# MacroMicro EP208: Separate the Long Lead Items From the Short Ones, and You Can Read the AI Cycle
> The main contribution of MacroMicro's After Meeting EP208: it splits the order-cut risk for the second half into two layers. Cuts to ordinary components only move the manufacturing cycle around; cuts to genuinely scarce parts are what a real turn looks like. Plus TSMC's inventory rising for two straight quarters and why that number has two meanings, Korea's leverage-on-leverage unwind, and why Kimi K3 did not trigger another DeepSeek moment. Educational notes, not investment advice.
Published: 2026-07-29
Locale: en
Tags: macromicro, podcast-notes, macro, inventory-cycle, ai-capex, education
TL;DR: The most useful tool in this episode is the split between long lead items and short ones: cuts to the former only produce cycle noise, cuts to the latter mark a real turn. Pair that with TSMC's inventory rising two quarters running, a number that reads as good news in an active restock and bad news in a passive one. Also why cheaper models didn't scare hardware this time: Kimi K3 is cheap but enormous, its weights need about 1.4TB, so it demands rack-scale hardware rather than less of it.
> *From the side, a ridge; head-on, a peak.*
> *Near or far, high or low, never the same.*
> —— Su Shi, "Written on the Wall of Xilin Temple" (Song dynasty); translation mine
> Personal notes after listening to MacroMicro's After Meeting **EP208** (released 2026-07-26). Not a transcript, not official show content. Please support the original if you want the full version.
## What this episode is about
The valuation reset that started in early July ran until roughly the 20th and looked to be finding a floor, with the baton about to pass to earnings season so fundamentals could speak. Then a flare-up between the US and Iran pushed oil back up. So the episode covers a handover moment: **attention moving from expectation to verification.**
The part worth writing down isn't the market recap. It's the framework the research team offered. They break the second half's risks into three layers, and deliberately separate the ones that make the cycle wobble from the one that makes it turn. That distinction is more useful than any bullish or bearish conclusion.
## The main points
- **The macro still holds.** Oil is back above 80 dollars and briefly approached 90 on the geopolitical flare-up, with US crude inventories already below the lowest July level since 2018, right as summer driving season arrives. But June consumption came in unusually strong (retail sales up more than 6 percent year on year) and June inflation undershot, with core goods and services ex-shelter both contributing negatively. Not slowing. Contracting. That buys some cushion.
- **The consumption number needs an adjustment, and the show made it.** Part of that strength came from the World Cup. Comparing May and June credit-card spending, the increase concentrates almost entirely in host cities. Strip those out and the rest is roughly flat. So: not weak, but not as strong as the headline. That habit of actively removing known noise before drawing a conclusion is the most transferable thing about this show.
- **Korea is a live case study in leverage.** Borrow through a personal credit line, use the proceeds to buy a leveraged ETF, then add margin on top of that. Leverage on leverage on leverage. On the way down, all three contract at once. And the measures intended to protect retail investors (raising margin requirements, restricting leveraged products, a July rate hike) added to the selling pressure instead. The Korean market is down close to 20 percent from its high. That figure is as of the 24 July recording, after the market had steadied; measured peak to trough instead, the index fell roughly 30 percent from its mid-June high, and July was its worst month on record. Which convention you use changes how painful this looks.
- **TSMC beat across the board and the whole supply chain still fell.** The concern is margin pressure in the early ramp of a new node, and guidance did flag roughly three to four points of dilution as N2 scales. The show reads this as a short-term worry, since N3 demand is strong enough to offset some of it, and the company's stated "customer success" pricing stance means it was never going to push price the way memory suppliers have.
- **The signal actually being missed is TSMC's inventory rising two quarters in a row.** Very little market discussion of it. In an active restocking phase, rising inventory reads as orders being pulled in. But the next phase is passive restocking, which marks the top rolling over. Same number, two meanings.
- **Three risks for the second half, two short-dated and one long.** First, power and land constraints at data centres delaying chip delivery and piling up inventory through the chain (the show notes finished-goods inventory rising at some suppliers, not just work in progress). The place to watch is server assemblers: if their inventory builds sharply, cuts to ordinary long lead items could travel back up the chain. Second, weak consumer and automotive demand, with only high-end models selling, Chinese subsidies rolling off and autos weaker still, so brands facing higher input costs they can't pass on will cut orders. Third, and the biggest: if AI applications advance more slowly than performance expands, capex gets revised down and even the genuinely scarce short lead items get cut. The show puts the window for that third one at 2028 to 2030, on the logic that the fabs being built now come online then, and one chip leader's roadmap targets edge computing around that point. If edge AI hasn't materialised by then, nothing replaces the current demand.
