The earnings season is drawing to a close, and the market has once again rolled out even larger projections for the amount of capital major cloud providers will invest this year to expand artificial intelligence data centers. Goldman Sachs estimates that global related spending could reach $1 trillion by 2026; JPMorgan Chase forecasts the U.S. market alone will see $697 billion; and Bank of America suggests that industry spending could approach $1.2 trillion by 2027. However, money alone cannot clear all obstacles.
The bottleneck is not a shortage of funds. On September 23, 2025, a visit to the OpenAI data center campus in Abilene, Texas, showed multiple buildings still under construction. The campus plans to have a total of eight data center buildings. The computational resources themselves are immensely costly. Memory chip prices are known to keep surging, and strong demand allows NVIDIA to set its own prices for the latest graphics processing units (GPUs) and software. Companies must also purchase land, construct buildings, deploy servers, and install cooling systems. Even as major providers continue to expand production, chip shortages persist. Construction contractors admit there is a significant shortage of skilled workers, making it difficult to meet client deadlines. Furthermore, public resistance to data centers has led to increasing regulatory restrictions: New York has imposed a one-year moratorium on new data center construction, and Texas has initiated reviews for power grid connections. Electricity, however, may be the most severe bottleneck of all.
Bloomberg New Energy Finance estimates that if the industry maintains its current growth trajectory, data centers will face a 19-gigawatt power gap by 2035. "We don't just need equipment and construction permits; we also need sufficient manpower," said George Janarikas, power industry analyst at Canaccord Genuity, in an interview. "At the same time, public opposition to data center construction is ongoing. I'm sure you've seen the news: protests, construction moratoriums, and project cancellations are everywhere. Considering all these factors, it seems to us that the grand goal of data center companies securing enough electricity to train large language models is unlikely to happen as quickly as expected." Wood Mackenzie recently disclosed that data center power service providers are trying to increase their approval rates by submitting duplicate applications to multiple power companies. The energy analysis firm points out that due to a large number of "inflated duplicate applications" and submissions from many inexperienced operators, the final power capacity approved by grids and utilities may only account for 28% of total applications.
These signals suggest that the construction of data centers will at least proceed much slower than the most optimistic market expectations, characterized by intermittent and uneven expansion—and this is the optimistic scenario. In the worst-case scenario, generative AI clients might shift to lower-cost open-source weight models or adjust their business models to adapt to constrained computational resources. At that point, the supply-demand dynamic would reverse, turning the bottleneck into overcapacity, impacting the entire supply chain from GPUs to natural gas generators. No such signs have emerged yet. Major cloud providers have repeatedly emphasized during the earnings season that demand far exceeds supply. Amazon predicts AWS business revenue could reach $1 trillion, with CEO Andy Jassy stating that the "scale of demand for computational resources in 2028 is truly astonishing"; CoreWeave CEO Michael Intrator mentioned on an earnings call that short-term capacity is "basically already sold out." This reflects a systemic supply-demand imbalance that will persist for several years and is unlikely to be resolved quickly.
Who will consistently benefit from this supply-demand imbalance? The winning sectors are broad: GPUs, central processing units, memory chips, servers, HVAC equipment, power equipment, power generation companies, and utilities have all seen successive upswings. Some are short-lived, but many industries continue to show strong trends. Janarikas remains bullish on the stocks he covers, such as nuclear power companies Oklo (OKLO) and NuScale Power (SMR). Ivana Delevska, founder and chief investment officer of ETF firm Spear Invest, suggests that chip equipment and optical communication manufacturers will be the next wave of beneficiaries, with the cycle already underway. A risk to note: this pressure will not be immediately apparent, but it is worth monitoring closely—if the expansion of infrastructure slows, the upstream suppliers, currently eagerly chased by investors as the so-called "picks and shovels" sellers, could face significant impact on their performance.
Comments