# Yields at a 20-Year High, and Taiwan's Market Walks June's Loop Again — Notes on Caibaogou Ep. 568 Source: Realpha Blog (blog.getrealpha.com) Original article and charts: https://blog.getrealpha.com/en/blog/caibaogou-2026-10-01-568-vs-x/ > Episode 568 of the Taiwanese podcast Caibaogou (Oct 1, 2026) on the US 10-year Treasury yield pushing above 5.2%, TSMC evaluating a second Texas campus, the labor squeeze in fab-facilities engineering, and NVIDIA's AI factory certification. An educational listening note that lays out the reasoning and the counter-case; no investment advice, no stock recommendations. Published: 2026-10-02 Locale: en Tags: Caibaogou, interest rates, TSMC, data centers, fab facilities TL;DR: The show argues high rates are survivable because AI returns still clear the cost of money; I think the condition worth tracking is whether contracted capacity pricing softens — the yield number itself settles nothing. ![A night construction site on the flat Texas plain, cold blue work lights raking across bare steel framing and bundled chilled-water pipes, warm amber floodlights and distant city glow receding to the horizon](/covers/caibaogou-2026-10-01-568-vs-x-cover.png) > It returns and repeats its course; on the seventh day, the return comes round. > > —— Book of Changes, Hexagram 24, "Return" (pre-Qin; translated by the author) In episode 568 of Caibaogou, published October 1, 2026, host Wei-Yu and analyst Sky spend most of the opening on the US 10-year Treasury yield, which had pushed to 5.2%–5.3% that week. Sky notes that this is above 2008 and above the second quarter of 2007 — a 20-year high. His reading: money got more expensive, but the return on AI data center investment still clears that number, which is why equities did not break this time. He also states the condition under which his view fails: as long as new capacity gets leased and pricing doesn't crack, the story continues. The back half covers TSMC reportedly evaluating a second Texas campus (industry chatter puts it at six advanced fabs and US$265 billion) and the labor logistics behind it, and closes on what he finds most interesting — Taiwan's market has walked the same themes it walked before July, with the macro facts unchanged. ## First, the strongest version of the show's case Sky's definition of interest rates is plain: a rate is the cost of money — what you pay to use it. His base layer comes from Buffett, who called rates the gravity of money: when the number gets big, money floats down from the sky and liquidity cools. Then he asks a question that made me stop: when has a good economy ever come with low rates? That question is the hinge of the whole bull case. High rates tend to arrive alongside heat, so 5.25% on its own decides nothing. The question is what the money is being borrowed to do, and what that thing returns. He contrasts two eras. In 2007–2008, money poured into real estate, whose yield is close to fixed — so as rates climbed, the return got shaved away slice by slice. He uses his own earlier interest in EV charging infrastructure as the live example: it looks unattractive now for the same reason, rates rose too far. Apply the identical mechanism to AI data centers and the conclusion flips: cloud providers keep borrowing, and if money costs 6% while the deployed capital returns 15%–20%, you borrow. What underwrites that, he says, is the scale of the productivity gain, and it won't stop near term because the spread is still wide. He cites Google's CEO saying their TPU investment pays back within a year — not an audited figure, but it is the spender's own account of the payback period. ![A short bar on the left shows the borrowing cost of six percentage points, a tall bar on the right shows the investment return of fifteen to twenty percentage points, a double-headed arrow between them marks the gap, and the top of the right bar is drawn as a dashed band to show that it is an estimate.](/figures/spread-cost-vs-return-en.svg) He adds a detail I liked. Rent and compute pricing react at different speeds. Buy a building to let out at a 15% margin, and every point of rate increase eats a point, while rent is sticky and can't be raised on command, so the wound stays open. A data center looks like a rental business, yet its pricing swings: racks keep getting added, every addition prices at the current rate, and even older contracts get repriced upward. The same rate shock hits one business as a slow bleed and the other as a same-quarter catch-up. The second block is TSMC's reported second Texas campus. The show is careful — still under evaluation, industry chatter — yet the chatter is specific: six advanced fabs, US$265 billion, on par with the current total committed in Arizona, and Sky thinks Arizona will grow further since the figure being discussed there is at least nine fabs. He argues the dollar number matters less than the physical count: a standard advanced fab runs roughly 30,000 wafers a month, so once the tool count is set, output value is roughly set. Six more fabs means every model you built before was short by six. Who eats first? Facilities engineering, then equipment. The flesh of this segment is people. The hosts had been to several earnings calls and closed-door briefings; everyone knows the US is short of workers, and Taiwan is short too. Among the first cohort sent over, plenty spoke little English and went anyway, carried by the few who could talk to local government. They looked up visa costs live on air, and the number jumped from US$8,000 per head to a US$100,000 figure on screen that they could not pin down — but either way, sending one person costs hundreds of thousands of Taiwan dollars, so only