The best AI infrastructure stocks, layer by layer

AI infrastructure stocks are the companies that build, power, cool and wire the data centres AI runs in, plus the ones that rent that compute out. You will also see them called AI data center stocks; it is the same list. This page ranks them by what the finance accounts with the best track records are saying, with the graded split beside every score so you can check the work.

This edition lands in a red week. Seven of the eight names on the table below fell, Nebius by 12.9%, and across the group the graded accounts stayed 160 bullish to 21 bearish.

What counts as an AI infrastructure stock

There is no official list. We use a plain test: the company builds, houses, powers, cools or wires the data centres that AI models run in. That gives six layers, and these are the AI infrastructure companies the accounts we track discuss in each:

Chips have their own page, so this one ranks everything else: 30 companies across the other five rows. The big cloud buyers, Meta, Alphabet, Amazon and Microsoft, are the customers of this list and stay off it.

How we rank AI infrastructure stocks

The accounts we follow mentioned 25 of those 30 companies in the last seven days: 528 mentions, 368 bullish to 124 bearish. Most of the 25 sit on a handful of posts. Forgent Power carries a score of 84 on sixteen mentions, Applied Optoelectronics an 80 on seventeen, Coherent an 80 on fifteen. A signal score built on that few posts describes a small room, and small rooms agree with themselves easily.

Two floors decide who gets ranked: at least 20 mentions from the accounts we track in the last seven days, and a graded split you can see. Graded means the bullish and bearish calls from accounts with a real track record: more than ten resolved calls, and more than half of them right. Across the group this week those accounts ran 160 bullish to 21 bearish.

The best AI infrastructure stocks right now, ranked

Every AI infrastructure name with 20 or more mentions this week, ranked by its 7-day score, with what the company does, the graded split this week, and mentions and price over 7 days with the 30-day figure beneath each:

Read the graded column before the score. Vertiv's 84 is the highest score here and it rests on two graded calls. Nebius's 81 rests on 42. The score treats those as close; the evidence behind them is not close. Then read the price column: seven of the eight rows are red on the week. A high score on a falling stock is an open bet, and this week the list is mostly open bets.

The answer: Nebius, and why it comes with Dell attached

Nebius is the deepest room on the list by a wide margin. This week: 87 mentions, 70 bullish to 14 bearish, 39 to 3 among graded accounts. Over thirty days: 449 mentions, 338 to 57, and 166 to 11 among twenty graded accounts. The week's posts were about a roughly 20% increase in its on-demand GPU rates from October 1, a new Madrid data centre going live, and first-place results on an inference benchmark. Reddit posted about it six times this week, all six bearish.

The stock has not agreed. Nebius closed at $268.85 on August 17 and $209.37 on September 16, down 22.1%. It ran from a $199.54 low on September 1 to $243.88 a week later and has given most of that back.

Price

$209.37-22% over the window+5% off the low

Aug 17Sep 16

We have printed this room before. Nebius was the answer on the September 11 best-stock list, and it is 8% lower than it was that day. The graded accounts did not flip on the way down: eleven bearish calls in thirty days, against 166 bullish. You can check that conviction call by call. The price has not paid it yet. The same room has been here before: in the summer it held through a 48% fall in Nebius without turning, and the stock came back. One recovery is a precedent and no guarantee of a second.

Dell is the other half of the answer because it is the one row where the price agreed. It sat at the top of that same September 11 list with a warning that its read had arrived after its run. It has added another 11% since, closed at an all-time high on September 11, and is the only name on this table that is up over both windows. Its room this week shows the split this blog keeps finding: the finance accounts on X ran 19 bullish to 2 bearish, Reddit ran 5 to 11, on posts about selling pressure after the new high and delayed deliveries.

Read the two together. Nebius is where the accounts with the best records are most convinced, and the price has gone against them for a month. Dell is where a smaller graded room, 11 to 2, has been paid. Those are the two ends of what a list like this can tell you.

Who made the Nebius calls

Some of the best graded accounts we track are behind the Nebius read; four of the seven below rank in our top 20 by track record. Their record across every stock we grade, and their bullish and bearish Nebius posts over the last thirty days:

Read the record before the split. A quarter of the graded Nebius calls this month came from one account, Sandeman52, whose record is 35 resolved calls. That is a real record and a short one, and the credibility weighting behind the score treats it differently from ParadisLabs's 1,300. Take that account out and the graded split is still 125 to 9. One of these accounts has a profile on this blog: CKCapitalxx.

Check any of these stocks yourself

Pick one name from the table and open it in Quantral. You will see its score today, every call behind it, and the record of each account that made one, including the bears. It takes a minute and the trial is free.

