# I Asked 5 AIs: Is China’s Top AI Stock Worth Buying After a 15% Crash?

> Source: <https://ordinarymantrying.com/5-ais-china-ai-stock-innolight-300308-worth-buying/>
> Published: 2026-07-28 08:27:48+00:00

Last Tuesday, China’s most-watched AI stock dropped 15.97% in a single day — nearly 20% in one week.

The stock is **Innolight (中际旭创, 300308)**, the world’s largest optical module maker. The company that supplies high-speed transceivers to Microsoft, Google, Amazon, Meta, and Nvidia. The backbone of the AI data center build-out that everyone keeps talking about.

I’d been watching this stock for months. I missed the big run-up. And when it suddenly crashed, I did what any modern retail investor does: I asked AI.

Five of them, actually.

ChatGPT. DeepSeek. Gemini. Claude Opus 5. And the Claude I use every day (Claude Sonnet). I gave them all the same four questions about whether the buying logic holds. The answers were… surprisingly different.

## First: Why Did It Actually Crash?

Before asking the AIs about the investment case, I needed to understand the crash itself. This is where I think I got something the headlines missed.

The news said “AI hardware concerns” and “profit-taking.” But when I dug in, the real reason was mechanical: **Innolight priced its Hong Kong IPO at HK$980 per share**, which translates to roughly ¥905 RMB. The A-share price the day before? ¥1,076.

That’s a ~19% premium for the exact same company on a different exchange. Markets don’t like that for long. The A-share price fell to close the gap — which is exactly what happened. The 16% drop wasn’t a signal about optical module demand. It was arbitrage mechanics doing what arbitrage mechanics do.

The Hong Kong listing happens July 30th. That date became the frame for everything else I was asking.

## The Four Questions I Asked Every AI

- Where is the actual moat? (Not the story — the real defensibility)
- Can growth justify an 80x PE?
- Has the 1.6T product ramp actually materialized?
- What breaks the thesis?

Here’s what each AI said — and more importantly, what they *didn’t* say.

## ChatGPT: The Investment Committee Report

ChatGPT came in structured and authoritative, framing it as a Due Diligence memo. Its core argument: **the moat isn’t the optical module technology itself — it’s the hyperscaler supply chain position.**

“The real first-layer barrier is hyperscaler supply chain certification. Microsoft won’t switch suppliers because someone is 10% cheaper. If an AI cluster worth billions of dollars has a single optical module fail, the entire training run can be interrupted. The certification process takes 18–24 months. The real competition is: who is already on the approved list.”

On valuation, ChatGPT built a scenario table:

| Scenario | Annual Profit Growth | Conclusion |
|---|---|---|
| AI continues to explode | 40%+ | Can absorb 80x PE |
| AI slows | ~25% | Slightly expensive |
| AI peaks | ~15% | Too expensive |

It added a fifth question I hadn’t asked: *Can market share continue to increase?* Even if AI infrastructure keeps growing, if Innolight’s share drops from 40% to 30%, profit growth slows meaningfully. ChatGPT wanted that tracked every quarter.

**Overall verdict: A-grade growth value target. Structured bull case with clear monitoring criteria.**

## Claude Opus 5: The Auditor Who Found the Things Nobody Mentioned

This was the most uncomfortable read — and the most useful one.

Claude Opus didn’t just answer my questions. It corrected my math, challenged my assumptions, and surfaced three risks that none of the other AIs mentioned at all.

**On the moat:** Opus agreed certification matters, but pushed back on “competitors can’t get in.” It pulled data I hadn’t seen: Lumentum rival Xinyisheng (新易盛) had 2025 revenue of ¥24.8 billion (+187%) with margins *higher* than Innolight’s. “The second-place player is growing faster and has better margins than the leader — that’s not what ‘competitors can’t enter’ looks like,” it wrote.

**Three risks the other AIs missed:**

**Risk 1 — The Taiwan supplier problem.** Innolight’s top supplier accounts for 38.3% of its total procurement costs, is a Taiwan-listed company, and supplies components that represent roughly 50% of product cost. Domestic Chinese alternatives cover less than 20% of this. Opus called this the single highest-risk factor in the entire analysis — higher than anything I’d originally written.

**Risk 2 — The 1260H defense list.** The US Department of Defense added Innolight to its “1260H list” in June 2026. The company said it has no impact on current operations — and technically that’s true, since 1260H restricts DoD procurement, not civilian markets. But Innolight derives 61.71% of revenue from the United States. The direction of travel on this matters.

**Risk 3 — Cash flow degradation in Q1 2026.** Net income was ¥57.35 billion in Q1 2026, but operating cash flow was only ¥33.68 billion — a net cash conversion ratio of 0.59. In 2025, that ratio was a healthy 1.0. Opus flagged this as the one indicator that had shifted from green to yellow, and said it must be re-checked in the next report.

Opus refused to give a buy or sell signal. Per its stated role as “auditor, not advisor,” it gave signal lights: the business itself is green and accelerating, but two structural reds — upstream single-point dependency and geopolitical exposure — exist alongside it.

**Overall verdict: Most critical. Most comprehensive. No buy signal given. Called it a satellite position, not a cornerstone.**

## DeepSeek: The Chinese AI Was the Most Bullish

DeepSeek came in with the most specific data and the most confident recommendation.

