A Correction in AI Valuations Is Looking More Likely, and That View Now Comes From a Central Bank, an Asset Manager and a Bond Investor.

Slide 1

tigzig.com·the analyst's tool shed

August 2026

AI valuation red flags

Correction likely, rational prices or not

European Central Bank · 17 August 2026

The CAPE ratio is 41.2, against a dot-com peak of 44.2

The Chair of the Financial Stability Board warns that leverage and AI cross-investment could amplify the fall

The build runs at 3% of US GDP a year to 2029, twice the telecom boom that ended in 2000

Cover on hyperscaler bonds has gone from nearly 5x in February to below 2x in July, and the vendor has begun financing the customer

Magnificent Seven free cash flow fell $148bn to $59bn in two quarters, and three of the seven are now negative

The counterweight: the build is still under half the size of the housing boom, and the technology is real

Their words Financial Stability Board · European Central Bank · Apollo · KKR · Brown Advisory

Reported by Reuters · CNBC · Axios

Our data Shiller CAPE and SEC filings

Amar Harolikar, ACA · Decision Sciences & Applied AI tigzig.com

---

Slide 2

tigzig.com

The argument

One chain, four links

A dozen sources, all published since early July, each looking at a different point on the same chain. Read together they make one argument.

01

The price

Valuations are back at levels last seen in 2000, and the research says a correction follows whether or not the price is rational

02

The build

A build twice the size of the telecom boom, running faster than either of the last two cycles

03

The money

Bond investors, in rising size and at falling enthusiasm, with the vendor now financing the customer

04

The payoff

Every number above rests on cash flows that do not arrive until 2028

SourcesFull citations on the last page, every one of them linked.

Amar Harolikar 2 / 20

---

Slide 3

tigzig.comThe price

The price · 01

Back at dot-com levels

The CAPE ratio prices the market against ten years of inflation adjusted earnings, which strips out the cycle. The series runs monthly back to 1881. In all that time it has been above 40 in only two stretches: twenty-one months across 1999 and 2000, peaking at 44.2 in December 1999, and again since May this year.

0

10

20

30

40

50 1881 1929 1960 2000 2026

44.2 in Dec 1999

41.2 today

Shiller CAPE ratio, monthly, 1881 to August 2026. Our chart, from the Tremor macro tracker

"Valuations on the US stock market, as measured by the CAPE ratio, are currently close to their historical peak."

European Central Bank · ECB Blog, 17 August 2026

SourcesTremor, our own macro tracker | Rational enthusiasm or the next dot-com bubble? · ECB Blog, 17 Aug

Amar Harolikar 3 / 20

---

Slide 4

tigzig.comThe price

The price · 02

A correction either way

The blue line is what happens if today's prices are entirely rational. It still goes through a correction before the next rise. Their reason: a new technology is one company's bet, but once everybody depends on it the risk belongs to everybody, so investors charge more to carry it. Charging more means paying less today.

Fundamental Perpetual boom Plus a bubble

Stylised boom-bust scenarios. Chart: European Central Bank, 17 August 2026

"The case for expecting a correction is not dependent on whether today's prices are rational or irrational. We should be aware of that and prepare."

European Central Bank · Andersson, Breckenfelder, Corradin, Nikolov and Viola

SourceRational enthusiasm or the next dot-com bubble? · ECB Blog, 17 Aug

Amar Harolikar 4 / 20

---

Slide 5

tigzig.comThe price

The price · 03

It would not stay in America

Most European exposure to the Magnificent Seven is not held directly. It sits inside investment funds and low cost ETFs, which means the holder often does not know how concentrated it is. That structure is itself a transmission channel, because a fall forces the funds to sell to meet redemptions, which pushes prices down further.

"Euro area households, which are increasingly channelling funds into low-cost ETFs, have around €440 billion of exposures to US technology equities without necessarily being aware of the associated concentration risk."

