Investors start counting the cost of AI. Firm capacity is becoming the new currency in the race

US technology companies are pouring tens of billions of dollars into artificial intelligence data centers every quarter, but money alone cannot secure electricity. PJM, the grid operator serving 13 US states, failed to procure enough capacity in its latest auction even at the price cap, with secured resources falling 6,831 MW short of the reliability requirement. Hyperscalers are therefore signing energy contracts lasting decades, guaranteeing equipment orders for future resources and even buying flexibility from households. Firm capacity is thus becoming the new currency that companies are competing for.

Friday’s slide on Wall Street was not just a reaction to weaker quarterly results. Investors were also considering whether massive AI spending could be turned into the necessary infrastructure—operational data centers—in time. The Nasdaq Composite fell 2.8% on October 30, its worst day since June. The sell-off was triggered in part by results from Amazon and Google parent Alphabet, which raised investor concerns about the cost of developing AI, according to UPI.

Money is not the only constraint, however. Although companies’ capital spending is outpacing the cash their businesses generate, the grid’s physical limits are a more formidable obstacle. Even a company with enough money cannot quickly secure electricity for new data centers. Grid capacity is therefore determining the pace of AI development just as much as budgets are.

Money is not enough; capacity is arriving too late

Even ample funding does not guarantee that investment will translate into capacity as quickly as a company needs. Companies have to focus intensively on the availability of key resources. Microsoft said that deliveries of some capacity from external providers were delayed in the past quarter. According to Chief Financial Officer Amy Hood, the delays mainly affected providers of graphics processing units as a service. Hood also reiterated that demand continues to exceed supply. Analysts had expected investment of USD 50.67 billion. Microsoft’s positive cash flow therefore partly reflects the fact that the company was unable to spend everything it had planned to.

The availability of energy also plays a role in this development, alongside chips. Alphabet’s commitments illustrate this. At the end of September, the company had yet-to-commence leases, primarily for data centers and energy infrastructure, with future payments of USD 130.5 billion. A year earlier, the figure was USD 42.6 billion, and in the second quarter of 2025 it was USD 23.9 billion. The total has therefore increased roughly fivefold in less than five quarters.

Alphabet signs energy supply contracts for its data centers lasting between three and 26 years, with commitments extending through 2054. They typically include take-or-pay clauses, requiring payment for a minimum volume regardless of actual consumption. The company has also provided guarantees of up to USD 16.2 billion so that its counterparties can order long-lead-time equipment for future power purchase agreements. In March, it bought renewable energy developer Intersect for USD 5.9 billion, saying the acquisition would accelerate the development of both data centers and energy infrastructure. According to a Moody’s analysis this year, the five largest hyperscalers have USD 662 billion in lease commitments yet to commence, which have not yet appeared on their balance sheets.

Another tech giant, Amazon, is also securing large volumes of electricity. It has around 690 million MWh in energy contracts recorded as derivatives, with an average remaining term of about 18 years.

PJM auction shows capacity is scarce even at the price cap

The capacity shortage is most clearly visible in the PJM capacity market. A capacity payment compensates a generator for its commitment to produce electricity whenever the grid operator calls on it. It is separate from payment for electricity actually generated. PJM, the transmission system operator for 13 states and Washington, procured 138,317 MW of capacity in its auction for the 2028/2029 period at a capped price of about USD 325 per megawatt per day. The cap was introduced by federal regulator FERC after prices in the 2024 auction surged by 880%. PJM will pay generators around USD 16.4 billion in total.

Even at the maximum permitted price, however, PJM failed to procure enough capacity. Secured resources fell 6,831 MW short of the reliability requirement. The reserve margin dropped to 14.7%, compared with the preferred 20%, as demand grows faster than supply. The price therefore hit the cap before supply and demand could balance. According to PJM, demand from planned data centers is the main driver of rising offer prices.

Independent market monitor Monitoring Analytics estimates that data centers account for around USD 6.3 billion of total capacity costs. But according to its president, Joe Bowring, those costs are borne not by the data centers themselves but by all customers in the PJM region. Retail electricity prices in the region have risen by about 49% over five years, compared with 33% across the US, according to Heatmap News. State regulatory commissions are therefore facing pressure to set tariffs that prevent households from subsidizing data centers.

Google buys capacity from households

When new capacity cannot be built quickly, it can be obtained by freeing up existing resources. On October 22, Google expanded its agreement with Voltus for a 1 GW virtual power plant in the PJM system. A virtual power plant connects batteries, smart thermostats and industrial equipment at customer sites, temporarily reducing their consumption during periods of high demand. The expansion adds 1,070 MW to the original 100 MW agreed in June, but the companies have not explained how that figure aligns with the stated gigawatt. They have not disclosed the financial terms.

Participating customers will receive bill savings and direct payments funded by Google’s capacity purchase. This is not equivalent to new, continuous generation, however, because the benefit depends on how much consumption participants actually reduce and whether they respond reliably. In any case, the hyperscaler is no longer just a customer; it is beginning to bring its own capacity to the grid.

Microsoft expects its investment to accelerate in the second half of the fiscal year, partly thanks to external capacity deliveries shifted into that period. Amazon is maintaining its investment forecast for this year at USD 220 billion. Its CEO, Andy Jassy, has indicated that spending will be significantly higher in 2027. Federal regulator FERC is waiting for grid operators to report on how they will adjust the process for connecting large loads; the final version is expected to arrive in August.

If Microsoft’s delayed capacity comes online and its free cash flow nevertheless remains positive, it will confirm that the main constraint is not a lack of money but the supply of capacity and hardware. If the next PJM auction also hits the price cap, firm capacity will remain in short supply, and companies will continue looking for solutions like Google’s agreement with Voltus.