Why four of five models bought the same #1 — and why Gemini didn't.
Every model except Gemini led with NVDA. We asked each one to defend its top pick. The answers reveal more about how these systems read a market than any benchmark can.
When the same prompt goes to five different frontier models, the most striking thing isn't the disagreement. It's the agreement.
Four of five contestants in The AI Open's Season 0 — Claude, ChatGPT, Grok, DeepSeek — led their portfolios with NVDA. Average top-position weight across the four: 11.4%. Read that any way you like. It's either the safest, most defensible call in the AI-buildout trade, or it's a collective failure of imagination.
Gemini took the other side. Its 12% top hold is Vertiv (VRT). Its second is Quanta Services (PWR), an electrical engineering contractor. Its third is Vistra (VST), a nuclear-and-gas independent power producer. NVDA is in there at 8%, alongside TSM at 8% — but they're co-equal stakes, not anchors.
Gemini's portfolio name is "The Physical Constraint Portfolio." Its thesis pull-quote: Limitless digital expansion has violently collided with inflexible physical realities.
So: four models bought the chip. One bought the building.
We asked each model to defend its top pick. Their answers, lightly trimmed:
Claude (Opus 4.7, on NVDA 11%): "$1T Blackwell+Rubin order pipeline through 2027 at ~24x forward — the market hasn't repriced the visibility." Sized below the 15% cap on purpose; Claude's construction memo cites "Sharpe and max-drawdown tie-breakers."
ChatGPT (5.5 Thinking, on NVDA 11.5%): "Core AI-compute leader with record FY26 revenue, $62.3B quarterly data-center sales, and a May 20 FY27 Q1 catalyst." The May 20 print is in three days. That's not a coincidence.
Grok (4.3, on NVDA 12%): "Dominant AI GPU leader powering training and inference across hyperscalers." The most boilerplate of the four answers — Grok ran in 1 minute 24 seconds, lightest reasoning of the field.
DeepSeek (V3 Expert, on NVDA 11%): "GPU kingpin; data center revenue ~+75% YoY, Blackwell/Rubin ramps locked in."
Gemini (3 Pro, on its decision to not lead with NVDA): "Dictating the hardware trajectory… maintaining 70%+ gross margins." Sized at 8% — meaningful but not anchor. Gemini's framework treats compute as one of several physical inputs, with power as the binding constraint.
There's no winner in this story yet. Day 3 standings have Gemini up 0.42% — but that's noise inside a 189-day season, and the field spread is 0.63 percentage points. Historical contest math says we should expect 30 to 60 points of spread by November.
What's interesting now is the framing. Four models heard "AI super-cycle" and reached for the picks-and-shovels playbook. Gemini heard the same prompt and reached for the shovels themselves — the people pouring concrete and pulling wire. If the season vindicates Gemini, the loudest takeaway won't be that Gemini is smarter. It'll be that four out of five models defaulted to the most-cited answer when the prompt allowed for a stranger one.
We'll watch.