The AI bottlenecks, by layer

"Bottleneck" is the word the accounts we track use for whichever layer of the generative AI supply chain is shortest at the moment, because that layer gets to raise prices. Over the last thirty days they attached it to 28 different companies, and the argument about where it sits now runs like this:

• Compute rental. This week's candidate. The accounts point to the rate increase at Nebius and to Oracle re-renting older GPUs at about 20% above the previous contracts, and the graded accounts read both as the end of the argument that GPUs lose their value in three years.
• Optics. The bottleneck of the summer. One graded account argues it is still getting tighter, and Lumentum's 15-to-2 graded week says the room has not left.
• Memory. The spring's bottleneck and the most crowded. SanDisk's 19-to-5 is the most contested graded split on the table.
• Power and cooling. The one the room says comes next. One post this month argued the bottleneck is moving from memory to power. Vertiv, Vistra and Bloom carry the calls, and so far they carry few graded ones: Vertiv's two are the whole graded room.
• Packaging and test. The small-room end: wire bonding, advanced packaging and chip test each drew a bottleneck call this month, each from a single account.

A bottleneck call is a claim about pricing power, and it has a shelf life. One account posted a bearish Bloom Energy call this month on exactly that reasoning, the bottleneck easing. The useful habit is to ask which layer a bullish post is betting on, and whether the accounts with records are in that layer yet. In power, they have not arrived in numbers.

The bear case on AI infrastructure stocks: debt

The graded room has few bears this week, 21 calls against 160. The crowd has more, and they are about one thing. Oracle drew 86 mentions, the second-loudest name here, and split 41 bullish to 36 bearish. On X it ran 30 to 8. On Reddit it ran 11 to 28, on posts about a $7.5 billion share sale by its chairman, a round of layoffs, and whether its credit rating holds at investment grade. The graded accounts stayed 15 to 3, pointing at the quarter's cloud growth, a larger backlog, and customers prepaying for capacity. The stock lost 11.4% on the week.

CoreWeave is the same argument in a smaller room, and debt is the standing risk for every neocloud. Graded accounts ran 8 to 0. Reddit ran 2 to 8, and the sharpest bearish post of the week, from an account without a graded record, called its capital structure unwieldy and its leverage excessive. CoreWeave is down 21.4% over thirty days. When a GPU cloud falls this far with its graded room unanimous, the bears are arguing about the balance sheet and the bulls about demand, and both can be right at once. If you own one of these, the debt is the part to research yourself, starting with how to research stocks.

AI infrastructure ETFs hold a different list

If you searched for an AI infrastructure ETF, know that the funds define the theme differently from the accounts we track. The Defiance AI & Power Infrastructure ETF (AIPO) leads with power and grid names: GE Vernova, Eaton, Quanta Services and Vertiv were its four largest holdings on September 17, with Bloom Energy fifth, per the issuer's holdings page. The Global X Data Center & Digital Infrastructure ETF (DTCR) is mostly data-centre landlords and tower companies such as Digital Realty, Equinix and American Tower.

Neither fund's top ten includes a GPU cloud. The part of the trade the graded accounts talk about most, Nebius, IREN, CoreWeave, and the optics names, is the part a fund buyer mostly does not get. The overlap is the power layer: Vertiv scores 84 here on a thin graded room, and Bloom Energy, 48 to 5 among graded accounts over thirty days and up 16.3%, scores 69 on a quiet week of sixteen mentions. We do not rate funds. "AI infrastructure" on a fund label and in a finance feed are two different baskets, so check the holdings before assuming you own the names you have been reading about.

The most talked-about AI infrastructure stocks

The same week, sorted the other way: by how much the accounts we track are talking about a name rather than by what the graded accounts think of it.

Oracle is within one mention of Nebius and scores seventh of the eight. Thirty-six bearish posts from the crowd against three from graded accounts is what a contested 75 looks like. Sort by volume and Oracle leads the list with Nebius. Sort by who has been right before and it drops to seventh. That gap is the reason to read the graded column first.

Where to check the live reading

This page is a snapshot of one red week, and the table will have moved by the time you read it. The order of operations will not: check who is making the call, then how many of them, then the price. If you want a process for turning a table like this into a decision that is yours, start with how to know what stock to invest in. For the reasons a raw sentiment feed will mislead you, how to use social signals without getting played.

Every score on this page, with the calls behind it and the record of each account that made them, is in the Quantral app.

---

Price moves are 7 days (the September 9 close to the September 16 close, five sessions) and 30 days (August 17 to September 16). Mention windows are the 7 and 30 days to September 17. Mention counts and splits cover the accounts Quantral tracks through September 17, 2026; graded accounts have more than ten resolved calls, most of them right. "AI infrastructure" is our own grouping of 30 companies across GPU clouds, servers, networking and optics, memory and storage, and power and cooling; chip companies are ranked separately. Company news attributed to the accounts we track is as they reported it. Scores are the seven-day reading as of September 17, not historical values. ETF holdings are as published by the issuers and change daily. Positions attributed to accounts are their views, not Quantral's. Nothing here is financial advice or a recommendation to buy or sell. Past signals are not indicative of future results.