Home
Dossiers
AI Was Supposed to Cut Costs. First, We Have to Build It

AI Was Supposed to Cut Costs. First, We Have to Build It

AI may eventually reduce costs through productivity. Before that happens, data centers, chips, energy and debt are already making the technology intensely physical.

Artificial intelligence is usually sold as software: a model, a chatbot, a feature added to an application. Its economic expansion is much more physical. Behind every generative service sit data centers, chips, power grids, construction sites, memory, cooling systems and capital. When all of those resources are demanded at enormous scale at the same time, AI begins to affect more than the technology sector. It can begin to affect prices across the economy.

The issue has now entered the US monetary debate. In minutes from the Federal Reserve’s September 2026 meeting, released on October 7, officials cited AI-related investment as one of the pressures supporting inflation. The Washington Post reported that data centers, computers and chips are adding to demand while productive capacity struggles to expand at the same speed.

There is an uncomfortable paradox here: AI is marketed as a technology that will lower costs, but building it requires investment that can raise costs first.

AI is an investment before it becomes productivity

The economic promise is simple: companies and workers will produce more value with less time. Between that promise and the result, however, sits an infrastructure phase that is easy to hide behind the interface.

A data center needs buildings, transformers, grid connections, cooling, servers, storage and thousands of accelerators. When many companies invest at once, demand rises for components, specialized labour and energy.

Those pressures can push up prices and wages before the efficiency gains promised by AI spread through the economy. It is a familiar pattern in large infrastructure cycles: the benefit arrives later; the spending arrives now.

Chips and memory have become macroeconomic goods

The AI race has turned components once confined to electronics into strategic resources. GPUs, high-bandwidth memory and networking systems are bought in huge volumes by hyperscalers and newer infrastructure operators.

The effect can cascade. If manufacturers shift capacity toward higher-margin data-center components, other industries may compete for the same production lines or related materials. The price of an ordinary laptop or server can therefore be affected indirectly by AI demand.

The pressure is visible even in metals. Reuters reported on October 7 that demand linked to technology investment had helped push tin — widely used in electronics solder — toward historically high prices.

The cost of AI also passes through debt

Data centers are not financed only from Big Tech profits. The scale of the build-out is increasing reliance on credit and bond markets.

On October 8, Reuters noted that companies in the AI race were raising capital at a scale large enough to compete with government issuance. When very large corporations absorb a growing share of available financing, they can raise the cost of capital for others and change the dynamics of bond markets.

This dimension is often missing from the AI debate. We are not merely buying better software. We are building industrial infrastructure financed by enormous amounts of capital.

Energy and construction connect AI to the physical economy

A data center cannot be placed anywhere. It needs available power, grid connections, land, permits and cooling capacity. In some regions, clusters of projects can increase competition for electricity and construction resources.

AI therefore enters prices through several channels at once: electronic components, power infrastructure, industrial real estate, specialized labour and financing.

That does not mean data centers must be inflationary over the long term. If AI genuinely raises productivity, higher output could reduce costs and offset some of the initial pressure. This is precisely the uncertainty raised by Fed officials: both the magnitude and timing of future productivity gains remain difficult to measure.

The risk is paying today for a benefit that may arrive tomorrow

The economics of AI depend on a temporal bet. Companies spend now because they expect future demand and productivity improvements to justify the investment. If adoption grows fast enough, this infrastructure may become the base of a new economic phase. If revenue remains below expectations, part of the build-out could become excess capacity financed by debt that is difficult to sustain.

The risk is beginning to appear in the valuation of so-called neoclouds, companies created specifically to sell AI infrastructure. Reuters Breakingviews pointed on October 8 to the difficulties surrounding the IPO of Firmus Technologies, backed by Nvidia and Blackstone, as a sign of increasing investor caution toward aggressive valuations.

The real cost of AI is not on the chatbot subscription page

A twenty- or fifty-euro monthly subscription hides a much larger chain. Behind the service price are semiconductor fabs, data centers, power plants, transmission grids and financial markets.

That is why the economics of artificial intelligence cannot be measured only in cost per token. The more important price may be the real resources required to build the infrastructure before we even know how much value it will generate.

AI may eventually become deflationary through higher productivity. But first it has to pass through a phase in which it consumes capital and physical resources on an extraordinary scale. The distance between investment and return is where one of AI’s most consequential — and least visible — costs is hiding.