The AI Trade Moved to the Grid: What 13Fs Show About the Nuclear Power Bet

The cleanest way to understand the last eighteen months of institutional positioning is this: the AI trade stopped being about chips and started being about electricity. Not because the chip thesis broke, but because the bottleneck moved. You can buy accelerators with money. You cannot buy grid interconnection with money, at least not quickly, and the funds that figured that out early rotated into the companies that already own generation.

Nuclear is where that rotation concentrated. And it is a useful case study in what 13F filings are good for — and where they fail badly.

The contracts that created the trade

The thesis is not speculative. It is contractual, and the contracts are public.

Constellation signed a 20-year agreement to supply Microsoft from the planned restart of the 835 MW Crane Clean Energy Center, and a second 20-year deal with Meta for 1,121 MW from the Clinton Clean Energy Center. Amazon expanded its offtake from Talen's Susquehanna plant to 1,920 MW running through 2042. In January 2026, Vistra and Meta announced agreements covering existing nuclear capacity in the PJM grid, including an option for Meta to take power from a potential new 300 MW small modular reactor at a Vistra site.

Those are multi-decade, investment-grade counterparties buying baseload power. For an equity analyst, that converts a commodity-exposed merchant generator into something closer to a contracted utility with a growth option attached. That re-rating is most of the move.

What the filings actually show

Institutional flow into this theme is real but messier than the narrative suggests. Oklo — the pre-revenue SMR developer with roughly $3 billion in cash, no commercial plants operating, and a valuation built entirely on future data center contracts — is the sharpest example.

In Q2 2026, Marshall Wace opened a new position of about 1.07 million shares, an estimated $55.8 million. In the same quarter, State Street cut roughly 948,000 shares, down 28.5%, about $49.6 million. Morgan Stanley had added 797,647 shares in Q1, a 42.3% increase.

Read those three lines together and you get the honest picture: this is not consensus. It is a fast-money position being built by discretionary funds while at least one large holder trims. And the State Street number carries an important asterisk — a passive-heavy filer's "sell" is often index mechanics, not a view. Distinguishing intentional adds from mechanical ones is the difference between a signal and a coincidence, which is what the flow view is built to separate.

The broader sleeve tells a similar story. The Global X Uranium ETF held roughly $6.3 billion in net assets across nearly 60 holdings as of August 2026 — meaningful, but small relative to the megacap AI complex that drives the demand thesis.

The risk 13Fs cannot price

On the regulatory side, the trade has a single point of failure, and it showed up already.

FERC rejected a request to increase the power the Susquehanna plant could dispatch to Amazon's adjacent data center campus. Constellation fell more than 12% on the news, Vistra roughly 3%, Talen more than 2%. Nothing about the underlying reactors changed. What changed was the regulator's willingness to let a hyperscaler pull existing baseload off the shared grid for private use.

That is the actual thesis risk, and no quarterly holdings filing will ever warn you about it. A 13F tells you a fund owned 1.07 million shares on June 30. It does not tell you whether the fund had hedged the regulatory outcome, whether the position was one leg of a pair trade against a utility short, or whether it was sold on July 1. Filings arrive 45 days after quarter end. FERC does not wait 45 days.

The macro backdrop compounds this. August payrolls came in at 162,000 against roughly 53,000 expected, pushing rate expectations higher and bond yields with them. Capital-intensive, long-duration power assets are precisely the wrong thing to own into a higher-for-longer repricing — a tension worth checking against how filers are positioned overall on the macro consensus view.

How to use the data anyway

The useful question is not "who owns Oklo." It is whether a name is held by many independent filers or by a handful making the same trade — a check the overlap tool answers in seconds. Six unrelated funds independently reaching the same conclusion is a rotation. Six funds crowding one thesis with the same regulatory dependency is a correlated position that unwinds together, exactly as the Susquehanna ruling demonstrated.

Nuclear may well be the durable second act of the AI trade. But treat Q2 filings as archaeology, not a live feed — and treat the regulatory calendar as the thing that actually moves these stocks.

---

Data sourced from public SEC 13F filings. Educational research only — not investment advice.