Large institutional investors cooled on the physical infrastructure of the artificial-intelligence boom in the third quarter, even as they continued to add modestly to AI and semiconductor positions, according to a Reuters review of nearly 6,400 SEC 13F filings.
The filings, which large money managers must submit within 45 days of the end of each quarter, showed more funds closing out data-center and utility holdings than opening new ones. Reuters found that 138 funds fully exited data-center positions, compared with 119 that established new ones. Across eight major names in the sector, the total number of institutional holders slipped.
That marks a subtle but notable shift. For much of the past two years, the market's AI enthusiasm spilled well beyond chipmakers into the companies that supply the power, land, cooling and construction needed to run massive computing clusters. Data-center landlords, electrical equipment makers and utilities with exposure to rising electricity demand all benefited from that spillover.
Why data centers and utilities became an AI proxy
The logic was straightforward. Training and running large AI models requires enormous amounts of computing power, and that computing power has to live somewhere. Data-center real estate investment trusts, or REITs, own and lease the buildings. Equipment suppliers provide the power management, cooling systems and backup infrastructure. Utilities stand to sell more electricity as those facilities come online.
For investors who wanted exposure to the AI theme without concentrating entirely in a handful of mega-cap technology stocks, these names offered a way to play the buildout. They also appealed to investors looking for steadier cash flows than pure-play chip stocks typically provide.
But infrastructure trades come with their own risks. Data-center construction is capital-intensive and sensitive to interest rates, since these companies often carry significant debt. Utilities face regulatory approval processes that can slow or reshape projects. And equipment suppliers are tied to the timing of construction cycles, which can be lumpy from quarter to quarter.
What the filings actually show
It is important to be careful about what 13F data can and cannot tell you. These filings capture long positions in US-listed equities held by institutional managers with more than $100 million in assets. They do not show short positions, derivatives exposure, or positions in non-US securities. They are also backward-looking, reflecting holdings as of the end of the quarter, not today.
Even so, the direction of travel matters. When more funds close a position than open one, it suggests the marginal institutional buyer is stepping back. That does not necessarily mean the underlying business is deteriorating. It can simply mean the trade has become crowded, valuations have stretched, or portfolio managers are rotating into other opportunities.
The Reuters review also found that buying in AI and semiconductor names remained slightly positive. In other words, institutions did not abandon the AI theme wholesale. They trimmed the infrastructure layer while maintaining or modestly adding to the chip and AI layer.
That distinction is worth noting. It suggests a rotation within the AI trade rather than an exit from it. Investors may be gravitating back toward the companies at the center of the technology story, where earnings growth has been more visible, rather than the peripheral beneficiaries whose gains depend on capital spending plans materializing on schedule.
What it means for investors
For everyday investors, the takeaway is less about copying what big funds did and more about understanding the shape of the trade. When a theme becomes popular, money tends to fan out from the core companies into adjacent ones. That broadening can create opportunities, but it can also push valuations in secondary names beyond what their fundamentals justify.
If institutions are now consolidating back toward the core, that is a signal worth watching. It may indicate that the market is starting to differentiate between companies with direct AI revenue and those with a more indirect, longer-dated connection to the buildout.
Investors holding data-center or utility exposure as an AI proxy might ask themselves a few questions. Is the investment thesis based on confirmed contracts and signed leases, or on expectations of future demand? How sensitive is the company to interest rates and financing costs? And how much of the expected growth is already reflected in the share price?
It is also worth remembering that 13F data is a lagging indicator. By the time filings are public, the quarter is over and positions may have changed. Treating them as a real-time signal is a mistake. Treating them as one input among many, alongside earnings reports, guidance and capital spending announcements, is more sensible.
Related coverage on the AI infrastructure theme has been active across markets. New US AI chip rules have reshaped semiconductor ETF flows, while companies such as Prysmian have cited data-center demand in their own strategic decisions. Analysts have also flagged how data-center margins underpin broader targets at construction firms, and how AI data-center pitches are shaping corporate outlooks even when headline numbers disappoint.
The broader backdrop matters too. Interest rates remain a key variable for capital-intensive sectors. If borrowing costs stay elevated, the economics of building and operating data centers become less attractive, and utilities face higher financing costs for grid upgrades. That could weigh on the very names that benefited most from the AI buildout narrative.
For now, the filings suggest institutions are being more selective. The AI story is not over, but the easy money in every adjacent name may be. Investors would do well to focus on where actual revenue is being generated, rather than where it is merely expected.


