KKReckons research brief · 28 October 2026

China popped. America is leaking.

The AI boom is being tested in two very different ways. China's AI shares have crashed. In the US, the strain is showing up in the debt of the most heavily borrowed builders, while the biggest spenders keep spending. This brief checks the numbers behind this week's headlines and sets out what to watch.

IOU
–40%

China's STAR50 AI and chip index in under four months, about $480bn erased

1,148bp

CoreWeave's five-year default insurance, against 63bp for US investment grade

$8.5tn

Gartner's estimate of revenue AI data centres need by 2030, against at most $3.6tn they may earn

The verdict in 30 seconds

Three markets, three readings

China · equities

Burst

The STAR50 is down 40% and still trades near 100 times past earnings. Retail investors bore the losses.

US · credit

Leaking

Neocloud and Oracle debt is pricing real default risk. No neocloud has defaulted yet.

US · Big Tech

Still spending

Amazon, Microsoft and Alphabet plan up to about $600bn of capex this year on growing cloud backlogs.

1 · China

The bubble that already burst

Reuters reports that China's AI rally ran higher than any in the West and has now unwound fast. AI stocks make up 42% of China's market value and 57% of turnover, Goldman Sachs estimates, so few small investors escaped.

Traders named no single trigger. They pointed to stretched prices, tax bills and the threat of US and other bans on Chinese equipment in data centres. Beijing's 9 October guidelines on high-tech growth told officials to avoid bubbles and herd-like investment. The securities regulator has announced no rescue.

How far the darlings fell

Approximate falls from recent highs, to scale. Time windows differ, as labelled. Source: Reuters, 26 Oct 2026.

Innolight, Eoptolinkoptical parts, since Jun–Jul
~60%
Unitreerobots, since August listing
~50%
MetaX, Moore Threadschipmakers, since July
~50%
STAR50 indexunder four months
40%
CXMTmemory chips, past month
>25%
CSI300 blue chipssince June highs
~20%
Nasdaq, for contrastfrom record three weeks earlier
6%

Even after the fall, the STAR50 trades at about 100 times trailing earnings (down from nearly 250), against about 30 for the Nasdaq 100.

2 · US credit

Where the air is getting out

Neoclouds rent out GPU capacity they bought with borrowed money. When rates rise, their sums get worse even if customers keep coming. That is why the stress shows first in their bonds and credit default swaps (CDS), the price of insuring their debt.

The cost of insuring AI debt

Five-year CDS spreads in basis points (100bp = 1% a year), to scale. 26 Oct unless noted.

CoreWeaveneocloud
1,148
Oracle19 Oct, a record
315
Nvidia
127
Meta
119
US investment-grade indexCDX benchmark
63

Sources: S&P Global Market Intelligence via Benzinga (26 Oct); CNBC (Oracle, 19 Oct). Above 1,000bp is generally read as distress. Default odds implied by CDS depend on assumed recoveries, so treat them as rough.

  1. 14 Oct
    CoreWeave's 2032 bonds slip below 80 cents on the dollar, the usual distress line (Bloomberg).
  2. 15 Oct
    CoreWeave CDS hits 1,194.5bp intraday; Oracle's sets a record 288bp.
  3. 22 Oct
    Jensen Huang says neocloud demand is "off the charts", and in the same breath that capital is scarcer than demand.
  4. 25 Oct
    Elon Musk predicts "guppy-level death and consolidation" among 200-plus neoclouds and says SpaceX won't buy any.
  5. 26 Oct
    CoreWeave, Nebius and IREN fall before rebounding; Meta announces a $95bn share sale to fund AI.
  6. 27 Oct
    No neocloud default so far. Short sellers made about $6bn of paper profit on the group in October (ORTEX, via Reuters).

Musk competes with neoclouds (SpaceX sells compute); Nvidia supplies and invests in them. Read both quotes with that in mind.