European Central Bank · ECB Blog, 17 August 2026

"A US AI fallout would not remain a US problem."

European Central Bank · the closing line of the same post

SourceRational enthusiasm or the next dot-com bubble? · ECB Blog, 17 Aug

Amar Harolikar 5 / 20

---

Slide 6

tigzig.comThe price

The price · 04

Leverage is back, aimed at the same names

Bailey names three at once. Leveraged ETFs, momentum strategies including retail money, and hedge funds sitting in equities and sovereign debt together. That last one is his contagion channel.

"The issue is not simply that investors are borrowing more, but that leverage is interacting with high valuations and market concentration, in particular the increasing cross-investment between artificial intelligence (AI) companies and hyper scalers, in a way that could amplify a future market correction."

Andrew Bailey · Chair, Financial Stability Board

"A growing footprint of leveraged entities, such as hedge funds in equity markets, which are also exposed to sovereign debt, increase the scope for contagion risk ... I remain concerned therefore that a large shock or combination of shocks could concurrently trigger multiple vulnerabilities."

Andrew Bailey · the same letter

SourceFSB Chair's letter to the G20 · FSB, 28 Aug

Amar Harolikar 6 / 20

---

Slide 7

tigzig.comThe build

The build · 01

Three per cent of GDP, every year

Five companies, spending an amount equal to three cents of every dollar the United States produces, for three years running. In 2019 the same figure was three tenths of one per cent.

0%

0.5%

1%

1.5%

2%

2.5%

3%

3.5%

0.3% 2019

0.4% 2020

0.5% 2021

0.6% 2022

0.6% 2023

0.8% 2024

1.4% 2025

2.4% 2026

3.1% 2027

3.2% 2028

3.1% 2029

consensus forecast

Hyperscaler capex as a share of US GDP. Our chart, redrawn from Apollo Chief Economist, 6 August 2026. Hyperscalers are Amazon, Meta, Oracle, Microsoft and Google. Sources: FactSet, BEA, Haver Analytics

"The consensus expects hyperscaler capex to run at roughly 3% of GDP every year from 2027 to 2029, up from 0.3% of GDP in 2019 and 1.4% in 2025."

Torsten Slok · Partner and Chief Economist, Apollo

SourceThe AI Capex Boom Is Building Twice as Fast ... · Apollo, 6 Aug

Amar Harolikar 7 / 20

---

Slide 8

tigzig.comThe build

The build · 02

Twice the size of the fibre boom

Apollo lines the two cycles up year by year. The orange bar is the telecom and fibre buildout of the late nineties.

Telecom and fibre, from 1995

Hyperscalers, from 2022

0%

1%

2%

3%

4%

0.8%

0.6% 1995 2022

0.8%

0.6% 1996 2023

0.8%

0.8% 1997 2024

0.9%

1.4% 1998 2025

1.0%

2.4% 1999 2026

1.2%

3.1% 2000 2027

1.0%

3.2% 2001 2028

0.7%

3.1% 2002 2029

forecast

Both buildouts as a share of US GDP, lined up year by year. Our chart, redrawn from Apollo Chief Economist, 6 August 2026

"This is more than twice the peak of the telecom and fiber buildout of the late 1990s, which topped out at 1.2% of GDP in 2000 before collapsing and tipping the economy into the mildest post-war recession."

Torsten Slok · Partner and Chief Economist, Apollo

SourceThe AI Capex Boom Is Building Twice as Fast ... · Apollo, 6 Aug

Amar Harolikar 8 / 20

---

Slide 9

tigzig.comThe build

The build · 03

And still smaller than housing

The counterweight, and it is Apollo's own third chart. Residential investment reached 6.6 per cent of GDP in 2005, more than double where hyperscaler capex is forecast to peak.

