For the past two years, tech giants have been racing to build the digital backbone of the artificial intelligence boom, spending tens of billions of dollars on chips and new data centers. Now, as construction cranes swing and server racks pile up, early signs of restraint are emerging.
Amazon Web Services (AMZN) has paused negotiations on some new leases for data centers, particularly overseas, according to a Wells Fargo report published last week. Microsoft (MSFT) made a similar move in February, canceling plans for approximately two facilities’ worth of computing power, according to TD Cowen.
Both companies have emphasized that signed deals remain in place and described the moves as normal capacity management. Microsoft said it was still set to spend $80 billion in its fiscal year that ends in June. AWS’ vice president of global data centers wrote on LinkedIn that “there haven’t been any recent fundamental changes in our expansion plans.”
The AI buildout is real, but the pace may be shifting. While cloud providers publicly maintain that expansion plans are unchanged, the recent lease pauses suggest a more cautious recalibration behind the scenes — a signal that the AI boom may not be advancing at the unrelenting speed companies like Amazon and Microsoft have projected.
One explanation for the shift is simple overcommitment. A UBS (UBS) report last week concluded that Microsoft’s pullback likely stems from overcommitting during the initial AI rush, according to CNBC. Microsoft’s leased capital expenditures jumped 6.7 times in two years, with lease obligations now totaling roughly $175 billion, UBS wrote. With a better idea of how the tech is actually being used and what the power needs are, Microsoft is now cutting early-stage projects that no longer make immediate sense. UBS said it found little support for the idea that a sudden demand lull was responsible for the change in strategy.
Cost pressures inside the AI ecosystem are adding up. A single query to OpenAI’s most advanced models can cost up to $1,000 in computing power alone. Despite charging $200 per month for premium access to ChatGPT, OpenAI CEO Sam Altman said in January that the subscription service is not yet profitable.
Even tech executives are acknowledging the gap between hype and outcomes.
Microsoft CEO Satya Nadella recently conceded that, so far, AI has not yet produced much measurable value. His comments reflect broader skepticism about whether generative AI can deliver sustainable returns — or whether infrastructure spending is getting too far ahead of real-world demand.
External factors are compounding the challenge.
President Donald Trump’s proposed tariffs have raised the prospect of dramatically higher costs for imported equipment, while tech stocks have been under pressure amid broader market volatility. At the same time, many regions are facing grid constraints that limit the ability to add new data centers, and local opposition to large-scale facilities is intensifying as communities push back against rising power demands, land use, and water consumption.
The scale of future AI infrastructure could amplify these strains significantly.
If current trends continue, the leading AI data centers could each cost $200 billion, contain two million AI chips, and require as much power as nine nuclear reactors by 2030, according to a recent study from researchers at Georgetown University, Epoch AI, and the RAND Corporation.
The pullback comes at a time of unprecedented investment.
More than 500 data center facilities are currently in the planning and construction phases worldwide, according to Synergy Research Group. Amazon, Microsoft, and Google (GOOGL) now account for 59% of all hyperscale data center capacity, the firm said. Each company has committed to pouring tens of billions of dollars into capital expenditures, largely aimed at powering generative AI models.
Meanwhile, the gap between AI infrastructure spending and AI-generated revenue continues to widen. In a June 2024 analysis, Sequoia Capital partner David Cahn estimated that the mismatch had ballooned into a $600 billion hole — up from a $200 billion gap just nine months earlier.
For now, cloud providers are sticking to their public message: Expansion plans remain on track. But the quieter signals — paused leases, early project cancellations, rising costs, and a fast-moving competitive landscape — suggest a more complicated reality behind the scenes.