The scalability problem behind AGI

Depending on how you define AGI, some argue we have already reached it. 

It is unarguable that the progress the AI industry has made is significant. In under four years, AI moved from conversational interfaces to autonomous action. But underpinning all of it remains a single foundation: pre-trained models. And if this is indeed AGI, it is AGI built on a foundation that cannot scale indefinitely.

The cost of compute

The computational requirements of frontier AI models have grown at a pace that bears little sustainable relationship to the value being extracted. Training compute for frontier LLMs has been growing at roughly 5× per year since 2020, doubling approximately every five months. 

The infrastructure investment required to sustain this is enormous. McKinsey projects that meeting AI compute demand through 2030 will require up to $5.2 trillion in data centre investment. Global data centre electricity consumption stood at approximately 415 terawatt-hours in 2024 (around 1.5% of total global electricity use) and is growing at more than four times the rate of overall global electricity consumption. 

Serving pre-trained AI models 

To remain useful, pre-trained models require retraining, and each retraining cycle costs millions of dollars, is resource intensive, and has to be repeated regularly. That level of investment is difficult to sustain indefinitely.

But training costs are only one part of the picture. For organisations deploying AI in production, inference (the ongoing cost of running a model) accounts for 80 - 90% of total expenditure over a model's active life. 

AI is already becoming embedded in how we work, communicate and how decisions get made, and we are still at the beginning. The systems we experience today handle a fraction of the tasks they are ultimately expected to handle, and already consume energy at a scale that is straining national grids and absorbing capital that dwarfs most industries.

Now if we consider what civilisational-scale AI actually looks like, that would be AI genuinely woven into healthcare, infrastructure, education, and governance across the world. If the foundation that gets us there is the same one we are building on today with static models, periodic retraining, and exponentially growing compute demands, the numbers start being incomprehensible.

The structural contradiction

A definition of AGI based on the fact that an agent can recover from failure and persist until a task is complete runs into difficulty. The agents are indeed autonomous but the models underneath them are not. They are static, eventually become out of step with the world they are being asked to navigate, and extraordinarily expensive to update.

But the question of whether we have reached AGI is, in some ways, beside the point. What matters is the direction of travel, and the current trajectory is unlikely to be the right path.