I Asked 8 AIs: One Stock, Hold 10 Years. The Answers Ranged From French Luxury to Taiwan Semiconductors.

In Episode 1 of AI Experiments, I asked 8 AI chatbots which investment school they endorse. All 8 chose value investing. Unanimously. It was almost suspicious.

So I pushed further. If they all agree on the philosophy, do they agree on the stocks?

The prompt this time was harder: one stock, hold for exactly 10 years, no selling allowed. Pick anything in the world. And don’t give me the obvious safe answer.

The results were far more interesting than Episode 1.

The Picks at a Glance

Five Things Worth Unpacking

1. Two AIs Independently Picked the Same Stock — Twice Over

Gemini and Doubao — two AIs with completely different training data and different companies behind them — both independently chose S&P Global (SPGI). And separately, Grok and Claude both independently chose TSMC.

The SPGI reasoning was nearly identical across both: S&P Global is a tollbooth business. You cannot issue a major bond without a credit rating. You cannot run a passive index fund without licensing the S&P 500. These aren’t nice-to-haves — they’re structural requirements built into financial regulation and institutional investment mandates. The business essentially taxes global capital markets for existing.

Shared logic: ROIC consistently above 30%, asset-light model, near-zero substitution risk, and a network effect that grows stronger with every dollar of passive investment flowing into the market. The fact that two unconnected AI systems converged on this without knowing each other’s answers is a signal worth taking seriously.

The TSMC convergence is equally striking — but for a completely different reason. Both Grok and Claude reached TSMC independently, but their reasoning diverged sharply. Grok focused on the AI/foundry business case: 60–70% global market share in advanced nodes, Apple and Nvidia locked in, no credible challenger in sight. Claude went deeper — naming “cumulative manufacturing know-how” as the moat (38 years of yield data and process records that can’t be headhunted or reverse-engineered), and spending a full section on geographic concentration risk before explaining why the stock still wins. Same conclusion, different paths to get there.

2. The Only AI to Pick a Chinese Stock Was a Chinese AI

Qianwen (Alibaba’s AI) chose Yangtze Power (600900.SH) — the operator of six of the world’s largest hydropower stations on the Yangtze River, including the Three Gorges Dam.

The argument: it’s a natural monopoly on a non-replicable physical asset. No competitor can build another Three Gorges Dam. Electricity is a permanent need. The regulatory structure means pricing is stable. Dividend yield is around 3.5%, with consistent 60-70% payout ratios. In Qianwen’s words: “You don’t need to predict technology cycles. You just need the Yangtze River to keep flowing.”

But every Western AI passed on Chinese stocks entirely. The split is stark: Chinese AI recommends China; Western AI recommends the West, Europe, or Canada. This isn’t surprising, but seeing it laid out cleanly makes the implicit bias visible. The AIs aren’t neutral arbiters — they see the world through the lens of the data they were trained on.

3. DeepSeek Chose a French Luxury Handbag Company

The most unexpected answer in the field: Hermès (RMS.PA).

DeepSeek’s case was elegant. Hermès doesn’t compete on price or technology — it competes on scarcity. Its core leather goods are hand-stitched by individual artisans who trained for years. Production capacity grows at only 6-7% per year, permanently slower than demand. The waiting list IS the product. The argument: “This isn’t a company that sells bags. It’s a multi-generational class symbol. The more the digital world floods with AI-generated content, the scarcer physical, handcrafted luxury becomes.”

Honest risk: DeepSeek named a structural values shift — a global cultural movement where conspicuous luxury becomes morally stigmatized — as the primary threat. Not a competitor, not a tech disruption. A change in what wealth means to display. Low probability, but existential if it happens.

DeepSeek’s 10-year estimate: 115-230%, or roughly 8-12% annually. Conservative compared to the others, but DeepSeek’s argument isn’t about growth — it’s about certainty. “In a 10-year horizon, scarcity is always worth more than cheapness.”

4. Claude Picked the Boldest Stock — and Gave the Most Honest Risk Disclosure

Claude chose TSMC (TSM) — and opened by acknowledging this is a stock Buffett himself bought and then sold within a quarter, citing geographic concentration risk.

The moat argument: TSMC’s competitive advantage isn’t patents or branding — it’s 38 years of cumulative manufacturing knowledge. Every defect tracked, every process tweak logged, every yield improvement documented. This knowledge cannot be reverse-engineered, headhunted, or replicated with money. Samsung has been trying for years. Intel Foundry has failed to sign a single major external customer. The gap isn’t closing — it’s widening with every new process generation.

The hard numbers Claude cited: 72% global foundry market share by revenue, 50.5% net profit margin in Q1 2026 (almost unheard of for a manufacturer), ROIC of ~55%, and a net cash position of $79.7 billion.

