The economics of AI are yet to be determined

The prevailing narrative is that AI is cheap, getting cheaper, and will remain so. That narrative is wrong. Or at best, dangerously premature.

Four structural forces are converging that challenge the assumption: rising energy costs for data centres, persistent compute scarcity, higher bills for more intensive model use, and a vast unresolved liability around the content that trained these models.

None of these forces are hidden. But few people are connecting them. When you stack them together, they reframe the economics of AI not as a settled trajectory, but as an open question with significant downside risk for businesses building on today’s pricing.

The subsidy layer

Right now, AI feels like a bargain. For a few hundred dollars a month, you can access models that perform work equivalent to tens of thousands in human labour. The maths seems obvious.

But this comparison only holds if you ignore the subsidy structure underneath it. The frontier labs are investing heavily to acquire users and build capacity. Their funding models differ: Google’s parent, Alphabet, also has substantial operating cash flow. A subscription price does not tell us whether a provider recovers its full development and infrastructure costs. That is the subsidy risk I am concerned about. The cost pressures are substantial:

• Power. Data centres consume staggering amounts of electricity, and demand is accelerating. In the US alone, the IEA projects data centre electricity consumption to rise by 130% between 2024 and 2030. That is a demand forecast, not a forecast that every provider pays more per unit of electricity. The pressure on capacity is the point.
• Hardware. Providers compete for advanced chips and data-centre capacity, using both NVIDIA hardware and their own accelerators. Training runs are only part of the picture. The ongoing compute required to serve millions of concurrent users at inference is an escalating operational cost.
• Development velocity. More capable models can consume more tokens and run longer workflows. Claude Opus 4.6 launched at the same base token prices as Opus 4.5, with premium pricing for prompts exceeding 200,000 tokens. Its deeper reasoning can also increase the bill per task. Those are customer prices and usage effects, not disclosed evidence of Anthropic’s internal cost per task.

My expectation is that the pressure to recover investment will surface in pricing, usage limits or product packaging. The question for a business is how much exposure it has when that happens.

The content liability

There’s a second structural cost that may prove even more consequential.

Today’s models draw on vast quantities of content created by writers, researchers, artists, and journalists. Some of that use is licensed: OpenAI’s March 2024 agreements with Le Monde and Prisa Media explicitly included training. Other uses remain contested by rights holders. Those agreements do not settle the wider question of how value returns to everyone whose work makes a model useful. You could reasonably describe the uncompensated part as extracted value at scale.

I’ve written about this value exchange problem elsewhere, so I won’t relitigate the argument here. What matters for the economics is this: the bill comes due. Whether through regulation (the EU AI Act already sets a direction), litigation (the New York Times v. OpenAI being the most visible case), or commercial negotiation, model providers will need to pay for access to the content that makes their products useful.

And as models require increasingly current information to remain relevant, this isn’t a one-off settlement. It’s a recurring liability that compounds.

The vendor control problem

There’s a third dimension that most businesses are overlooking entirely. When you replace a knowledge worker with an AI subscription, you’re not just changing a cost line. You’re shifting where control sits.

Dimension	Employee model	AI vendor model
Cost predictability	Salary, fixed and planable	Subscription, subject to pricing and contract terms
Capability stability	Consistent within role	Model versions deprecated, capabilities shift between releases
Operational control	Works within your systems, on your terms	Platform-dependent; terms of service govern access
Strategic dependency	Labour market risk, manageable	Single-vendor risk, concentrated

An employee operates within your business. An AI model is a vendor platform. The provider can change prices subject to its terms, retire a model you’ve built workflows around, or change how the product functions. Your business absorbs the impact.

This isn’t theoretical. We’ve already seen models deprecated mid-cycle, pricing structures restructured, and capabilities vary materially between versions. The more deeply you integrate AI into core operations, the more exposed you become to decisions made in someone else’s boardroom.

The efficiency counter-argument

To be fair, there are forces working in the other direction. Models will become more efficient to train and run. Standard queries will require less compute over time. Hardware will improve. Competition between providers will exert some pricing pressure.

But efficiency gains don’t eliminate structural costs. They moderate them. Energy consumption, hardware scarcity, content licensing, and continuous development aren’t going away. They’re baked into the economics of the technology itself. Efficiency is a slope; the structural costs are a floor.

The implication

The economics of AI are far from settled. My view is that today’s cheap access reflects a period of expansion funded by investors and major technology businesses, with unresolved content liabilities alongside it. It is not a reliable guide to the price of a future workflow.

AI will deliver value. That’s not the question. The question is whether businesses are pricing the risk correctly. Three things should be on every leadership team’s radar:

1. What is your true cost exposure? If your operating model assumes current AI pricing holds, what happens when it doesn’t? Model the scenario where costs double or triple within 18 months.
2. Where does vendor dependency sit? Map the workflows, decisions, and outputs that now depend on a single model provider. Understand the concentration risk.
3. How defensible is your data position? As content licensing costs flow through to model pricing, businesses with proprietary, structured data will have optionality that those relying entirely on third-party models will not.

The structural forces are clear. The timeline isn’t. But building strategy on the assumption that today’s pricing is permanent is a bet, not a plan.