Eighteen months inside the AI power trade — the theme that turned a Texas utility into a momentum stock, broke in a single session, fully recovered, and then quietly gave it all back. Five charts, real prices, and the part almost nobody writes down.
Vistra, Jan '24 → peak
+472.4%
$38.07 to $217.92 close, 22 Sep 2025
One session, 27 Jan '25
−28.3%
The day the theme was stress-tested
Vistra since 24 Jan '25
−12.3%
18 months later, still below
Nvidia, same window
+49.9%
The "obvious" trade won
Act one · The theme
The AI story everyone traded was compute. The story underneath it was electricity. Data centres need power, power needs generation, and generation is the one part of the stack you cannot spin up in a quarter. You can poach a researcher. You cannot poach a substation.
That reasoning is what a desk calls a second-derivative trade — you skip the crowded thing and buy the bottleneck feeding it. It worked spectacularly. Between the first session of 2024 and the September 2025 peak, Vistra closed up 472.4%, more than doubling the return of Nvidia over the same stretch.
The power trade beat the chip trade for nearly two years
Month-end closing price, indexed to 100 at January 2024. One axis — every series measured the same way.
Source: Massive.com month-end closes, Jan 2024 – Jul 2026. Indexed by Main St. Investors. Past performance is not indicative of future results.
| Month | VST | NVDA | SPY |
|---|---|---|---|
| Jan '24 | 100.0 | 100.0 | 100.0 |
| Feb '24 | 122.3 | 124.4 | 103.7 |
| Mar '24 | 169.1 | 133.0 | 104.9 |
| Apr '24 | 182.9 | 134.2 | 103.4 |
| May '24 | 222.4 | 165.7 | 106.7 |
| Jun '24 | 200.6 | 190.2 | 112.2 |
| Jul '24 | 162.6 | 162.1 | 107.8 |
| Aug '24 | 191.8 | 180.4 | 114.0 |
| Sep '24 | 310.7 | 188.8 | 115.9 |
| Oct '24 | 268.3 | 204.7 | 115.5 |
| Nov '24 | 358.8 | 209.0 | 121.9 |
| Dec '24 | 364.4 | 218.4 | 119.7 |
| Jan '25 | 377.2 | 181.5 | 121.7 |
| Feb '25 | 300.0 | 188.8 | 120.2 |
| Mar '25 | 220.1 | 142.5 | 102.2 |
| Apr '25 | 312.6 | 173.1 | 114.6 |
| May '25 | 360.4 | 204.2 | 119.2 |
| Jun '25 | 431.4 | 240.8 | 126.5 |
| Jul '25 | 467.0 | 262.6 | 125.8 |
| Aug '25 | 422.0 | 252.4 | 130.9 |
| Sep '25 | 453.4 | 283.6 | 135.4 |
| Oct '25 | 422.7 | 306.1 | 138.0 |
| Nov '25 | 375.2 | 275.7 | 138.7 |
| Dec '25 | 370.9 | 285.4 | 138.2 |
| Jan '26 | 355.4 | 288.9 | 140.0 |
| Feb '26 | 390.3 | 267.8 | 138.8 |
| Mar '26 | 339.3 | 268.1 | 132.7 |
| Apr '26 | 348.6 | 300.0 | 145.8 |
| May '26 | 333.9 | 310.0 | 149.2 |
| Jun '26 | 339.1 | 294.5 | 150.7 |
| Jul '26 | 376.1 | 322.8 | 151.3 |
Act two · The test
A cheaper training method landed and the market repriced the entire AI capex chain in one session. Here is the part that matters: the "safer" second-derivative trade lost more than the thing it was a derivative of.
Nvidia — the crowded, obvious, most-owned name in the market — fell 17.0%. Vistra fell 28.3%. Constellation fell 20.9%. The S&P 500 barely moved. Crowding is not measured by how famous a stock is. It is measured by who has to sell when the story wobbles, and the power names were owned by people whose thesis was one sentence long.
One session, four assets
Closing price change, 24 Jan 2025 → 27 Jan 2025. Vistra highlighted.
Source: Massive.com daily closes. Single-day price change, unadjusted for dividends.
| Asset | Ticker | Change |
|---|---|---|
| Vistra | VST | -28.3% |
| Constellation | CEG | -20.9% |
| Nvidia | NVDA | -17.0% |
| S&P 500 | SPY | -1.4% |
The tell was never in the story. It was in who else was telling it.