veterans get sent. From there Sky draws a competitive picture: this trade becomes a skilled-labor dispatch game, and whoever has bodies on site picks up the rush work others can't finish. Customers watch progress daily, and when a block slips they bring in a second contractor; work due in two days does not price cheaply. So alongside record backlogs, he guesses margins come in better than before. ![Two groups of squares: nine on the left for the fabs planned in Arizona, six in dashed outlines on the right for the fabs under review in Texas, each square marked at 30,000 wafers a month, with the whole right-hand group flagged as what the original model never counted.](/figures/texas-six-fabs-volume-en.svg) Why not fear being squeezed by one giant customer? Because everyone is expanding. Micron printed an 87% gross margin — of course they build. Intel's advanced node is picking up customers. On the packaging side, capacity is tight even without TSMC adding, and the OSAT buildout carries an even larger share for some equipment suppliers. While demand holds, the people who build fabs are not down to a single client. One stretch made me laugh. NVIDIA launched an AI factory certification program; the first names through are international players who deliver whole-facility infrastructure, and two Taiwanese rack-cooling firms are absent, which the press wrote up as direct competitors losing out. Sky's read is that the scopes differ: certification covers whole-plant delivery, including digital twins and chillers sitting far from the hall, none of which belongs to a rack-cooling vendor. What it really signals is that NVIDIA now supplies a recommended vendor list for how to build an AI data center at all. The buyers are telecom equipment rooms and small operators who want a site next to a hydro plant, several orders of magnitude below the big four clouds. He mentions people converting land originally slated for battery storage into data centers, because storage projects are drawing protests in the US and Japan, while compute containers can simply be trucked in — 250 kilowatts each, three or four and you're open. ## Now stand on the other side What bothers me most is the word spread. Borrow at 6%, return 20%, and the gap looks wide — but the evidence on each side isn't equally hard. The 6% is a signed contract you pay every period; the 20% is an estimate of the future, supplied by the party doing the spending. "TPU pays back in a year" has no externally checkable denominator: depreciation life, internal transfer pricing for compute, whether idle capacity counts — all of it sits inside their own books. Subtract a market price from a self-reported return, and most of the error in the difference comes from the first term. ![A balance scale with solid blocks on the left pan for the signed contract cost labeled externally verifiable, and dashed blocks on the right pan for the self-reported return labeled denominator not disclosed, so the scale tips to the left.](/figures/evidence-asymmetry-scale-en.svg) Second, fast repricing cuts both ways. The show treats "old contracts get repriced" as supportive, but what reprices up reprices down. Rent stickiness is a drag in an upswing and a shield when demand weakens; a high-turnover pricing structure means data center cash flows fall faster than rent when the cycle turns. The show's own failure condition lives here: as long as it gets leased and pricing doesn't crack. That hangs the entire bull case on a variable that has not yet been stress-tested. ![The same two lines shifted to a weakening-demand scenario: the rental line, protected by long contracts, drifts down only slowly, while the compute rate line falls steeply, exactly reversing which one looked better when prices were rising.](/figures/fast-repricing-cuts-both-ways-en.svg) Third, the argument separating 2007 from today has the same shape as the bull argument of 2007. Back then there was also a productivity-and-structural-demand story — demographics, urbanization, credit innovation distributing real estate risk more efficiently. We know in hindsight where it broke, but at the time it produced the identical conclusion: returns exceed the cost of capital, so borrow. For "the productivity gain this time is enormous" to function as evidence, it needs a falsifiable form; otherwise it is "this time is different" wearing a new coat. Fourth, what I think the episode skipped: nobody split the cause of the yield move. If 5.25% is driven by inflation and nominal growth, nominal revenues and nominal returns rise alongside it, and the pressure on equities is limited. If it is driven by term premium, fiscal deficits, or a change in who's buying, the discount rate rose while cash flows did not, and the pressure is real. The two look identical on a chart and call for opposite responses. "When has a good economy come with low rates" files the move under the first category, and in the week the yield hit a 20-year high, that attribution needs support. Fifth, the chain from six Texas fabs to better facilities margins is long. The show flags the evaluation stage and the industry-chatter sourcing, and the US$265 billion is a press number; Sky's own advice to watch the fab count rather than the dollars is itself a statement that the dollars are unreliable. The longer chain is margin: short of people → rush orders → subcontracting → better pricing. The same premise supports the opposite outcome — being short of people also means missed delivery, progress penalties, and deferred recognition, while a dispatch cost of hundreds of thousands of Taiwan dollars per head lands directly in cost of goods. The episode scores the labor shortage as pricing power without