EM equity fund inflows
$1.91B
Week through Nov 4, 2026; fifth consecutive weekly inflow
EM bond fund flows
-$299M
Same week — carry money left while equity money stayed
CoreWeave 5-year CDS
~1,038 bps
Record territory, last quoted Oct 15, 2026
Taiwan, week to Nov 6
-12.2%
Biggest weekly fall since 1991
SK hynix Q3 operating margin
80%
Quarter ended Sept 30, 2026; revenue up 328% year over year
AI-linked debt issued, 2026
$466B
Estimated through late October, versus $101B a year earlier
The rotation, Nov 4
Money Left AI. It Did Not Buy Cheaper AI.
Investors are describing a rotation into cheaper AI in emerging markets. The fund-flow data through Nov 4 describes something else: money leaving the AI complex rather than buying it. Global equity funds absorbed $38.63 billion, a seventh straight weekly inflow and the largest since Sept 23, while technology-sector funds took just $623.61 million, a three-week low. Capital rotated out of AI into non-AI equities and bonds.
Where the money went, week through Nov 4, 2026
Global equity funds posted their largest weekly inflow since Sept 23, but technology-sector fund inflows fell to a three-week low. Emerging-market equity funds absorbed money for a fifth straight week while emerging-market bond funds saw an outflow — the carry leg was being sold, not bought.
Unit: US$ billions
Global equity funds
Seventh straight weekly inflow
38.6
US equity funds
14.8
European equity funds
13.3
Global bond funds
Largest weekly inflow since Oct 7
9.9
Asian equity funds
8.3
Emerging-market equity funds
Fifth straight weekly inflow
1.9
Technology-sector funds
Three-week low
0.6
Emerging-market bond funds
Outflow
-0.3
- Emerging-market equity funds absorbed $1.91 billion, a fifth consecutive weekly inflow, even as emerging-market bond funds sold $299 million.
- US equity funds drew $14.81 billion and European funds $13.25 billion, while global bond funds took $9.91 billion, their largest weekly inflow since Oct 7.
- Cash funds took $140 billion in the week to Nov 4, the seventh-largest weekly inflow on record, and the dollar rose almost 0.8% in its biggest one-day jump since June.
- Three-month euro cross-currency basis swaps touched about -8 basis points on Nov 5, the lowest since late 2023, before recovering to roughly -3.5 basis points by the following Monday.
Nov 5-6
The Selling Hit the Companies That Fund AI
| Market or name | Move | Detail |
|---|---|---|
| Nasdaq Composite (Nov 5) | -5.5% to 24,315 | S&P 500 -4.4% to 7,353.73; the seven largest US technology names shed about $900 billion in the final hour, one chip designer accounting for $374 billion |
| Taiwan TAIEX (week to Nov 6) | -12.2% | Worst week since 1991; closed Nov 6 down 4.5% after being more than 8% lower intraday |
| TSMC | Limit-down circuit breaker | Hit the daily limit Nov 6 in Taipei |
| KOSPI (Nov 6) | -6% close, -8.7% intraday | Korea Exchange triggered a circuit breaker during morning trading |
| CoreWeave | -10% | Five-year CDS past 1,000 basis points to a record |
| Indian IT services (Nov 6) | Up to +3% | Infosys up more than 2%, Wipro up 0.68%, against a 12% fall in the Philadelphia Semiconductor Index |
| US dollar (Nov 5) | +0.8% | Biggest one-day jump since June; euro cross-currency basis at its lowest since late 2023 |
The damage landed where AI capital is funded, not where AI is consumed. Foreigners sold a record $31.78 billion of Taiwanese shares in October, surpassing the March high, and have withdrawn $131 billion from Korean equities this year — the largest outflow among major Asian markets — even as the TAIEX was still 56.1% higher year to date at 45,211.83 on Oct 30. The selling hit the companies that fund AI, not the ones that sell AI services.
There was a specific catalyst, and it was political rather than technical. After the Democratic sweep in the US midterms, investors worried about delays to data-centre construction, and that fear jumped straight into the capital budgets of the companies that build those data centres. The first thing repriced was the timing of AI capital expenditure, and the listed order books moved accordingly.