Residential investment, from 1996

Hyperscalers

0%

2%

4%

6%

8%

4.4%

0.6% 1996 2022

4.4%

0.6% 1997 2023

4.6%

0.8% 1998 2024

4.8%

1.4% 1999 2025

4.7%

2.4% 2000 2026

4.8%

3.1% 2001 2027

5.1%

3.2% 2002 2028

5.6%

3.1% 2003 2029

6.1% 2004 2030

6.6% 2005 2031

6.1% 2006 2032

4.8% 2007 2033

3.5% 2008 2034

forecast

Both as a share of US GDP, lined up year by year. Our chart, redrawn from Apollo Chief Economist, 6 August 2026

"The data-center buildout is still less than half the size of the housing boom, which peaked at 6.6% of GDP in 2005."

Torsten Slok · Partner and Chief Economist, Apollo

SourceThe AI Capex Boom Is Building Twice as Fast ... · Apollo, 6 Aug

Amar Harolikar 9 / 20

---

Slide 10

tigzig.comThe build

The build · 04

The unwind is the risk, not the build

The level is the wrong thing to watch. What made 2008 severe was not how high housing got, it was how fast it came down. And this cycle is going up faster than either of the last two.

"The bottom line is that the data-center buildout is smaller than housing in level but larger in the change in share of GDP, and faster than either previous cycle."

Torsten Slok · Partner and Chief Economist, Apollo

"The same arithmetic runs in reverse: housing's unwind, from 6.2% of GDP in early 2006 to 3.0% by the end of 2008, is what made that recession severe, while telecom's much smaller reversal produced the mildest one ... A cycle that builds at 0.85 percentage points a year can unwind at a similar pace, and that, rather than the buildout itself, is the macro risk if AI demand disappoints."

Torsten Slok · the same note

The unwind Apollo is pointing at (residential investment, % of GDP)

Housing, early 2006

6.2%

Housing, end 2008

3.0%

SourceThe AI Capex Boom Is Building Twice as Fast ... · Apollo, 6 Aug

Amar Harolikar 10 / 20

---

Slide 11

tigzig.comThe money

The money · 01

The buyers are backing away

The cover ratio counts how many dollars of investor orders a bond deal draws for every dollar issued. On hyperscaler deals it has more than halved since February, while every other investment grade borrower has barely moved.

All hyperscaler deals February 2026

nearly 5x July 2026

below 2x Amazon's own bond sales March 2026

3.4x July 2026

1.6x

Dollars of investor orders per dollar of bonds issued. Our chart. Every figure is one the source states in words, from Apollo, 15 July 2026 and Reuters, 29 July 2026

"For hyperscalers, it has fallen from nearly 5x in February 2026 to below 2x in July, suggesting investors may need wider spreads to absorb additional hyperscaler supply."

Torsten Slok · Partner and Chief Economist, Apollo

SourcesCover Ratios for Hyperscaler Bonds Declining · Apollo, 15 Jul | Reuters, 29 Jul

Amar Harolikar 11 / 20

---

Slide 12

tigzig.comThe money

The money · 02

Borrowing more, at a worse price

Amazon, Alphabet, Meta and Oracle issued about 194 billion dollars of bonds in 2026 to 7 July, up 79 per cent on the whole of 2025. Goldman expects the five, adding Microsoft, to reach roughly 250 billion this year and 400 billion next. Of the bonds issued this year with comparable pricing, 78 of 91 now trade above the yield they were sold at.

Median new issue concession, basis points

The extra yield a borrower must offer, against its existing debt

2025

2.25 bp

2026

12 bp

"I don't see sort of a point in time in the near future where that sort of supply goes away. Technically, that puts pressure on spreads ... We're already seeing fatigue within credit markets in supporting this massive debt issuance."