— The DeskAct three · The round trip
Five months after the crash, Vistra was back above where it started. Two months after that it printed a new high of $217.92. Everyone who held through the drawdown was vindicated — and that vindication is precisely what made the exit impossible to take.
Then it went sideways, then down. As of 22 July 2026 the stock trades at $167.65 — below the level it held the day before the crash, eighteen months earlier. The drawdown you survived was not the risk. The round trip you sat through was.
Vistra: crash, full recovery, new high, round trip
Month-end closing price with key sessions marked. Reference line: $191.11, the close before the 27 Jan 2025 repricing.
Source: Massive.com and FMP daily closes, Jan 2025 – Jul 2026. Latest price 22 Jul 2026.
| Month | Vistra close | vs $191.11 |
|---|---|---|
| Jan '25 | $168.03 | -12.1% |
| Feb '25 | $133.66 | -30.1% |
| Mar '25 | $117.44 | -38.5% |
| Apr '25 | $129.63 | -32.2% |
| May '25 | $160.57 | -16.0% |
| Jun '25 | $193.81 | +1.4% |
| Jul '25 | $208.54 | +9.1% |
| Aug '25 | $189.11 | -1.0% |
| Sep '25 | $195.92 | +2.5% |
| Oct '25 | $188.30 | -1.5% |
| Nov '25 | $178.86 | -6.4% |
| Dec '25 | $161.33 | -15.6% |
| Jan '26 | $158.35 | -17.1% |
| Feb '26 | $173.89 | -9.0% |
| Mar '26 | $150.33 | -21.3% |
| Apr '26 | $157.84 | -17.4% |
| May '26 | $160.23 | -16.2% |
| Jun '26 | $158.63 | -17.0% |
| Jul '26 | $167.65 | -12.3% |
The scoreboard
Measured from the last close before the theme was tested, the trade that was "obviously crowded" is up half. The bottleneck trade is down. And Constellation — the cleanest nuclear expression of the same idea — never came back at all.
Total price change, 24 Jan 2025 → 22 Jul 2026
Closing price basis, dividends excluded. Power names shown against the chip trade and the index.
Source: Massive.com / FMP closes. Price return only; excludes dividends and financing costs.
| Asset | Ticker | Change |
|---|---|---|
| Nvidia | NVDA | +49.9% |
| S&P 500 | SPY | +23.1% |
| Vistra | VST | -12.3% |
| Constellation | CEG | -21.0% |
We want to be exact about what this does and does not say. The theme was correct — electricity really was the bottleneck, and it still is. The trade stopped working long before the thesis did. Those are different things, and confusing them is the single most expensive mistake in thematic investing.
What we'd do differently
Buying the bottleneck instead of the crowded thing feels defensive. It isn't. The bottleneck name has a thinner shareholder base, a simpler story, and less conviction underneath it — so it gaps further when the story is questioned. Measure crowding by who has to sell, not by how obvious the name is.
→ 28.3% versus 17.0%, in one session.
The recovery is what killed this trade. Once you have been proven right through a 46% drawdown, no level feels like a reason to sell. An exit condition defined before the position — a price, a multiple, a change in the story — is the only version of that decision made by a calm person.
→ The new high in September 2025 was the exit. It didn't look like one.
Power demand from data centres is a decade-long story. That does not entitle any particular stock to go up this year. When a thesis is measured in years and a position is measured in months, you need a written rule for what the position does while the thesis is still true but the price isn't cooperating.
→ Right theme, wrong holding period, negative return.
WHAT WE GOT RIGHT
Power as the AI bottleneck was a real, durable, and correctly identified constraint — and it produced a 472% move before it stopped.
WHAT WE GOT WRONG
We treated the derivative trade as the conservative expression. The one-day test said the opposite, in public, with numbers.
WHAT IT COST
A position that was up triple digits ended eighteen months later below where the test began. No exit rule, no realised gain.
Straight talk
This is a case study, not a track record. It is a teardown of a well-documented market episode using public closing prices, published so you can check every number yourself.
Nothing here is a recommendation. We are not telling you to buy or sell Vistra, Constellation, Nvidia, or anything else. Prices are historical and past performance says nothing about what happens next.
We publish the losses too. A desk that only shows you its winners is showing you marketing. This one cost real money and the write-up is the point.
Free · Six mornings a week
Themes before tickers. The trade structure a desk would actually run. The risk stated before the reward, and the write-up when we're wrong. Sunday through Friday, five minutes a morning.
Free forever. No spam. Unsubscribe in one click.