3 · The maths

A $4.9tn hole, on Gartner's model

Gartner estimates data-centre owners need $8.5tn of cumulative revenue by 2030 to cover announced investment, financing costs and a 12% return on capital. It sees revenue attributable to AI data centres reaching no more than $3.6tn.

Revenue needed vs revenue expected, through 2030

To scale. Gartner, 14 Oct 2026. These are model estimates, not results.

Revenue needed
$8.5tn
AI data-centre revenue, at most
$3.6tn
$4.9tnshortfall, the difference between the two bars. Gartner says more than $3.5tn of spending could face write-downs by 2030, and over 80% of neoclouds may stop operating independently.

One boom, three different spending figures

They measure different things, so they don't contradict each other. Always say which one you mean.

~$1tn

All AI infrastructure spending by AI providers in 2026, including neoclouds and labs (Gartner).

~$800bn

Hyperscaler capex expected in 2026, as cited by Neeraj Kaushal. Not all of it is AI.

up to $600bn

Amazon, Microsoft and Alphabet combined in 2026, from company guidance (Reuters).

4 · Leak or burst?

Ed Yardeni's slow-puncture thesis

The veteran strategist told CNBC that air is coming out of the bubble without it bursting. His case: the bond market is rationing money to weak borrowers faster than rating agencies, so weak players fail or sell while stronger ones carry on.

Yardeni's AI Bubble Meter, 21 Oct

Two dials out of 100. Conditions for a bubble are high; the triggers that usually pop one are not yet.

83

Conditions

34

Triggers

The one trigger in play: the two-year Treasury yield

Rise from its December low, to scale (0–400bp). Red marks show the rises before the 2000 and 2006 busts.

air is coming out of the bubble without the bubble bursting
Ed Yardeni, CNBC, 26 Oct
guppy-level death and consolidation
Elon Musk, on X, 25 Oct, about 200-plus neoclouds
off the charts
Jensen Huang, on neocloud demand, 22 Oct
a protracted investment bust
BIS, Annual Economic Report, June 2026, as a risk if returns disappoint

Why it might stay a slow leak

  • The biggest spenders fund most capex from cash flow.
  • Cloud backlogs keep growing: nearly $1.7tn at the big three.
  • Stronger firms can buy weak ones' assets cheaply.

Why it could still burst

  • Credit is where past bubbles broke.
  • A December Fed hike would push the trigger dial up.
  • If Big Tech trims spending, neocloud debts get harder to pay, and chipmakers feel it next.
5 · History

Investors lose; the rails stay

Past infrastructure manias destroyed capital but left assets that were used later. GPUs may be the exception.

UK · 1840s

Railway mania

~1/3

of about £200m invested (roughly 40% of GDP) was lost to investors, Andrew Odlyzko estimates. The network was built and used.

US · 1996–2002

Telecom fibre

~95%

of laid fibre was reported dark in 2001. Global Crossing went bankrupt with $12.4bn of debt. Much of the fibre was lit later.

Now

AI data centres

3–5 yrs

is roughly how long a GPU stays competitive. Unlike fibre, stranded chips may age out before a second life. This is our inference.

6 · The wider economy

Is AI crowding out other investment?

Paul Krugman says it is. Columbia's Neeraj Kaushal points out that total US investment has barely moved, so AI may be reshuffling capital rather than starving the rest.

US gross domestic investment, share of GDP

New York Fed, 14 Oct 2026. Bars start at zero, to scale.

Five years before Covid
21.3%
2023 to Q1 2026
21.5%
Latest four quarters
21.1%

Scale 0–25%. The NY Fed finds investment was reallocated rather than expanded: factories, other buildings and housing weakened as AI rose. Kaushal's column dates the later period from 2022; the NY Fed post uses 2023 to Q1 2026.

Signs of crowding out

  • Non-AI business investment fell as AI rose.
  • Data centres compete for power, land and builders; the BIS flags higher electricity costs.
  • Heavy borrowing adds to pressure on rates.