Colby Stilson · head of fixed income, Brown Advisory

SourceHyperscaler debt binge pushes yields up ... · Reuters, 29 Jul

Amar Harolikar 12 / 20

---

Slide 13

tigzig.comThe money

The money · 03

The vendor is funding the customer

In late July, Nvidia was reported to be weighing a 250 billion dollar guarantee on an OpenAI data centre in Ohio, and separately a deal to finance OpenAI's purchase of 350 billion dollars of Nvidia chips. Its shares fell 4.5 per cent that morning, and the cost of insuring its bonds against default posted its sharpest intraday rise since they began trading actively in November. Andrew Bailey named this same structure on page 6, calling it cross-investment between AI companies and hyper scalers.

"You buy from me, I invest in you, and everything's fine unless one of us has a problem, in which case both of us have a problem."

Axios · on why the possibility of more circular deals soured the debt market that Monday

"It also renews concerns that the increasingly 'circular' nature of AI deals could create systemic financial risks."

Axios · the same report

SourcesNvidia reignites "circular" AI concerns · Axios, 27 Jul | FSB Chair's letter to the G20 · 28 Aug

Amar Harolikar 13 / 20

---

Slide 14

tigzig.comThe money

The money · 04

Now compute itself is a commodity

On 5 October, pending approval, CME and Silicon Data list two futures contracts on the hourly rental price of Nvidia H100 and Blackwell chips.

The case for it

"Compute futures give the market ... a public, tradable reference price for the resource every AI system runs on."

Carmen Li · chief executive, Silicon Data

"You can think about it as a revenue stream, and you can securitize it ... and sell it to investors."

Waldemar Szlezak · head of digital infrastructure, KKR

The note of caution

Oil and electricity trade this way. A chip is different, because it ages. Michael Burry argued late last year that the hyperscalers overstate the useful life of their AI chips and understate the depreciation.

"When Wall Street starts getting noticeably excited about securitizing physical assets, a natural question emerges: What could go wrong?"

CNBC · on the same announcement

SourcesAI computing power is a tradable asset class · CNBC, 11 Aug | Wall Street endorsed Huang's big concept · CNBC, 11 Aug

Amar Harolikar 14 / 20

---

Slide 15

tigzig.comThe payoff

The payoff · 01

Everything rests on 2028

Consensus has hyperscaler free cash flow falling sharply in 2026, recovering slowly through 2027, then more than doubling by 2030. Every valuation, every bond and every spread on the pages before this one is priced off the right hand side of this chart.

Free cash flow for Google, Meta, Microsoft and Amazon, actual and consensus forecast. Chart: Apollo Chief Economist, 9 July 2026. Source: FactSet

"But what if the payoff takes longer than consensus assumes?"

Torsten Slok · Partner and Chief Economist, Apollo

SourceA Slower AI Payoff Would Be Everyone's Problem · Apollo, 9 Jul

Amar Harolikar 15 / 20

---

Slide 16

tigzig.comThe payoff

The payoff · 02

Three of the seven are already negative

The same turn shows up in our own numbers, on a wider basket. Across the Magnificent Seven, net free cash flow fell from $148 billion in the December quarter to $59 billion in June. Alphabet, Amazon and Tesla were each below zero.

107 Q4 24

90 Q1 25

78 Q2 25

114 Q3 25

148 Q4 25

98 Q1 26

59 Q2 26

Net free cash flow, US dollars billion

Seven quarters to June 2026. Our chart and our data, from SEC EDGAR XBRL filings

Free cash flow, June quarter 2026, $bn. Nvidia is the quarter ended 26 July

Apple +$31.9B

Nvidia +$21.4B

Microsoft +$19.6B

Meta +$1.7B

Tesla -$1.1B

Alphabet -$5.9B

Amazon -$8.8B

Alphabet has never had a negative quarter before.

Its first negative quarter in 22 years as a public company.

SourcesOur own pull from SEC EDGAR | Alphabet Q2 2026 earnings release

Amar Harolikar 16 / 20

---

Slide 17

tigzig.comThe payoff

The payoff · 03

And the moat is leaking

Two things have to hold for that 2028 cash flow to arrive. Prices have to stop falling, and the American models have to keep their share. Chinese models were five of the world's fifty most used in January 2025. By May 2026 they were about twenty, read off Apollo's unlabelled chart.

