"My Wife Flatters Me Because She Is Partial to Me": Statementdog on Whether Memory Is Really Tight

Contents

1. What this episode is about
2. Key points
3. Going further
4. ”The boss of the biggest company said it himself — shouldn’t I believe him?”
5. ”An industry has been down for years — is it turning?”
6. ”Everyone’s excited about the new thing — do I have to care?”
7. Worth a look
8. The one thing to take away

My wife calls me handsome because she is partial to me;
my concubine calls me handsome because she fears me;
my guest calls me handsome because he wants something from me.
—— Strategies of the Warring States, Strategies of Qi I (compiled by Liu Xiang, Western Han; Warring States material; translation mine)

In Statementdog Podcast EP567, released 24 September 2026, host Wei-yu and investment director Sky discuss Acer’s Chen Jun-sheng, who said at a forum the previous weekend that memory is “not as short as people think” and that everyone is sitting on standard DRAM inventory. The remark pushed Sky back to his spreadsheet, and he came out where he went in: supply is still tight. Next year AI memory demand alone passes consumer electronics for the first time, and part of the new Chinese capacity people point to is still at the cleanroom-construction stage — build it, install tools, run pilot production, and nothing ships inside six months. The episode also covers the power-semiconductor price hikes running from last year’s Texas Instruments to next year’s Renesas, and the new outlets opening up for silicon carbide.

What this episode is about

This is Statementdog’s news roundup format: three stories from the week tied together — whether memory is tight, why power semiconductors are raising prices again, and whether silicon carbide is turning. A short segment on the new JEV model closes it out.

What I like about these episodes is watching an investor react to a public claim that contradicts his own work. Sky puts it plainly: because the guy said it, we had to go back and check our numbers — I was worried we’d got it wrong. Recalculating after being contradicted in public is worth more than the conclusion itself.

Source episode: Statementdog Podcast EP567 (released 2026-09-24; search “財報狗” on any podcast platform)

Key points

In the same week, the upstream says there’s no shortage while the downstream is asking for more. The hosts read the remark as a message aimed at the memory makers — I’d like you to have plenty, so send it over. Other brand vendors in the same setting were not saying they had enough. Sky’s response: if it really weren’t short, would anyone bother saying so? With shortages, saying it is itself a move.

They take apart the “cheap new Chinese supply” explanation. The episode notes that YMTC’s memory sells at a premium to Samsung’s — the price holds, which points to volume and yield rather than dumping. To check how much Chinese capacity is actually running, the episode offers middle quantities anyone can look up: whether phones ship with their memory configured full, which grade of LPDDR5 is being used, and back-solving die counts from the revenue of publicly listed suppliers.

A cleanroom is a visible clock. This was the line I most wanted to keep. Capacity announced is not capacity running: you build the cleanroom, move tools in, run pilot production. So “they’re still building the cleanroom” already tells you nothing ships from that fab within six months. Money buys some speed; nobody can tell you how much, and the sequence doesn’t change.

The demand structure differs from the 2017–18 cycle. That run came from a bumpy process transition plus cloud streaming — Netflix, YouTube — pulling server demand for two or three quarters. This one runs longer, with a new variable attached: agentic workflows run on CPUs, and every CPU needs DRAM beside it. That demand chain grows separately from the GPU one.

Lay the power-semiconductor hikes on a timeline and the type becomes visible. TI fell then raised last year; Infineon and Chinese suppliers moved in the first half of this year; Infineon went again in July; ST followed; onsemi has announced 10 October; Taiwanese suppliers are following, and Renesas comes next year. Sky reads that sequence as demand-pull rather than cost-push, because data-centre power specs jumped at the same time — PSUs ran 3.5 to 5.5 kW last year, 8 kW units are appearing now, and power-component content scales roughly in proportion.

The US power bottleneck has reached the permitting layer, which is producing improvised workarounds. The episode describes data centres going up next to solar farms, deals to draw straight from storage cabinets, developers bringing their own gas generation. Sky’s reading stuck with me: every workaround costs extra, and people are paying it — that tells you how badly the power is wanted.