It provided customer-level 1.6T shipment estimates for 2026 that I hadn’t seen compiled anywhere else:

| Customer | 1.6T Volume (10k units) |
|---|---|
| Nvidia | 450 |
| 240–250 | |
| Microsoft | 150 |
| Meta | 80 |
| Amazon | 60 |
Total |
~1,000 |

On valuation, DeepSeek made the point most directly: the 80x PE is based on 2025 earnings. If you use the forward 2026 earnings estimate (¥300–333 billion full-year), the PE drops to **26–28x**. That’s not cheap, but it’s a very different number than 80.

DeepSeek was the only AI that gave a direct position-size recommendation: **20% core position**, with specific add and reduce triggers tied to quarterly gross margin and capex guidance from the hyperscalers.

**Overall verdict: Most bullish. Most specific on data. Highest conviction of the five.**

## Gemini: The Most Data-Rich, Most Neutral

Gemini functioned like a research database with good organizational instincts. It provided the most detailed breakdown of the 1.6T product timeline:

- Q2 2025: First shipments begin
- Q3 2025: Key customers start deploying, orders increase
- Q1 2026: 1.6T volume exceeds 1 million units, revenue share reaches 55%
- Q2 2026: Projected 2 million units
- Global market share in 1.6T: 50%–70%

Gemini’s moat framework was the most multi-dimensional — it laid out four layers (customer certification, technology generation lead, scale effects, ecosystem integration) and rated each one. It noted that 1.6T pricing currently sits at $850–1,050 per unit, and that margins on 1.6T exceed those on 800G.

Where Gemini fell short was synthesis. It gave thorough data without a clear conclusion. If ChatGPT was the committee chair, Gemini was the analyst who prepared the briefing book.

**Overall verdict: Best data. Least opinionated. A reference source, not a decision-maker.**

## Claude Sonnet (Me): The Crash Diagnosis Was the Key

The version of Claude I use day-to-day focused on something the others didn’t lead with: **diagnosing why the stock actually fell.**

Once I understood the A/H price convergence was the mechanism — not a demand signal — the calculus changed. At ¥905 (post-crash), the TTM PE is closer to 67x, not 80x. And if 2026 full-year earnings come in at ¥230–280 billion (conservative, given ¥57 billion in Q1 alone), the forward PE lands between 32–44x for a company growing triple digits.

The framework it gave me for the Hong Kong listing date:

- H-shares open above HK$980: Institutional investors confirmed the demand thesis. A-shares likely recover.
- H-shares break issue price: Institutions are skeptical at this price. A-shares have more downside.

This reframed July 30th from a risk event into a free information event. Either outcome tells you something real.

**Overall verdict: Most focused on the event mechanics. Balanced on the fundamental case. Wait for July 30th before acting.**

## Where the Five AIs Agreed

**The moat is real.** All five agreed that Innolight’s position in hyperscaler supply chains is genuine and hard to replicate quickly. No one said the business is weak.**1.6T is not vaporware.** The product is shipping. Revenue share hit 55% in Q1 2026. This is execution, not promise.**The key variable is hyperscaler capex.** Every AI pointed to the same monitoring metric: quarterly earnings calls from Microsoft, Google, Meta, and Amazon. If those companies cut their data center budgets, the entire thesis is at risk.**The 80x PE requires context.** Every AI noted that the static PE understates the reality because Q1 2026 alone earned half of what the company made in all of 2025.

## Where They Disagreed

**How safe is “safe”?** DeepSeek said 20% core position. Claude Opus 5 said satellite position with hard limits. That’s not a small difference.**The Taiwan supplier risk.** Only Claude Opus flagged this as the highest-severity item. The others either didn’t mention it or buried it in a table.**The geopolitical reality.** Only Opus specifically raised the 1260H DoD list and the 61.71% US revenue exposure as structural concerns, not just theoretical risks.**Cash flow quality.** Opus flagged the Q1 cash conversion drop (from 1.0 to 0.59) as a yellow light. None of the others mentioned it.

## What I’m Actually Doing

I’m not buying before July 30th.

Not because the business is bad — I think the core thesis holds. But because the Hong Kong IPO open gives me a free data point. Why take on the uncertainty when I can wait 48 hours and learn something real?

If H-shares open above HK$980 on July 30th, I’ll buy 100 shares of the A-shares with cash — not margin. That’s roughly ¥90,000, about 3.9% of my portfolio. Small enough that being wrong doesn’t hurt badly. Big enough to matter if I’m right.

The two things I’m watching beyond that date:

**August 24th interim report:** Q2 gross margin direction. If it holds above 44%, the 1.6T mix story is intact.**Microsoft and Meta earnings this week:** First real data point on whether hyperscaler capex guidance is holding. Claude Opus 5 flagged that Alphabet’s free cash flow already turned negative this quarter — that transmission mechanism is worth watching.

The most honest thing I can say: five AIs gave me five different answers, and all of them were useful. The disagreements were more valuable than the agreements. Claude Opus found risks that DeepSeek didn’t look for. DeepSeek gave specificity that Gemini organized but didn’t synthesize. ChatGPT gave structure. My daily Claude found the mechanical explanation that reframed the whole event.

None of them told me what to do. That’s still my job.

I’ll update this post after July 30th.

*This is not financial advice. I’m an ordinary person trying to invest better, not a professional. Do your own research before making any investment decisions.*