AI's Real Bottleneck Is Power: How an AI Agent Screened 25 Energy and Grid Stocks with Stocklake

In October, Amazon was reported to be taking early delivery of about $5 billion of Nvidia GB300 racks before some of its data centers had the power to run them, and storing the hardware until capacity was ready. In the same weeks, Elon Musk talked about roughly 13 GW of AI compute by 2027, a 440,000-GPU cluster by mid-2027, and another 20 GW of power and cooling infrastructure next year.

Put the two together and the constraint is plain. For a growing share of AI buildouts, the limit is no longer the chip. It is the electricity, the grid connection and the cooling. Hardware arrives faster than the power to switch it on.

That leaves a practical question: who supplies that power, and which of those companies looked reasonable to own after the selloff? I gave the question to an AI agent with Stocklake connected and watched how it worked through it. This post is what it found on October 20, what made the answer useful, and what each company's next earnings report has to show for the ranking to hold.

Four layers of the power stack

The agent's first move was to split “AI power” into the four places money gets spent, from making the electricity to cooling the rack.

• Generators and nuclear: VST, CEG, NRG, TLN, CCJ, OKLO, SMR
• Grid and turbines: GEV, ETN, PWR, HUBB, BE
• Data-center power and cooling: VRT, NVT, MOD, FIX, EME
• Europe: PRY.MI, ENR.DE, SU.PA, ABBN.SW, LR.PA, NEX.PA, SIE.DE, RWE.DE

The European names are covered under their home-exchange tickers (Milan, Xetra, Paris, Zurich), in the same data shape as the US ones. That mattered later: comparing Prysmian with Eaton took one call, not a separate research project.

How the agent worked through it

1. Read the tape first. get_market_assessment (a Pro tool) came back cautious: regime CAUTIOUS, Fear & Greed at 28, oil at $97.81 after a 10% week, the 10-year yield at 5.12%. That set the tone. This was not a market paying for momentum, so the agent looked for quality that had been marked down rather than for breakouts.

2. One batch snapshot. A single get_stocks pass returned price, forward P/E, analyst targets, RSI and position against the 200-day average for all 25, plus three forensic accounting scores: Altman Z (balance-sheet distress), Piotroski F (financial strength) and Beneish M (an earnings-quality screen). get_stocks itself is free; the forensic scores in it are a Pro field.

3. Parallel deep dives. The agent split the work across four sub-agents, one per layer. Each pulled research, five years of statements plus quarterlies, earnings history and the trend in analyst targets for its names (all Pro; the free tier returns the latest fiscal year of statements), then scored each one on Quality, Forward-earnings value and Price attractiveness.

4. Synthesis. The four reports were merged into one ranked list and an earnings calendar.

What the data said on October 20

The whole AI-power complex had sold off together. Of the 16 names that ended up ranked or cut, 15 closed October 20 below their 200-day average, and most sat between 20% and 50% under their 52-week highs. Strong balance sheets had been marked down with the weak ones.

That is when a screen like this earns its keep. When a theme sells off as a block, the differences between companies stop showing up in the price, but they are still in the financials. One caution applied to every name: none of these downtrends had turned yet. Oversold describes where a stock is, not that it has stopped falling.

The top 10, ranked by our screen

How to read it: Comfort Systems is first because it scored well on all three axes at once: net cash, a margin that is still rising, and a growth-adjusted valuation under 0.5 after a third came off the price. GE Vernova has arguably the best order book in the group, but at 35 times forward earnings on a single-digit operating margin, the price score held it to ninth. The screen ranks what you'd be paying for, not just how good the business is.

What the screen left out

Being in the right theme was not enough. Six names were cut, each for a reason the numbers showed:

• Bloom Energy: 45x forward earnings on trailing EPS of $0.76, and the share count up 26% in a year (233.7M to 293.4M).
• Talen: a GAAP loss (trailing EPS −$4.13) and $9.6B of debt.
• NRG: 9x forward looks cheap only because forward EPS ($10.90) is nearly three times trailing GAAP EPS ($3.82). $23.5B of debt.
• Cameco: 43x forward and 136x trailing earnings.
• RWE: forward EPS (€3.36) below trailing (€4.51), and negative free cash flow of −€11.1B from the build-out.
• Modine: negative free cash flow (−$74M) and a 9.0% operating margin, 55% off its high.