Signs it isn't

  • The overall investment rate is flat.
  • Business saving has held up because the spenders are cash-rich.
  • Foreign capital can flow in, as Kaushal notes.
7 · Protection

Hedging AI is getting harder

An FT editorial argues the bigger the boom grows, the fewer places there are to hide. Here is how the usual hedges look now.

HedgeStatusThe catch
Credit default swapsCostlyInsuring CoreWeave costs about 11.5% a year; protection on Amazon, Alphabet, Meta and Nvidia hit record prices too.
GPU rental futuresNot tradingCME and Silicon Data contracts were still under regulatory review on 26 Oct. They track chip scarcity, not share prices.
TreasuriesMixedThey help in a slowdown, but not if the threat is a Fed hike. The 10-year yield is near 5%.
GoldVolatileNear $4,000 an ounce; it has swung hard this year.
Value stocks, ApplePartialLighter on AI capex, but Apple sits in the same mega-cap indices.
CashErodingInflation near 3.7% eats into real returns.

Market levels are approximate quotes from 27 Oct and will have moved. Not investment advice.

8 · The India angle

For India, AI has been a bystander trade

Foreign investors have pulled money from India to chase AI and chip stocks in the US, Taiwan and Korea. When AI sold off, Indian IT briefly became the place to hide.

First half of October

Index gains, 1–15 Oct, to scale. SBI Securities via Business Standard.

Nifty IT
+10.9%
Nifty 50
+1.7%

IT was the only large sector with foreign inflows (₹2,599 crore) while foreign investors sold ₹47,955 crore overall. Year-to-date foreign outflows reached a record ₹3.08 lakh crore by 12 Oct. The IT rally faded on 26 Oct.

The real risk for Indian IT

  • AI is pushing prices down: HCLTech's CEO says the industry is in the early phases of deflation.
  • Infosys cut its growth guidance to 1.5–3%, even as AI reached 10.5% of revenue.

The lesson from China

  • The danger is retail money piling into a theme officials seem to bless.
  • India's closest parallels are small and mid-caps and AI data-centre listings, not the Nifty.
  • Goldman's basket of 42 Indian AI enablers is up about 60%.
9 · What to watch

The next six weeks

28 Oct

Fed decision, 2pm ET (11:30pm IST); a hold is expected. Microsoft and Alphabet report after the close.

29 Oct

Amazon results, with heavy capex expected. US GDP and PCE inflation data.

3 Nov

US midterm elections, with data centres a local issue.

Week of 16 Nov

Anthropic shares reported to start trading on Nasdaq as ANT.

8–9 Dec

Fed meeting. Markets price roughly 73–80% odds of a hike, depending on the day.

The other side

The bull case is not foolish

The spenders are among the most cash-generative companies ever, and unlike in 2000, business saving has not collapsed. Cloud backlogs keep growing, and Anthropic's revenue rose roughly 14-fold in the first half. Gartner's ceiling is a model that bulls dispute. Even in the bear case, history says the infrastructure gets used and the losses fall on leveraged early financiers rather than the whole economy.

KKReckons' read: a credit event at the edges, not yet a systemic bust. But credit is exactly where past manias turned, which is why the run from 28 October to 9 December matters.
Method and caveats

How we checked

This brief starts from four pieces KK shared (Reuters, two Economic Times articles and an FT editorial) and checks their numbers against primary sources where possible. Figures are attributed and dated; estimates and forecasts are labelled.

  • Not independently verified: Goldman's 42% and 57% shares (seen only via Reuters), the chip index's 12% monthly fall and Morgan Stanley's quarter-point estimate (Barron's only), and talk of an OpenAI listing next year.
  • Fed odds moved daily in October, between about 73% and 90%.
  • Several CDS readings come from outlets citing S&P Global or Bloomberg data.
  • The GPU shelf-life point and the India read are KKReckons' analysis.

Sources