The supply chain starts with producers and designers of AI infrastructure: firms like TSMC and Samsung, which fabricate chips; Nvidia, which designs them; and Cisco, which provides connectivity. Then come the hyperscalers like Amazon, Google and Microsoft. They are building data centres both for their own AI models and to sell compute (processing power) to others. In addition to the hyperscalers, there are more specialised companies like Equinix (data centres) and, of course, Anthropic and OpenAI, the developers of foundational LLMs.
Finally, there are the individual and corporate end users of AI services. Individual use is growing rapidly, and corporate use in some areas (such as software development and customer support) is exploding.
But most large businesses, while experimenting extensively, have yet to implement end-to-end use cases. Many still need to organise their historical data to train AI for their own purposes, and to restructure their traditional operations so that AI can be deployed to improve with experience. Moreover, many firms rightly worry about data security, AI errors, and hallucinations that could destroy their brand image. Still, as less-conservative younger companies adopt more AI uses, they will put competitive pressure on older, larger firms to change.
The AI rollout could nevertheless be interrupted in several ways, posing risks for debt-funded players. For instance, if graphics processing units, CPUs, and memory chips become faster and more energy efficient, the equipment filling existing data centres could depreciate rapidly, making it harder for them to amortise their costs. And LLMs, which have become extraordinarily capable through essentially next-word prediction, could plateau until a new technique emerges.
See also: Mistral signs Airbus and BMW as it brings AI to manufacturing
For now, AI labs are investing massive sums to train newer, larger models, on the assumption that the first model to reach some magic point where it becomes self-improving will rule the AI world and reap enormous profits. But this scenario seems implausible. Even if there is such a point, competitors could still match the first mover’s model (including by hiring away key employees to obtain technical trade secrets).
So far, no AI model has gained a sustained advantage. Unless Gemini (Google), Claude (Anthropic), and ChatGPT (OpenAI) can eventually differentiate themselves by appealing to specific user segments (or by merging or colluding), it is hard to see where the profits justifying their enormous training investments will come from.
Moreover, although politicians have been largely standing on the sidelines so far, policy interventions to address AI risks and concerns are inevitable. Since data centres consume tremendous amounts of power — driving up power prices for everyone — state and local governments will face increased political pressure to limit their construction. In Indiana, for example, multiple counties recently enacted a moratorium on data centre construction.
See also: AI boom fuels record US$14.5 bil in Taiwan tech firm borrowing
Projections for next year already suggest that hardware makers and data centres will be unable to supply enough US compute capacity. And as compute shortages mount, end users will have more reasons to delay implementation. You cannot reorganise all your operations around AI if you have good reason to worry about future access reliability or reasonable pricing.
Worse, whereas broader use may take longer than many expect, malevolent use by hackers and deepfakers, as well as unsupervised use by children, is growing rapidly. It is not difficult to imagine disaster scenarios — such as a deadly cyber incident, gross data misuse by AI agents, or poorly trained AI models advising children to commit acts of violence against themselves or others (something that has already happened). The chorus demanding regulation and more liability for AI models will only grow louder. The risks posed by rogue AI could even prompt a sorely needed dialogue among major powers, perhaps leading to some kind of AI Geneva Convention.
Perhaps the most important trigger for political intervention would be massive AI-related job losses. Fearful of political or social backlash, even firms inclined to adopt AI may hesitate to shed redundant employees outside a recession, thereby reducing the gains from AI deployment and diffusion.
Given all these uncertainties, it is far from clear how widely and quickly AI will be rolled out, and who will profit. Hardware manufacturers and designers seemed well-positioned, given the tremendous demand for computing. But if data-centre construction is interrupted, that could shift profits to hyperscalers and AI labs. They might reduce the amount of compute dedicated to training better models, which gives them only fleeting advantages, and shift to selling the compute they have sewn up to firms using their already capable models. Such shifts are also likely if model capabilities plateau. Regulation might also force modellers to spend more effort on improving the training and safety of existing models, building broader public trust.
The good news is that a more limited, careful AI rollout could give firms more time to find labour-augmenting (as opposed to labour-displacing) uses, and governments and workers more time to adjust. The bad news is that euphoric visions of quick exceptional profits could be unfounded, a particular problem for AI firms that have to make unforgiving debt payments. AI advances will likely pay off eventually. But not every provider will profit, or even survive. —
©Project Syndicate
Raghuram G Rajan, former chief economist of the International Monetary Fund, is a professor of finance at the University of Chicago Booth School of Business
