The AI buildout is no longer just a semiconductor story. Power, water and grid access are becoming part of the same investment test.
The latest warning on AI infrastructure is not that companies have stopped spending. It is that the physical world is starting to matter as much as the chips. A new United Nations University report puts hard numbers around that problem, estimating that data centers linked to AI could help drive global data center electricity use to about 945 terawatt-hours by 2030, a level roughly comparable to Japan's current electricity consumption.
That figure is striking on its own, but the market reaction around Broadcom this week shows why it matters to investors. Broadcom reported record second-quarter revenue of $22.19 billion for the period ended May 3, 2026, with AI semiconductor revenue of $10.8 billion, up 143% from a year earlier. Yet the shares still fell sharply after the company left a key long-range AI chip sales forecast unchanged and guided current-quarter AI chip revenue a little below some analyst expectations.
This is what a crowded trade looks like when expectations get too clean. Broadcom is not a weak company. Its custom AI accelerators and networking products sit close to the heart of hyperscaler spending plans, and CEO Hock Tan said the company expects AI semiconductor revenue to reach $16 billion in the current quarter. But when valuations assume a perfect march upward, even strong numbers can look ordinary if they do not raise the ceiling.
According to a recent Reuters report, Broadcom stuck with its long-range forecast of $100 billion in sales from AI chips and said it now expects to ship more than 10 gigawatts worth of AI chips in 2027. That language matters because the industry is increasingly measuring compute in power terms, not just server counts or chip units. The bigger the model, the larger the cluster. The larger the cluster, the harder it becomes to ignore electricity procurement, cooling and grid connection timelines.
The International Energy Agency has already framed the problem clearly. Data centers used around 415 TWh of electricity in 2024, about 1.5% of global consumption, and that demand is projected to more than double by 2030. The United States accounted for the largest share in 2024, followed by China and Europe. In some local markets, the pressure is far larger than the global percentage suggests, because data centers tend to cluster where fiber, land, tax incentives and power access line up.
For chipmakers, this changes the quality of growth. Demand for accelerators may remain strong, but customers cannot install unlimited compute if substations, transmission lines and cooling systems move slower than purchase orders. A hyperscaler can commit billions to custom silicon, but the return on that spend depends on turning those chips into working capacity. That makes energy availability a direct input into revenue timing, not a side issue for sustainability teams.
The environmental bill is broader than carbon
The UNU-INWEH report pushes the debate beyond electricity alone. It warns that AI's environmental footprint includes water used for cooling and power generation, land required for energy infrastructure and supply chains, and carbon emissions tied to the grid mix. The report estimates that by 2030, AI-related water demand could be comparable to the basic domestic water needs of 1.3 billion people in Sub-Saharan Africa.
That comparison is not meant to suggest AI is useless or that data centers should simply stop being built. The practical point is sharper. Different energy choices create different trade-offs. Lower-carbon power can still increase land or water pressure depending on how it is produced, stored and delivered. A clean-looking procurement contract may not solve the local burden if the surrounding grid, water system or community absorbs the strain.
This is where the investment story and the public policy story meet. Big technology companies want faster permitting, more power purchase agreements, new nuclear options, gas backup, geothermal projects and larger renewable portfolios. Utilities want certainty before building infrastructure that customers may or may not fully use. Regulators want growth, but they also have to decide who pays for upgrades when a single facility can demand as much power as a town.
Broadcom's sell-off does not prove the AI cycle is ending. It does show that the easy phase of the trade is narrowing. Investors once rewarded almost any company with exposure to AI infrastructure because the direction of spending was obvious. Now the question is more specific: which companies can convert demand into profitable revenue while their customers navigate power constraints, supply chains and local resistance?
The next phase of AI will still need chips, memory, networking equipment and software. It will also need transformers, transmission lines, cooling systems and credible energy plans. That is a less glamorous story than model breakthroughs, but it may decide who actually captures the economics. Watch the capex guidance from hyperscalers, but watch the grid connection queues too. They are becoming part of the same signal.
Also read: China has turned brain implants into a commercial medical race • Seattle is moving to pause new data centers for one year • AI leaders ask Congress to tighten synthetic DNA screening
