A single AI forecast inflates US power auctions by $5.7B

Australia's latest inflation figures showed that prices in the category covering audio, visual, and computing equipment rose 6.1% in a single quarter.

That's by far the largest increase in the Australian Bureau of Statistics series, which began in 1980. Computers, TVs, and electronics in general have gotten cheaper and delivered more performance for the same money.

Customers around the world, including Australia, have (reasonably) expected this trend to continue.

This is why the third quarter came as such a surprise, as it interrupted a consumer expectation that has been in place for decades.

The Reserve Bank of Australia linked higher prices for some computers and other consumer electronics to rising memory-chip costs associated with demand for AI infrastructure.

The same components, or the manufacturing capacity needed to produce them, are now being pursued by companies whose shopping budgets look nothing like yours.

On Nov. 3, the RBA raised its cash rate to 4.85%, citing AI-related investment as one of several sources of price pressure. Higher fuel costs and domestic capacity constraints were also involved, and the bank said AI's effect on overall consumer prices remained modest.

The technology hasn't single-handedly taken over Australian monetary policy, although it has progressed from a promise in investor presentations to a consideration in decisions that affect ordinary borrowers.

We've also seen this in the US, where AI demand has reached much further into the system that determines the cost of electricity. PJM, the grid operator serving all or parts of 13 states and Washington, D.C., runs auctions that pay power plants and other resources to be available when the grid needs them, years before the electricity is actually used.

In its analysis of this summer's auction, PJM’s independent market monitor calculated that power providers would receive about $16.4 billion for making capacity available during the 2028-29 delivery year.

The monitor then reran the auction without the forecast growth in demand from new and expanding data centers, while leaving existing target="_blank" rel="noopener">AI's effect on inflation is: the spending starts producing identifiable, quantifiable consequences long before we see any productivity gains.

The costs also reach people who aren't buying AI services, while the institutions assigning those costs are still working out how much responsibility the companies requesting the infrastructure should carry.

Your new computer has some very wealthy competition

Data centers are buildings packed with computing equipment, along with the cooling systems and electricity connections needed to keep it working.

However, AI demand extends well beyond the GPUs that attract most of the attention.

Filling a data center with expensive chips also requires the hardware that feeds them information and keeps them running, creating competition for parts and factory capacity that other electronics manufacturers depend on.

The RBA's account follows that competition through to Australian shops, where higher memory costs contributed to more expensive consumer electronics.

The ABS's 6.1% quarterly equipment-price increase covers a particular statistical category, so it can't be treated as the price increase for every computer, much less a precise measurement of AI's contribution. It does, however, give the central bank's explanation an observable consumer outcome.

Again, something similar has been happening in the US. The September consumer-price report recorded a 9.7% annual increase for computers, peripherals, and smart-home assistants, while smartphone prices were 9.9% lower than a year earlier.

The San Francisco Fed calculated that prices for information-processing equipment rose 13% over the year to August, but added only 0.2 percentage points to that month's 3.4% headline inflation rate under the personal consumption expenditures measure. Its economists found limited evidence that those pressures had spread broadly across the economy, where energy costs and tariff-related goods prices were doing considerably more of the damage.

The Bank of England was also monitoring demand for AI components as a source of global cost pressure, but judged the potential effect of these and other non-energy risks on UK inflation to be much smaller than the energy shock and little changed since September.

Its quarter-point increase in Bank Rate to 4% was justified by energy-driven inflation risks, alongside stronger-than-expected activity.

Another complication in software is that a subscription can become more expensive when it starts doing something its previous version couldn't.

Federal Reserve researchers have warned that parts of the inflation data may not adequately account for those improvements. Paying more can still hurt a household budget, even when part of the increase reflects better quality.

Electricity is another way costs can spread because both new and existing customers depend on the same system. Someone replacing a laptop can postpone the purchase, whereas households connected to a grid that accommodates substantial new demand have far less control over what gets built and how the resulting costs reach them.

The electricity bill starts with a forecast

PJM coordinates the power system across much of the eastern US, and one of its jobs is to make sure enough electricity can be supplied when demand is high.

Alongside markets for delivered electricity, it runs auctions that pay suppliers to commit capacity for a future year, meaning they agree to be available under the market's rules when the system needs them.

That availability has a separate price from the electricity eventually generated. Power stations need sufficient revenue to remain in service or justify investment, while the operator has to plan before households and businesses switch on equipment that will require supply.

Forecasting how much power customers will need therefore becomes part of a financial transaction.

PJM's July auction for the 2028-29 delivery year produced approximately $16.41 billion in auction revenue. In a recent analysis, its independent market monitor, Monitoring Analytics, reran the auction without the forecast growth in demand from new data centers and expansions of existing facilities, while leaving other assumptions unchanged.

The resulting revenue fell to approximately $10.7 billion, a modeled difference of $5.7 billion for that delivery year.