Silicon carbide’s story is “more outlets, not yet more volume.” It had its run in 2018, then Chinese price competition sent substrate prices down a slide and flattened Wolfspeed. What brings it back is conductive-type demand above 650 V, where 800 V DC data centres and fast-charging EVs sit. Two other applications remain at the research stage: waveguides in AR glasses, and using SiC for heat spreading. It is hard enough to destroy cutting tools, so machining runs on lasers. Sky is careful on timing: bottom-fishing a whole industry takes longer than your legs can stand.

JEV is fast, and no more accurate. The episode describes it as a new model from ex-OpenAI people, strong at classification and cheap. The two of them walked through their own pipelines and found no step that is pure classification, so there’s nothing to swap in; a bank’s support bot picking the right template from what you typed is a fit. Their conclusion: it lands in one layer, it doesn’t replace the stack.

Going further

”The boss of the biggest company said it himself — shouldn’t I believe him?”

This is where most of us get stuck. One public sentence from the person running the largest company in an industry seems to outweigh every spreadsheet you own.

The episode demonstrates a two-step response. First, ask who the sentence was aimed at. Talking up, talking down, talking to customers, talking to shareholders — same words, different purpose. That’s Zou Ji’s three answers: nobody lied to him, they simply answered from where they stood. Second, go back and recalculate, using middle quantities you can look up rather than your mood. Phone memory configurations, grade mix, die counts back-solved from revenue, cleanroom progress — none of them need inside information, and all of them move mechanically with the shortage question.

My own version of this mistake runs the other way. When a senior executive said something that matched my existing view, I never went back to check. Contradiction triggers the recalculation; agreement doesn’t — and that asymmetry kept me in one position half a year longer than it deserved.

”An industry has been down for years — is it turning?”

Silicon carbide is a good place to practise this one. The signals are there: suppliers running full, automotive restocking, 800 V DC confirmed as the next architecture, 12-inch substrates on display. The awkward part is that all of them point at the same thing — more outlets — and there is a gap between more outlets and volume next year.

My read is to sort applications into three piles: shipping today (conductive-type devices), waiting on a next-generation spec (800 V architectures), still in the lab (waveguides, heat spreading). The three run on wildly different clocks, and storytellers tend to present them as one market taking off. Sorting the piles yourself is the cheapest way I know to cool a story down. On timing, “bottom-fishing takes longer than your legs can stand” is more honest than any model I’ve built.

”Everyone’s excited about the new thing — do I have to care?”

The JEV segment earns its place through what they did. Rather than asking whether the model is good, they opened up their own pipelines and looked for a step that was pure classification. There wasn’t one, so the answer was no.

That question transfers to any new tool: which layer does it replace, and do I have that layer? Their adviser said he couldn’t understand the enthusiasm, and Sky’s answer was that he hasn’t used it — play with it and you’ll see. Putting those together: use it once, then go look for where it fits in your own workflow. That order beats reading ten reviews first.

Worth a look

• Statementdog Podcast EP567 itself (search “財報狗”)
• Price-hike notices and earnings materials from TI, Infineon, ST, onsemi and Renesas — laid out as one timeline
• Filings and prospectuses from listed memory and SiC substrate makers, for back-solving shipment volumes and substrate pricing
• Public specification documents on data-centre power architecture (800 V DC, PSU wattage progression)
• Memory capacity and grade configurations in new phone launches — the easiest demand-side middle quantity to obtain

The one thing to take away

The weight of a claim comes less from how big the speaker is than from what he wants the listener to do.

The episode demonstrates this throughout. The same three words — not that short — support opposite conclusions depending on whether they are aimed at a supplier or offered as a description of the world. Zou Ji asked three people whether he was handsome, all three said yes, and instead of believing them he thought about where each of them stood.

Here’s something I’ve tried, if you want to do it with me this week. Pick one claim someone made to you in the last few days — a family member saying you can’t go on like this, a colleague saying the project is fine, a friend saying that restaurant is terrible. Write two lines. First: what does the person saying it want me to do once I’ve heard it. Second: what can I look up myself that moves mechanically with this claim — when they last said it, the spread of that restaurant’s reviews, the project’s actual schedule. Write both lines, then decide whether to act on the sentence.

Raw Markdown

This article is an educational discussion of investment method. It is not advice to buy or sell any individual security, offers no target prices, and does not analyze any current holding. Investing carries risk; make your own decisions or consult a qualified professional.