• AI inference costs have fallen 47% per quarter since 2023, a 13-fold annual decline, according to Epoch AI—six times faster than computing costs and 18 times faster than lithium battery costs.
  • U.S. and Chinese tech firms are projected to spend nearly $1 trillion on AI infrastructure this year, raising doubts about whether the spending can generate adequate returns.
  • While demand for AI services is strong, the timing mismatch between upfront investment and future revenue remains a critical risk.

A Paradox of Plenty

The cost of delivering a fixed level of AI capability is collapsing. Epoch AI’s September 22 report estimates that the cheapest available way to achieve a given benchmark performance has fallen about 47% per quarter since 2023—equivalent to a 13-fold annual reduction. That’s roughly six times faster than the decline in computing costs and 18 times faster than lithium battery costs. The finding measures performance-adjusted inference costs across five principal benchmarks covering mathematics, science, and puzzles, accounting for tokens consumed by reasoning models.

Yet even as AI output gets cheaper, the industry is pouring unprecedented sums into infrastructure. CNN (WBD) reported on October 13 that U.S. companies are on track to invest approximately $800 billion in AI data centers this year, while listed Chinese firms are expected to spend $139 billion—a combined total of roughly $939 billion. Deutsche Bank (DB) research cited in the same report puts capital expenditure by Alphabet (GOOGL), Amazon (AMZN), Meta (META), Microsoft (MSFT), and Oracle (ORCL) at almost $800 billion, though that figure reflects corporate capex, not exclusively AI spending.

The result is a widening gap between the cost of AI services and the cost of building the capacity to deliver them. “What institutional investors like us are really focused on is regulatory stability,” said Andrea Valeri, Blackstone (BX)’s country chairman for Italy and chief investment officer for Blackstone Credit and Insurance’s private credit business in Europe and APAC, at Bloomberg’s Future of Finance conference in Milan. “Italy in this regard has been on a very steady growth trajectory.” While Valeri’s comments addressed Italy’s investment climate, the sentiment echoes a broader theme: investors are watching for signs that the AI buildout will yield durable returns.

Demand Is Real, but So Are the Costs

Evidence of commercial demand is mounting. Microsoft reported fiscal 2026 revenue of $331.8 billion, up 18%, with operating income of $155.2 billion, up 21%. Its fiscal fourth-quarter revenue reached $90 billion, up 18%. Amazon’s AWS unit saw second-quarter 2026 revenue rise 37% year over year to $42.2 billion, signaling expanding cloud demand—though AWS revenue is not exclusively AI revenue.

Still, the scale of investment raises questions about whether returns will justify the outlay. “The danger is a timing mismatch: companies incur construction and equipment costs before the resulting services generate enough cash,” Deutsche Bank’s reported analysis warns, highlighting pressure on hyperscaler free cash flow. The bank also cautioned about an “AI air pocket” in which spending arrives before expected revenue.

Physical constraints add to the challenge. Power availability is becoming a major competitive factor. CNN reports that U.S. projects face an aging grid and connection constraints, while China is attempting to place computing facilities near plentiful renewable energy as part of its “Eastern data, Western compute” strategy. China’s weaker access to advanced chips remains a countervailing disadvantage, but Beijing announced a $560 billion information-infrastructure plan in September covering the next five years.

Accounting Changes Muddy the Picture

Microsoft’s latest earnings discussion identified an important accounting change: shifting some future data-center leases from finance leases to operating leases brings its calendar-2026 capex-and-finance-lease outlook to approximately $175 billion, while underlying investment expectations remain unchanged. The company also extended the estimated useful life of office and data-center buildings from 15 to 25 years. These adjustments complicate comparisons between headline spending estimates.

Epoch’s findings also suggest that the cost of newly achieved state-of-the-art performance initially falls particularly quickly—an average 66% per quarter across its five main benchmarks—before slowing as that performance level matures. The proposed explanation is that competitors rapidly catch up and erode an initial pricing premium. That dynamic could make it difficult for providers to sustain premium pricing.

What to Watch

The debate over whether AI investment will pay off is far from settled. Epoch cautions that its dataset spans only about three years and is incomplete, so the 47% figure should be treated as a rough estimate rather than a universal law. Real users may save less than the estimates suggest because they do not continually switch to the cheapest suitable model.

The most useful tests, according to analysts, include incremental revenue and operating cash flow, utilization of completed data centers, revenue growth relative to price declines and operating costs, equipment replacement needs and access to power, and whether enterprise customers can demonstrate durable productivity gains.

Nvidia (NVDA)’s declining share of China’s data-center AI-chip market—falling to 34% in the first half of 2026, according to IDC—shows how policy is redirecting infrastructure purchases. Meanwhile, China’s renewable-powered computing buildout shows energy becoming a strategic asset. None of these developments individually settles the ultimate return-on-investment question, but together they frame the central challenge: can an industry spending nearly $1 trillion a year generate enough value to justify the cost?

Correction: An earlier version of this article misstated the annual decline in AI inference costs. It is 13-fold, not 13% annually.