If you want to understand why AI tools feel dramatically more capable every year, the single most useful number to watch is compute — the raw amount of calculation used to train a model. Over the past decade, the computation behind leading AI systems has grown exponentially, roughly doubling on a remarkably short cadence. This is not a minor technical detail; it is the engine behind nearly every improvement you have noticed in the tools you use.

Scale Has Been the Story

The headline finding from years of AI research is almost embarrassingly simple: bigger has tended to mean better. Models trained with more compute, on more data, with more parameters, have consistently outperformed their smaller predecessors. Breakthroughs that once required clever new algorithms increasingly come from simply scaling up what already works. That is why each new generation of writing and image tools tends to leap ahead of the last.

The Cost Behind the Curve

Exponential compute growth comes with an exponential bill. Training a frontier model now demands enormous clusters of specialised chips, vast amounts of electricity, and infrastructure that only a few organisations can afford. This is why the most powerful models come from a small number of well-funded labs, and why access to those models — rather than ownership — is what most businesses will rely on.

What It Means for Your Toolkit

For marketers, the compute curve is good news with a caveat. The good news is that capabilities keep improving and trickling down into affordable products. The caveat is that the frontier is controlled by a handful of players, so it pays to avoid over-committing to any single provider. Build your workflows so you can swap the underlying model as the landscape shifts — because it will keep shifting as long as compute keeps climbing.

Diminishing Returns and What Comes Next

No exponential lasts forever, and there are real questions about how long pure scaling can keep delivering gains before costs, data limits or efficiency concerns force a change of approach. Researchers are already exploring ways to get more capability from less compute, which would broaden access and lower prices. For now, though, the compute curve remains the clearest explanation for why each generation of tools outclasses the last — and a useful reminder that the pace of improvement is tied to physical and financial limits, not magic.

Source: Our World in Data — Artificial Intelligence.