel, and later TSMC and Samsung, actively worked to fulfill by investing enormous resources in research and manufacturing improvements, effectively turning the prediction into a self-reinforcing roadmap for the entire semiconductor industry.

Why Transistor Density Matters So Much

Packing more transistors onto a chip of the same size generally means more computing power without a corresponding increase in cost, size, or energy use per calculation. This compounding effect is what allowed computers to shrink from room-sized machines in the 1960s to the smartphones everyone now carries, while simultaneously becoming dramatically more powerful. Each doubling of transistor density has historically translated, with some lag, into faster processors, larger memory capacity, and new categories of devices that would have been physically or economically impossible before. The personal computer, the internet as we know it, smartphones, and cloud computing all depended on decades of transistor scaling that followed roughly the curve Moore described.

The Link Between Moore’s Law and AI Progress

Modern AI, and deep learning in particular, is extraordinarily compute-hungry. Training a large neural network involves performing enormous numbers of matrix multiplications across massive datasets, and the scale of that computation has grown dramatically with each new generation of models. The affordability of that scale has depended directly on decades of semiconductor progress: cheaper, denser, more powerful chips meant that researchers could train bigger models on more data without costs becoming completely prohibitive. The 2012 ImageNet breakthrough, where a deep convolutional neural network called AlexNet dramatically outperformed previous computer vision approaches, was only possible because consumer graphics processing units, originally built for rendering video games, had become powerful and affordable enough to train deep networks in a reasonable amount of time. Without decades of transistor scaling driving down the cost of raw computation, that moment, and the deep learning boom that followed it, would have arrived much later, if at all.

GPUs, TPUs, and Specialized AI Hardware

As AI workloads grew, the industry did not just rely on general-purpose chips getting denser; it began building specialized hardware optimized specifically for the kind of parallel computation that neural networks require. Nvidia’s graphics processing units became the default hardware for training deep learning models throughout the 2010s, while Google developed custom chips called Tensor Processing Units, or TPUs, starting in 2016, designed from scratch to accelerate the specific mathematical operations behind neural networks. This shift represents an important nuance: even as traditional Moore’s Law transistor scaling has slowed, overall computing power available for AI workloads has continued to grow rapidly through architectural specialization, parallelism, and software optimization, a pattern some researchers now call the era of “more than Moore” progress.

Is Moore’s Law Slowing Down?

Since the mid-2010s, many semiconductor engineers and industry observers have argued that traditional Moore’s Law scaling is slowing considerably. As transistors shrink toward the scale of just a few nanometers, manufacturers run into fundamental physical limits, including quantum effects that make transistors unreliable at extremely small sizes, as well as skyrocketing costs for the manufacturing equipment needed to produce chips at that scale. Gordon Moore himself acknowledged in later years that the pace would eventually have to slow due to these physical limits. Even prominent chipmakers like Intel have publicly stated that the traditional two-year doubling cadence no longer holds as reliably as it once did, though the broader semiconductor industry continues to find ways to extract more performance through new transistor designs, three-dimensional chip stacking, and advanced packaging techniques.

What a Slowing Moore’s Law Means for AI’s Future

If raw transistor scaling continues slowing, the AI industry faces a genuine question about how it will keep training ever larger and more capable models. Some of the largest AI labs have responded by shifting focus from simply building bigger models toward improving efficiency: better training techniques, smarter architectures, and algorithms that extract more capability from the same amount of compute. Others have responded by investing directly in massive data center buildouts and chip supply, treating compute capacity itself as a strategic resource worth securing years in advance, a dynamic visible in the enormous capital investments major AI companies have made in custom silicon and dedicated data centers throughout the mid-2020s. Whether through new chip architectures, quantum computing research, or simply smarter software, the AI field’s continued progress remains closely tied to the same underlying question Moore first raised in 1965: how much more computing power can be packed into the same physical and economic constraints, and for how much longer.

Why This History Still Matters

Moore’s Law is often treated as a historical curiosity, a line on a chart that technologists like to cite. But its real significance lies in how directly it connects to the AI systems people use every day. Every large language model, every image generator, every recommendation algorithm running in the background of a mobile app depends on a vast amount of affordable computing power that exists only because of six decades of relentless semiconductor improvement. Understanding Moore’s Law, and understanding that it is now bending rather than breaking cleanly, is essential context for understanding both how AI got this far and what constraints might shape where it goes next.