- **Kimi K3 didn't trigger panic, and the reason is worth keeping.** It benchmarks close to the frontier at roughly a third of the price, and the gap between Chinese and US models has compressed from six months or more to a matter of weeks. But cheap is not the same as small. At 2.8 trillion parameters, the weights alone need about 1.4TB of memory. It won't fit in a single box, so it requires rack-scale hardware, and splitting weights across many chips leans hard on interconnect bandwidth. Cheaper model, more demanding hardware.
- **The profit pool across the AI stack is being redistributed.** The show's frame has five layers: energy, chips, models, infrastructure, applications. Chinese open-source models compress pricing power and the valuation premium at the model layer, but the profit released by cheaper inference flows to applications (more tokens affordable) and to infrastructure (the model is free, the thing running it is not). One layer losing often means another gaining.
- **Sovereign AI has become a budget line.** The show notes that in June the US Commerce Department briefly halted the export of one frontier model, which pushed several governments to think harder about what happens if a single country can switch off a capability their defence, healthcare, and financial systems depend on. The buyers are clear (anywhere sensitive data can't leave), and so is the spending: models, compute, and compliant operations.
## Extensions
**One: long versus short lead items is the most useful tool this episode hands out.**
Most arguments about AI go nowhere because the two sides aren't discussing the same thing. One says inventory is about to blow up, the other says demand is still strong, and both can be right once you separate the parts: ordinary components have started to accumulate while genuinely scarce ones haven't loosened at all.
The value of the split is that it gives you an **order of observation** rather than a conclusion. Watch server assembler inventory first, then whether end brands can still pass on price increases, and only then whether capex guidance actually comes down. The first two loosening means cycle noise. The third means the trend has changed. A framework that ranks its own signals beats ten slogans about direction.
**Two: an inventory number has no direction on its own. It needs a partner variable.**
Inventory rising two quarters running is good news during active restocking and bad news during passive restocking. Looking at inventory alone will never resolve which one you're in, because that depends on what demand is doing. This is the same disease as the point I made in the Gooaye notes about markets pricing the rate of change rather than the level: **a number only means something once you know what it's paired with.**
In practice that argues against reacting to any single indicator turning. Inventory, gross margin, capex all have at least two readings, and what decides the reading is the variable sitting next to them. So the thing worth building isn't a list of indicators. It's a table of pairs.
**Three: whether cheaper models help or hurt hardware depends on whether the models get bigger or smaller.**
This is the sharpest reasoning in the episode. The implicit logic behind the panic in early 2025 was: models get cheaper, so less compute is needed. But this round of cheapness came from architectural efficiency, and the model itself got larger. Weights of about 1.4TB won't fit in one box, so it goes to a rack, and splitting the model across chips consumes interconnect on top of that.
So the test isn't price. It's **whether the weights fit on a single node**. What I like about that test is that it's checkable: parameter counts and weight sizes are published, so you don't have to judge whose narrative sounds better. Next time someone argues that a cheap model threatens hardware demand, asking that one question will filter out a good half of the noise.
## One honest boundary
The episode places the window for its biggest risk in 2028 to 2030. Worth saying plainly: **a forecast on that timeline is very nearly unfalsifiable in the near term.** It can't be wrong yet, because the date hasn't arrived, and by the time it does the roadmaps will have been revised several times over.
That doesn't make it worthless, but the value has to be filed in the right place. It works as a **watch post**, not as a reason to hold a position. The difference: a watch post says "if A and B both show up, I re-examine my thesis," while a position reason says "because of what happens in 2029, I'm buying now." The first can absorb new information. The second will torment you for the three years in between. The most common misuse of a long-horizon macro call is letting it justify a short-horizon position.
## Further reading
- Original episode: *MacroMicro After Meeting* EP208 (2026-07-26), on YouTube and major podcast platforms
- The other two sets of notes from the same week: Gooaye EP682: How to Read a Sell-Off With No Bad News and Statementdog EP541 Notes: Once Chips Got Big, Round Wafers Stopped Paying
- On inventory cycles: search active restocking, passive restocking, and the four-phase inventory cycle
---
*These are personal, educational notes written after listening to a podcast. They are not official MacroMicro content, not investment advice, contain no price targets, and make no recommendation on any current position. Companies and figures appear only as context from the episode and have not been independently verified. Investing carries risk; do your own research or consult a qualified professional.*