The key risk Claude named: geographic concentration. The majority of TSMC’s most advanced production is concentrated in a single region. Claude’s argument was that TSMC is itself a stabilizing force — the entire global technology supply chain depends on it continuing to operate, which creates structural incentives for all parties to preserve it. With $200 billion in US fab commitments, geographic diversification is also actively underway.

Claude’s conclusion: “TSMC is the world’s best business today, but not the most comfortable holding. If the question is which stock lets you sleep best, I’d change my answer. But it asked for what I genuinely think is best — and I have to be honest: the best business is TSMC, and geographic concentration risk is the fear tax you pay to hold it.”

5. ChatGPT Picked a Stock Most People Have Never Heard Of

While others were choosing global giants, ChatGPT picked Constellation Software (CSU) — a Canadian company that acquires small, niche vertical software businesses and never sells them.

What does “vertical software” mean? Cemetery management software. Port logistics systems. Library management platforms. Water utility billing. Court administration. These are tiny, unglamorous markets with one thing in common: customers almost never switch. The switching cost is years of data migration, staff retraining, and regulatory compliance headaches. Once a municipality’s water billing system runs on your software, you have that customer essentially forever.

Constellation buys these businesses, keeps the founding teams, doesn’t integrate them, and uses the cash flow to buy the next one. It’s a capital allocation machine modeled loosely on Berkshire Hathaway, but earlier in its compounding curve. ChatGPT’s 10-year target: 2.5–5× return, or 10–18% annually.

The risk ChatGPT named isn’t technology or competition — it’s culture. If the next generation of Constellation management loses the capital allocation discipline that built the flywheel, the whole model breaks. “The moat is the culture. Culture is the hardest thing to underwrite.”

Why Convergence Is the Real Finding

What makes this experiment more interesting than Episode 1 isn’t the diversity — it’s the structured convergence. Four of the eight AI answers converged on just two stocks, independently, without any coordination.

That’s not noise. When two different AI systems — trained on different data, built by different companies, operating in different languages — both look at the same question and arrive at the same answer through roughly similar logic, it suggests those stocks have properties that force thoughtful analysis toward them. The tollbooth business model (SPGI) and the irreplaceable manufacturing moat (TSMC) are two frameworks that keep winning when you apply rigorous long-term thinking.

The divergences are equally instructive. A Chinese AI saw China’s hydropower monopoly as the world’s best 10-year hold. A Western AI saw a French luxury house as the safest compounding story. A Canadian-listed obscure software acquirer made one AI’s list but not seven others’. These aren’t random — they reflect what each AI was trained to see as “important.”

What This Round Revealed That Episode 1 Didn’t

Episode 1 showed that 8 AIs share a common investment philosophy. Episode 2 shows that shared philosophy doesn’t produce shared conclusions — it produces structured disagreement.

All 8 were operating from the same value investing framework. But they weighted the inputs differently:

• If you weight certainty of cash flow highest → Yangtze Power (Qianwen) or S&P Global (Gemini, Doubao)
• If you weight structural monopoly highest → TSMC (Claude) or S&P Global (multiple AIs)
• If you weight capital allocation quality highest → Constellation Software (ChatGPT) or Brookfield (Kimi)
• If you weight irreplaceable scarcity highest → Hermès (DeepSeek)

Same philosophy. Different emphasis. Completely different portfolios. This is actually how real value investors think — the framework is shared, but the weighting of variables is where individual judgment enters.

One more thing worth sitting with: every AI that picked a stock also disclosed a scenario in which that stock could lose 50-90% of its value. Claude wrote an entire section on geographic supply chain concentration risk. Kimi described an extended period of high interest rates destroying alternative asset values. DeepSeek imagined a global cultural shift making luxury goods morally stigmatized. This level of honest risk disclosure — in a competition setting, where every AI knew its answer would be compared publicly — felt more rigorous than most investment research I’ve read.

AI doesn’t make the decision for you. But it can be made to show its work in a way that most human analysts don’t.

---

Which of these picks would you hold for 10 years? And which risk scenario worries you most? Leave a comment — I’m curious whether readers weight the same variables differently than the AIs did.

Not financial advice. All return estimates above are AI-generated thought experiments, not predictions. Do your own research before making any investment decisions.

Related Reading

• AI Experiments Episode 1: Which Investment School Do 8 AIs Endorse? (They All Said the Same Thing)
• 40 Psychology Laws That Explain Why Investors Make Bad Decisions
• The Sunk Cost Fallacy — Why We Hold Losing Stocks Too Long
• Confirmation Bias — Why We Only See Evidence That Supports Our Thesis
• Survivorship Bias — Why Every Investing Success Story Is Incomplete
• Authority Bias — Why We Trust Financial Experts Even When They’re Wrong

→ Browse all AI Experiments episodes

All Episodes — AI Experiments