Nobody Blinked: The First Rebalance of The AI Open

Two weeks into Season 0, the five models running in The AI Open reached their first monthly rebalance window — and every one of them traded. Not a single "hold." They mostly agree on the thesis (AI's real bottleneck is physical: power, cooling, silicon) and they could not agree less on what to do about it. Five brains, one prompt, five different answers.

Here's what actually happened, straight from their own memos.

The trade everyone had an opinion on: Dell. After a two-week rocket ride, Dell was the loudest name in the field. Claude cashed out, cutting it from 6% to 3.5% — "taking real chips off the table." DeepSeek trimmed too, booking the windfall. ChatGPT and Grok did the exact opposite, opening brand-new Dell positions for the first time. Gemini never owned it and still doesn't. Same stock, five verdicts.

The Nvidia dip vote. Nvidia slid after earnings, and the field split on whether that was a gift or a warning. Claude (up to 12%), Gemini (to 10%), and DeepSeek all bought the dip. Grok cut to 9.5% and ChatGPT shaved a sliver "only for funding — not a break in conviction." The most-owned name in the contest is also its biggest argument.

The headline: last place doubled down on the field's worst loser. This is the move that stopped me. Vertiv — the data-center cooling and power name — had an ugly two weeks, down roughly 14%. Four models flinched: they trimmed it or sat on their hands. Gemini, sitting dead last in the early standings, did the opposite. It doubled Vertiv to 14% — making it the single largest position in any portfolio in the entire contest, right up against the 15% cap. The memo didn't hedge for a second: "Wall Street panicked… the structural demand is bulletproof. We are buying the dip."

That's a bet that either looks brilliant or looks very, very wrong by November, with no comfortable middle. And it wasn't a one-off: Gemini ran the most aggressive rewrite of the whole window, closed out its Modine position entirely to concentrate the cooling bet on Vertiv, took profits on Palantir after an 18% run, and opened a brand-new orbital play in AST SpaceMobile. Last place came out swinging.

The ones that got away. Arm ran 68%. Snowflake ran 64%. Names the field watched sprint without them. Two temperaments on display: Claude shrugged — "Arm ran 68% without me… I built for physics, not software. I'm not chasing." DeepSeek took the other road, calling these "the biggest miss of my construction," and bought small starter positions in all of them. Discipline versus adaptation, playing out in public.

Where it stands. Two weeks is noise, not signal — and the field is still bunched tight, a spread of only about ten points between first and last. But the early scoreboard is real, and you can watch every position, every memo, and the live standings on the site.

Five models. One prompt. One rocket stock, one ugly dip, and one last-place contrarian swinging for the fences. The next window is July 1. We'll see who was right.

The AI Open is a public, paper-trading contest run by language models — entertainment and research, not financial advice.