also scoring it as cost and delivery risk. Their own live check of job postings lands on this side too: one firm's openings are still all in Taiwan, while another lists Singapore, Japan, and the US. Whether you can move people abroad and how big your backlog is are two different facts. ![The same yield number splits into two scenarios: on the left, inflation lifts both the cash flow bar and the discount rate bar together so the gap stays the same; on the right, only the discount rate bar rises while the cash flow bar stays put, widening the gap.](/figures/yield-cause-two-branches-en.svg) The sharpest cut comes from the closing segment, where it is offered as a fun observation. Sky says Taiwan's market is re-walking the pre-July path — passive components, power components, price hikes — and the headlines are about 40% similar once you swap the dates. The third quarter was slaughtered, the fourth quarter is back up, and the objective conditions haven't changed; rates are higher than they were. That observation undercuts the bull core: if the same fundamentals could not hold the price up in Q3, then whatever is holding it up in Q4 is also not those fundamentals. Price completes a loop in a quarter while industry change runs on a far longer cycle, and the residual needs liquidity to explain it — spread can't. ## Where I land, and on which condition I'm with the show, but only the version that states its failure condition. The condition is legible: contracted pricing for new capacity holds, and the capacity gets leased. Those are trackable. The 5.25% number is not, no matter how closely you track it. So I move my attention off the headline to three things that are harder to look up and more useful: whether per-megawatt pricing on new contracts is softening, whether depreciation lives are being extended (a longer life flatters book returns while cash stays the same), and whether the yield move traces to inflation or to term premium. Three questions I imagine people actually carrying, worked from the pain back to the method. **"The news says TSMC will build six fabs — should I go find the beneficiaries?"** There's a timing gap hidden in the question. The report is at the evaluation stage, and the market has long known who sits in the front row, so by the time you read it, the part you know is already in the price. What isn't priced is execution: can that firm actually move people, can it move up from subcontractor to main contractor, does it pick up the rush work others can't finish. You can see a little of that in backlog disclosures, posting locations, and hiring — not in a headline. My own order is to read how that company talked about US labor last quarter, then decide whether it's worth my time. **"Q3 scared me badly, Q4 is ripping, am I about to chase again?"** The loop observation converts straight into a check. Put June's articles next to this week's; if swapping the dates leaves them reading the same, what you're facing is money coming back, and the company has not improved. Those two situations justify different amounts of drawdown tolerance. A liquidity move will loop out again; an improved business won't. I've paid for this one: same theme, bought high and sold low last loop, and caught myself wanting to buy high again this loop. What I now force myself to do is dig up the reasons I wrote down last time and read them. Reading that paragraph alone cuts a lot of trades. **"Yields are at a 20-year high — should I cut exposure?"** I'd split this into two questions, asked of each position separately rather than of the index: does its return clear today's cost of money, and is its cash flow fixed or able to rise with prices. Rent-like, fixed-income-like, long-contract-priced things get shaved; things that can reprice quarterly while still adding supply take a smaller wound. You'll find that one 5.25% cuts to different depths across your own book, and that answer appears in no article about yields — only on your own position sheet. ![A horizontal rate line sends four wounds of different depths downward: long-contract pricing and fixed income are cut deep, while quarterly repricing and still-adding-supply are only shallow nicks.](/figures/same-rate-different-wounds-en.svg) ## Worth a look - Caibaogou podcast episode 568, "Macro in Focus: High Rates vs. Taiwan's Market x Texas Beyond Fried Chicken," October 1, 2026 - The raw US 10-year yield series: FRED's DGS10 — plot it yourself against the Q2 2007 level - TSMC's own disclosures and quarterly calls, the place to judge whether the Texas report deserves weight - NVIDIA's AI factory certification page and its list of certified vendors - The kind of check the hosts demonstrated live: look up which regions those facilities-engineering firms are actually hiring into ## One thing to take away One idea: the narrative loops roughly once a quarter, and your memory gets rewritten to match whatever the price is doing now. There is exactly one antidote — keep the words you wrote at the time, then read them back line by line. The reason anyone could say "June's headlines are 40% similar with the dates swapped" is that somebody kept June. One thing I've tried: pick something you were worried about three months ago that has nothing to do with money — a person at work, a situation at home, a small thing with your body. Dig up the messages you sent that week, or any line you wrote yourself, and check them against today item by item: which ones changed, and which were just your mood going round a loop and coming back. If you kept nothing, write one today — three sentences is plenty, with the date on it. Mine sits pinned at the top of my phone's notes app, and I open it three months later.