What made the answer useful

A list of ten tickers is easy to get. What I didn't expect was how much of the reasoning behind each line the agent could show, and check, in the same session.

It read the last earnings report, not just the price

get_stock_research and get_earnings_intelligence put each name's last quarter next to its multiple. For Comfort Systems: Q2 EPS of $12.53 against a $10.38 consensus, revenue of $3.26B against $2.94B, and a record $14.1B backlog, up 73% on the year. That is the context that makes 23x forward look different from 23x on a company that just missed.

It caught the debt the headline skips

get_stock_financials returns each quarter's balance sheet on Pro, so the agent could see what an acquisition did to it. Eaton's total debt went from $10.5B at the end of 2025 to $21.8B at the end of March 2026, around the Boyd Thermal purchase. Prysmian's Atkore deal closed on October 19, partly paid for with €850M of new shares sold at €121, well above where the stock closed on October 20. Neither knocked the company out. Both moved it down the list.

It didn't take a cheap forward P/E at face value

Forward estimates usually use adjusted earnings; trailing P/E uses GAAP. Divide one by the other and you get “growth” that may just be an accounting difference. NRG was the clearest case: 9.0x forward against 25.6x trailing, because forward EPS of $10.90 is 2.9 times trailing EPS of $3.82. The agent flagged the gap instead of calling the stock a bargain.

A forensic score sent it to the receivables line

nVent's headlines were good: analysts were raising targets while the stock fell, and JPMorgan started coverage at overweight on October 20. But its Beneish M-score sat at −1.779, right at the −1.78 line where the model asks for a closer look. The agent pulled the statements to see why: in fiscal 2025, receivables grew 62% ($528M to $854M) while revenue grew 29.5% ($3.01B to $3.89B).

That isn't evidence of anything wrong. Fast-growing businesses build receivables, and the most recent quarter cut the other way: in Q2 2026 receivables were up 29.6% on the year against revenue up 52.8%. But it is the reason nVent made the list at tenth with a question mark rather than higher, and it is a specific line to read in the next report.

Europe priced the same theme lower

Prysmian at 17.0x and Schneider at 19.6x forward against Eaton at 23.4x and GE Vernova at 35.2x, for exposure to the same grid and electrification spend. Schneider had also just reported 16.9% organic growth, against 13.8% expected. Because the home listings come back in the same shape as US stocks, the agent could put them in one table without converting anything by hand.

The agent kept watching

Nine of the ten were due to report within about three weeks of the screen. So I asked the agent for one more thing: give me each earnings date, remind me on the day, and when the numbers are out, tell me how the stock looks now against what the screen expected.

The first reminder is for Comfort Systems, which reports on October 22. This is what it looks like, built only from what the agent knows going in:

After each report, the follow-up runs the same calls again: get_earnings_intelligence for the result against expectations, get_stock_financials for the new quarter, get_analyst_trend for how targets moved, and get_stocks for where the price and multiple ended up. The table below lists what each of those checks looks for.

What each report has to show

A ranking is a claim about what the next numbers will look like. Here is what each report has to show for the screen's reasoning to hold.

The six that were cut report on these dates: Bloom Energy (reports Oct 22), Modine (Oct 27, estimated), Cameco (Oct 30), Talen (Nov 3), NRG (Nov 5, estimated) and RWE (Nov 11).

What I took from it

• Read the regime before the stock. A cautious tape changed what the agent looked for: marked-down quality, not momentum.
• The forensic scores are worth a look every time. One number at the edge of a threshold led straight to the line in the statements that mattered.
• Check where forward earnings come from. A forward P/E built on adjusted EPS against GAAP trailing EPS can make almost anything look cheap.
• Look abroad for the same theme. The European names offered the same exposure at lower multiples, and they were one call away.
• A screen is a starting point with a date on it. The useful part was setting up the follow-through: what each report has to show, checked on the day.

Free vs Pro for this workflow

You can start this on the free tier. The parts that made the answer worth reading need Pro.

Try it on your own theme

Everything here came from Stocklake MCP tools that any AI agent can call: get_market_assessment, get_stocks with forensic scores, get_stock_research, get_stock_financials, get_earnings_intelligence and get_analyst_trend, across 3,500+ US and international symbols. Add Stocklake as a connector with the server URL https://api.stocklake.dev/mcp, then give your agent its own headline to work from. Setup steps for each client are in the docs.

Frequently asked questions