China United States France Canada Unattributed

Origin of the world's 50 most-used AI models. Chart: Apollo Chief Economist, 9 July 2026. Source: OurWorldinData.org. Legend redrawn above at a readable size

"If Chinese models keep gaining and token prices keep falling, the hyperscaler cash flows expected may prove too optimistic."

Torsten Slok · Partner and Chief Economist, Apollo

SourceA Slower AI Payoff Would Be Everyone's Problem · Apollo, 9 Jul

Amar Harolikar 17 / 20

---

Slide 18

tigzig.comThe payoff

The payoff · 04

If it comes late, three things break

Apollo set out what a slower payoff does, in order.

1 Cash flows and earnings disappoint. The free cash flow surge slips later while committed capex and heavy depreciation hit on schedule, squeezing margins.

2 A Mag 7 sell-off that takes the market with it. Because those seven are so much of the index, the pain cannot stay contained. It spreads to chips, power, data centres and the S&P 500 as a whole.

3 Balance sheets stretch and credit risk rises. With internal cash unable to cover spending, the hyperscalers lean further on debt, inviting possible ratings downgrades if profits lag.

"AI has been the one thing holding up both the economy and markets, and with so much riding on so few names, a slower payoff wouldn't just be a sector problem, it would risk tipping the economy into recession and the S&P 500 into a correction."

Torsten Slok · Partner and Chief Economist, Apollo

SourceA Slower AI Payoff Would Be Everyone's Problem · Apollo, 9 Jul

Amar Harolikar 18 / 20

---

Slide 19

tigzig.comMy read

My read

I have been through a few of these

My own view is that we are at an overvalued point, and this is one of several bubbles waiting to burst.

The dot-com one I remember very clearly. That time it was all about page views, with no revenue, never mind profit. This time the revenues are there, that much is real. It is the profits and cash flows people are counting on that carry the question mark.

Both times the technology itself was real, and in the end all of the value did arrive. It just did not arrive in the amount the market was expecting, or inside the compressed timeframe it was pricing. So the thing went through its whole cycle first.

In the market since 1994. Ex derivative trader (full time, proprietary account).

The ECB economists made the same point in their own words: "The exact timing is unknowable in advance. These boom-bust patterns are only identifiable with hindsight."

More of my work on fundamentals, macro and markets tigzig.com/analysis

Amar Harolikar 19 / 20

---

Slide 20

tigzig.com·the analyst's tool shed

For the record

Sources

Every quotation is verbatim. Charts are either the publisher's own, reproduced with credit, or ours, drawn from the figures the publisher printed and labelled as such. All published since early July.

The price

FSB, 28 Aug FSB Chair's letter to G20 Finance Ministers and Central Bank Governors

ECB Blog, 17 Aug The AI boom: rational enthusiasm or the next dot-com bubble?

The build, and the payoff

Apollo, 6 Aug The AI Capex Boom Is Building Twice as Fast as the Housing Boom

Apollo, 15 Jul Cover Ratios for Hyperscaler Bonds Declining

Apollo, 9 Jul A Slower AI Payoff Would Be Everyone's Problem

The money

Reuters, 29 Jul Hyperscaler debt binge pushes yields up as investor demand cools

Axios, 27 Jul Nvidia reignites "circular" AI concerns as it weighs OpenAI financing guarantee

CNBC, 11 Aug AI computing power is becoming a tradable asset class as CME launches futures contracts

CNBC, 11 Aug Wall Street just endorsed Jensen Huang's 'big concept' for AI. What now?

Alphabet, 22 Jul Alphabet Q2 2026 earnings release, with its own free cash flow reconciliation

Our own data

Tremor Shiller CAPE, monthly since 1881. Free, no sign up, full data download

Amar Harolikar · more at tigzig.com/analysis 20 / 20