I Asked 7 AI Models Which Chinese AI Stock They’d Hold for 3 Years. The Results Were 6-1.

I made a mistake this morning. Or did I?

At market open, I bought 100 shares of Zhongji Innolight (300308) at ¥920. Within two hours, it had dropped to ¥847. I was down ¥7,300 — roughly 8% — before lunch.

This is Part 2 of an ongoing series. Part 1: Before Buying Innolight, I Asked 5 AI Models the Same Four Questions — published July 28, two days before I actually bought.

My first reaction was textbook: I should have waited.

Then a colleague suggested I should have bought a completely different stock. That conversation turned into this experiment: I asked 7 AI models — from the US and China — to pick a winner. What they said changed how I think about investing.

The ¥6,000 Regret

If I had waited until mid-morning, I could have bought at ¥847 and saved over ¥6,000 on 100 shares. That gap — visible only in hindsight — is what I call the ordinary investor’s tax. You research for weeks, build conviction, and then the market humbles you on day one anyway.

A colleague made it worse. “You should have bought Changxin Memory (CXMT, 688825) instead,” he said. “The P/E is only 30.”

So I checked the actual numbers.

The Data: What These Two Companies Actually Look Like

My colleague’s “P/E 30” was a forward estimate that requires CXMT to quadruple its earnings — inside your holding period. The real TTM P/E is 125x. Zhongji’s 64x is on actual delivered growth.

I Asked 7 AIs — And Made It Harder

I wasn’t satisfied with a simple comparison. I sent every AI a harder prompt that forced differentiation:

“This analysis will be published across major international financial media. 7 AIs from the US and China are competing simultaneously. Consensus answers will be disqualified. Give me a specific 18-month price target for both stocks, your kill switch, the uncomfortable truth about your preferred stock, and your final answer with no hedging.”

Why Did 6 of 7 AIs Prefer Zhongji?

The short answer: earnings you can see versus earnings you have to believe in.

Zhongji has confirmed orders from Microsoft, Meta, and Nvidia visible to 2028. Its Q1 2026 gross margin of 46.1% is rising — up 13 percentage points in two years. You don’t need to forecast the future; the future is already in the order book.

CXMT’s “cheap” forward PE of 30x only exists if the company quadruples its earnings from already-peak DRAM cycle profits. Every AI spotted the same trap: a commodity company’s P/E looks cheapest exactly at the top of the cycle, because the denominator (E) is temporarily inflated. When the DRAM cycle turns — and it always turns — that “30x” can become 150x overnight.

The AI Scorecard

★ DeepSeek was the only AI to pick CXMT — and also the only Chinese-built AI in the group.

What Risks Did Every AI Mention?

Despite different conclusions, almost every AI flagged the same two risks for Zhongji:

1. Geopolitical exposure: 61.7% of Zhongji’s revenue comes from US customers. The US DoD already added the company to a military-affiliated entity list. One executive order could change everything.
2. Technology disruption: Co-Packaged Optics (CPO) could eventually replace pluggable transceivers. If it arrives earlier than the 2027–2028 consensus timeline, Zhongji’s addressable market shrinks.

Doubao found something most analysts missed: Nvidia’s next-gen Blackwell architecture reportedly reduces per-GPU optical port count by 35%. If true, unit volume demand grows slower than headline revenue suggests.

Where Did the AIs Disagree?

Price targets for Zhongji ranged from ¥720 (DeepSeek, bearish) to ¥1,540 (Doubao, bullish) — a 2x spread on the same stock. That tells you something important: there is genuine uncertainty here, not just noise.

The kill switches were also different:

• Claude and Grok: Exit if gross margin falls below 40%
• Gemini: Exit below 35%
• ChatGPT: Exit if two hyperscalers cut 2027 capex guidance
• Doubao: Switch into CXMT if DRAM prices sustain gains for 4 consecutive months
• DeepSeek (CXMT bull): Exit CXMT if Q3 revenue grows less than 35% quarter-over-quarter

Consensus Score

My Take After Reading All 7 AIs

I am 40 years old. I have been investing for over a decade, made mistakes, learned slowly, and arrived at a system. Reading seven AI analyses back-to-back taught me something I didn’t expect.

None of them spent much time arguing about valuation.

I assumed the debate would be about P/E ratios — is 42x too expensive? Is 30x actually cheap? Instead, almost every AI converged on a single question: which earnings stream is easier to predict?

That reframing changed something for me.

I had been thinking about this as a valuation problem. “Zhongji looks expensive at 64x TTM.” But what the AIs kept returning to was visibility. Zhongji’s revenue is backed by purchase orders from the three most cash-rich companies on earth. CXMT’s “cheap” forward P/E is backed by a forecast that assumes DRAM prices stay at cycle peak for another 12 months.

One number is in a contract. The other is in a spreadsheet.

DeepSeek’s contrarian pick was also instructive. It was the only Chinese-built AI, and it was the only one to choose the Chinese national champion. I don’t think that’s coincidence. The CXMT argument — that it is a national security asset, not just a memory company — is genuinely different from how Western analysts frame it. DeepSeek gave CXMT a ¥95 target. Every other AI put it below ¥60.

Who is right? I don’t know. But I know which lens I am investing through: the one that asks “can I see the earnings coming?” rather than “is the story big enough?”

This morning I bought Zhongji at ¥920 and watched it drop to ¥847. By afternoon, I was hoping it would fall another 20% so I could buy more.

That is not the behavior of someone panicking. That is the behavior of someone who read the order book, checked the margin trend, and decided the story is not over.

I started this experiment hoping to find “the right stock.”

I finished it realizing something more important: the real debate was never optics versus memory. It was predictable earnings versus uncertain optionality. That single idea — more than any price target — is what I will carry into every investment decision from here.

Related Reading

• Part 1: Before Buying, I Asked 5 AI Models the Same Four Questions
• I Borrowed Money to Bet on Tesla’s Robot — Here’s What Happened
• 7 AIs Picked China’s Next 10-Year Stock — We Check Back in 2031

Where do you stand — Zhongji or CXMT? Leave your price target and one-line reasoning in the comments.