The funding-cost layer
The AI Build-Out Is Financed With Debt, and That Debt Repriced First
This is what makes the emerging-market 'discount' a funding cost rather than a valuation opportunity. AI-linked debt issuance ran to an estimated $466 billion in 2026 through late October, against $101 billion in the same period of 2025. CoreWeave's five-year credit default swap last printed near 1,038 basis points on Oct 15, its 9.75% bond due 2031 yielded 15.78%, and its debt changed hands at 77 cents on the dollar with implied default odds near 63%. The funding cost of the AI build rose faster than the price of its suppliers' equity.
The company itself confirmed the mechanism in a September filing: CoreWeave's $3.0 billion of convertible senior notes due 2033, with proceeds partly earmarked for repayment of indebtedness, capital expenditure and moving the enterprise credit profile toward investment grade. An issuer buying back its own credit risk is telling you the spread is the constraint. When the marginal funder has to buy back its own credit profile, equity multiples compress everywhere upstream.
I believe a 50 basis point hike in the policy rate should be considered at the December meeting.
The policy rate is the other half of the discount. Waller said he is leaning toward supporting a 25 basis point increase in December if October inflation improves as expected, that a 50 basis point increase should be considered if core CPI runs around 0.3% or higher, and that hiking only 25 basis points would put the Fed's credibility at risk. Core PCE inflation was 3.1% year on year in September, headline PCE 3.7%, unemployment slipped to 4.1% in October, and futures now price a December hike and another by March. Every basis point lifts the hurdle on the capital that builds the AI supply chain.
| Company | Capex / revenue | EV / EBITDA | Earnings yield | Return on equity |
|---|---|---|---|---|
| CoreWeave | 271% | 18.3x | -7.1% | -45% |
| SMIC | 75% | 20.0x | 0.8% | 5% |
| TSMC | 35% | 15.1x | 4.4% | 41% |
| Micron | 23% | 7.3x | 10.3% | 92% |
| SK hynix | 13% | 3.9x | 21.3% | 117% |
| Samsung Electronics | 9% | 3.5x | 16.1% | 46% |
| Advantest | 2% | 25.9x | 2.8% | 63% |
| Wipro | 2% | 10.1x | 6.7% | 16% |
| Infosys | 2% | 9.9x | 6.6% | 34% |
Who funds the AI build-out
Trailing capital expenditure as a share of revenue. The two heaviest users of capital are also the least profitable businesses in the chain, which is the opposite of what a funding-cost discount implies.
Unit: % of revenue, trailing twelve months
Put the multiples next to the capital intensity and the 'discount' dissolves. SK hynix trades on 3.9 times trailing EBITDA with an 80% operating margin because revenue is up 328% year on year; Micron trades on 7.3 times after an 86.8% gross margin in its September quarter. These are peak-cycle earnings, and trailing multiples computed on peak earnings are the cheapest kind of cheap. A multiple built on peak earnings is the most expensive kind of cheap.
The destination set
Displaced Capital Bought AI Monetizers, Not AI Hardware
On Nov 6, with the Nasdaq down 6% — its biggest fall since March 2020 — and the Philadelphia Semiconductor Index down more than 12%, Indian IT services rose as much as 3%. Infosys gained more than 2% and Wipro closed up 0.68% at 199.75 rupees. The distinction is capital intensity: Infosys spends 1.6% of revenue on capital expenditure and Wipro 1.8%, against 35% at TSMC and 271% at CoreWeave. The money went to the businesses that convert AI into revenue without paying for it.
- Infosys reported September-quarter revenue of $5.15 billion, which rose 1.3% at a 21.3% operating margin, with AI revenue now 10.5% of the total and up 29% sequentially.
- Large-deal bookings reached a $5.2 billion total contract value with 62% classified as net new, giving the AI services demand a backlog rather than a capital plan.
- Guidance was nevertheless cut to 1-2% revenue growth with margins held at 20-22%, which is why the December-quarter print, not the AI narrative, is the real test.
- Cash funds still hold $140 billion of dry powder from the week to Nov 4, which is what lets investors sit through a December rate hike without being forced to sell.
This is where the consensus story inverts. The emerging-market bid was real and persistent, but it landed in the parts of emerging markets that consume AI rather than manufacture it, and it arrived as equity money with bond money going the other way — because emerging-market bonds are precisely the funding a December hike reprices. This was a funding-cost trade, not a value hunt.
