Heather Stewart Argues Deferred Data-Center Bills Could Test AI Boom

Heather Stewart’s new Guardian analysis shifts the AI debate from model safety to financial timing: expensive data centers can be built now while much of the cost to their eventual users arrives later. Her argument is that falling AI prices and deferred compute commitments could make the industry’s growth assumptions harder to sustain.

The business model depends on two clocks

The scale of the upfront expansion is substantial. Stewart cites an estimate that Google, Amazon, Microsoft, Meta and Oracle will issue $132 billion in debt in 2026 to fund data-center construction. Debt alone does not establish that the projects will fail, but it raises the cost of being wrong about demand and future cash flow.

At the same time, the revenue side is under pressure. A Silicon Data index tracking what customers pay per million tokens, the units processed by large language models, has dropped by more than half since June to below $1, according to the analysis. Stewart also cites Bloomberg reporting that OpenAI has repeatedly cut fees to retain customers, while inputs such as semiconductors remain costly.

The commitments are not conventional debt

That structure means the obligation is not technically debt, but Stewart argues that it can create a similarly sharp financial test when capacity comes online. Groundbreaker’s analysis puts the broader “compute commencement wall” facing AI labs over the next several years at $1.5 trillion. That is an estimate, not a disclosed industry liability total, and its significance depends on contracts maturing as expected and revenues failing to keep pace.

Revenue growth is the assumption carrying the structure

Three pressure points in Stewart’s argument

• Cloud giants are adding debt to fund construction before all future demand is proven.
• Customers are paying less for AI processing even as major infrastructure inputs remain expensive.
• Frontier labs may face large contractual costs only after new facilities are ready to use.

The profitability measures cited in the piece add to the uncertainty. Stewart reports that Anthropic told investors its adjusted operating income was positive, while noting that this measure excludes many costs. That does not settle the companies’ financial health; it illustrates why definitions of profitability matter when infrastructure spending is still being financed and future obligations have yet to arrive.

The unresolved issue is therefore more basic than whether AI demand exists today. It is whether demand can become paid, durable revenue at a pace fast enough to absorb a huge wave of operating commitments. Stewart’s warning is that a safety debate about slowing AI development should not obscure the separate financial risk embedded in the race to build capacity.