The supply chain
One Transmission Chain, and It Runs From the Policy Rate to the Order Book
| Link | Who | Evidence of linkage |
|---|---|---|
| Upstream — logic and memory | TSMC, SK hynix, Micron, Samsung Electronics | TSMC October revenue up 59.7% year on year and $46.72 billion in Q3 at a 66.4% gross margin; SK hynix says HBM4 becomes mainstream in the fourth quarter and will pass HBM3E on volume and revenue; Micron spent $7.1 billion of net capital expenditure in fiscal Q3 2026; Samsung's HBM4 sales more than tripled in Q3 |
| Upstream — test and inspection | Advantest | First-half net sales of 845.2 billion yen, up 60.5%, with operating income up 96.5% to 456.8 billion yen and raised full-year guidance |
| Upstream — China capacity | SMIC | Highest capital intensity among AI-linked names at 75% of revenue on a 0.8% trailing earnings yield |
| Midstream — the financiers | CoreWeave | Capital expenditure at 271% of revenue, more than $51 billion of debt and obligations, a 0.46 current ratio, five-year CDS past 1,000 basis points |
| Downstream — hyperscaler budgets | The seven largest US technology names | About $900 billion of market value erased in the final hour of Nov 5, with construction-delay fears after the US midterms |
| Downstream — AI monetisation | Infosys, Wipro | AI revenue at 10.5% of Infosys' total and growing 29% sequentially; $5.2 billion of large-deal contract value, 62% net new |
| Flows — the foreign exit | Korean and Taiwanese equities | A record $31.78 billion net sold in Taiwan in October and $131 billion withdrawn from Korea this year, against a TAIEX still 56.1% higher year to date at Oct 30 |
The transmission runs in one direction and it is fast: a higher policy rate raises the hurdle on data-centre financing, which raises the cost of the compute that TSMC, SK hynix and Micron sell into and of the test equipment Advantest sells, which is what the Nov 5-6 reaction was pricing. Construction delays would appear in foundry and memory orders before they appeared in any demand forecast. The first signal of a delayed build-out is the order book, not the revenue line.
Horizons
Four Dated Catalysts Decide Whether the Discount Persists
Over the next quarter the calendar does the work: the October inflation print before the December meeting, TSMC's November monthly sales on Dec 10 and fourth-quarter results on Jan 8, 2027, CoreWeave's next debt sale, and Infosys's December-quarter results. Four dated events decide whether this is a funding-cost discount or a real one.
- The October CPI release, then the December FOMC meeting: futures pricing a December hike and another by March makes the 25-to-50 basis point range the base case.
- TSMC's November sales on Dec 10 and fourth-quarter results on Jan 8, 2027: the cleanest read on whether data-centre construction is being deferred or cancelled.
- CoreWeave's next bond sale: pricing inside the terms of its September convertible would signal that the AI credit cycle has stabilised.
- Infosys's December-quarter results: a second guidance cut with margins intact would confirm that AI services monetise without capital expenditure.
- SK hynix's HBM4 ramp, where the company expects HBM4 to surpass HBM3E on both volume and revenue in the fourth quarter.
Over one to three years the arithmetic reverses, and the company disclosure is unusually specific. SK hynix has signed long-term agreements with 14 major customers covering more than half of 2027 and 2028 production capacity, expects the memory shortage to persist through 2030, and plans roughly 100 trillion won of FY2026 shareholder returns including a 20,000 won quarterly dividend with a 10,000 won special portion from Kioxia investment gains — peak-cycle capital return by definition. If construction merely shifts a quarter or two, those numbers hold and the discount closes. The bull case survives a delay and fails on a cancellation.
Synthesis
What Would Break This Thesis
- If CoreWeave's CDS compresses back toward its pre-October level and its new debt prices near par, the credit-cycle leg of the argument disappears and the emerging-market hardware discount becomes real.
- If the Fed holds in December and cuts in the first half of 2027, dollar funding costs fall, emerging-market bond funds turn back into inflows and the carry unwind reverses.
- If Indian IT services keep compounding at high single digits with margins at 20-22% and capital expenditure near 2% of revenue, the destination set stays services rather than hardware.
- If Korean and Taiwanese foreign outflows stop and October's record $31.78 billion Taiwan pace does not repeat, the crowdedness argument weakens sharply.
None of that is knowable today, which makes this a choice about what to own rather than a forecast of where flows go next. The durable exposure sits with the toll booths — foundry, memory, test equipment — and with the services firms that turn AI into revenue without financing it. Own the parts of the chain that get paid whether or not the build-out accelerates.
The investable map: where the funding cost actually bites
- October revenue rose 59.7% year on year to 586.9 billion New Taiwan dollars even as the AI complex was being liquidated, with January-October volume up 43.3%.
- A 66.4% gross margin in the September quarter and roughly 70% foundry share let TSMC fund 35% capital intensity from cash flow rather than credit.
- Near term, the Dec 10 monthly sales and Jan 8, 2027 results separate deferred construction from cancelled construction.
- Over one to three years the toll booth is unassailable; the risk is a multiple that assumes flawless execution.
- Third-quarter revenue of 104.6 trillion won rose 328% year over year at an 80% operating margin, and net cash reached 98.7 trillion won.
- Trailing multiples of 3.9 times EBITDA are cheap only because the earnings base is at a cyclical peak.
- Long-term agreements with 14 customers cover more than half of 2027-28 capacity and the shortage is expected to persist through 2030.
- Shares are still about 59% below their 52-week high and fell 10.4% on Nov 6, so the cycle risk is priced and the cash-flow risk is not.
- Third-quarter operating profit of 107.4 trillion won jumped 783% year over year, with HBM4 sales more than tripling.
- A 16.1% trailing earnings yield prices in a memory upcycle that runs to 2030 rather than a durable re-rating.
- The shares fell 4.1% on Oct 29 despite the record profit, which shows the market discounting the cycle rather than the quarter.
- Capital intensity of 9% of revenue is the lowest among the memory names, which limits funding risk if the build-out slows.
- Fiscal Q4 revenue of 54.23 billion dollars at an 86.8% gross margin and full-year EPS up 811% mark the most cyclical profit pool in the AI chain.
- At 86.8% gross margin, the multiple has to fall before the earnings do, and a December hike raises the discount rate applied to peak margins.
- Trailing capital intensity of 23% of revenue and 7.1 billion dollars of quarterly net capital expenditure make it a funding-sensitive supplier.
- The shares are roughly 42% below their 52-week high of 1,255 dollars, so the de-rating has started but is not finished.
- A 0.46 current ratio and 271% capital intensity leave no room for a higher cost of capital, with negative working capital of 11.4 billion dollars.
- Five-year CDS near 1,038 basis points, 15.78% yields on the 2031 bonds and debt changing hands at 77 cents on the dollar price a refinancing problem, not a growth story.
- The September convertible issue, with proceeds partly aimed at moving the credit profile toward investment grade, shows the funder is the constraint.
- Horizon: the next bond sale is the catalyst; equity has already fallen about 51% from the June quarter's market value.
- September-quarter revenue of 5.15 billion dollars rose 1.3% at a 21.3% operating margin, with AI revenue at 10.5% of the total and up 29% sequentially.
- At 1.6% of revenue in capital expenditure and 9.9 times trailing EBITDA, Infosys monetizes the AI build without financing it.
- A 6.6% trailing earnings yield and a 7.3% free cash flow yield are paid whether or not data-centre construction accelerates.
- Horizon: the risk is a second guidance cut with margins intact, which would say AI is deflating pricing faster than it is scaling revenue.
- First-half net sales of 845.2 billion yen rose 60.5% with operating income up 96.5%, and full-year guidance was raised.
- Capital expenditure of 2.5% of revenue means Advantest sells the capex boom without carrying it.
- But 25.9 times trailing EBITDA and a 2.8% earnings yield make it the richest multiple in a chain now repricing for funding cost.
- It was among the names that led the Nov 6 Tokyo decline, so the market is already discounting a slower build-out.
- With capital intensity at 75% of revenue and a 4.6% return on equity, SMIC converts capital into the thinnest margins in the chain.
- A 0.8% trailing earnings yield on 20 times EBITDA is the opposite of a cheap AI complex.
- As the highest capital-intensity AI-linked name, it is the most exposed to any global funding-tightening pass.
- Horizon: its domestic Chinese build-out is the least substitutional revenue, so a Washington construction freeze hits it hardest.
