AI Bubble: a16z Says No. So Who's Left Holding It?

Click here to watch a video of this episode.
Andreessen Horowitz has published 90 slides arguing that AI is not a bubble, one month after closing a $1.1 billion fund for AI hardware. On slide 88, holographic Pokémon cards are up 110% since January 2024, against 72% for the Nasdaq. The real question is not whether there is an AI bubble, but who is holding it when the next buyer pauses.

On 30 September, a16z's Growth team, led by David George, published State of Markets II. The deck makes a serious case: tech delivered about 76% of the S&P 500's earnings growth this year, and valuations fell while profits rose. The twist is in the financing. Meta's Hyperion data centre sits in a vehicle 80% owned by Blue Owl's funds, the Chicago Fed puts large US banks' committed AI-related lines at about a quarter of their core capital, and the AI bubble debate has quietly moved from shareholders to banks, insurers and pension funds.

Aman Narain and Zubin Vandrevala boil the deck down to five mega trends and ten takeaways for financial services, score every argument, and walk through, day by day, what happens if the AI bubble bursts.

Key takeaways:
1. a16z is right about the companies and silent on who pays for them: in one quarter Alphabet and Amazon booked about $150 billion of paper gains on private AI stakes, roughly 12% of S&P 500 earnings on Goldman's estimate.
2. Finance has become the balance sheet of AI: Meta's Hyperion data centre sits in a vehicle 80% owned by Blue Owl's funds, with about $27 billion of bonds behind it and Meta as the tenant.
3. The Chicago Fed puts large US banks' drawn AI-related loans at about 9% of Tier 1 capital and committed lines at about 25%. The gap is what nobody sees until a bad week.
4. AI efficiency is real and gets competed away. The saving goes to the customer, and part of the AI budget comes straight out of headcount.
5. If the AI bubble bursts, pensions and insurers face a markdown rather than a wipe-out. The danger is that every loss arrives in the same week.

Topics covered:
- Why a16z's deck deserves both respect and a disclaimer: the umbrella seller's weather forecast
- Five mega trends, argued both ways and scored: earnings, capex, atoms, cheap intelligence, private markets
- The merry-go-round between cloud giants, AI labs and lenders
- Linda in Loudoun County, and what data centres do to a household electricity bill
- How Meta rents its own Hyperion data centre, and why Steve Eisman reaches for Enron
- Graham's pension: one capex cycle in four wrappers
- Chip-backed loans, data-centre securitisations and a CDO question from London in 2000
- Fintech funding splitting in two: Stripe and Revolut, then everyone else
- Agentic payments: machines paying machines before consumers do
- What happens to pensions and life insurance if the AI bubble bursts, and the likelier plateau
- Four dated predictions, to be checked on 6 October 2027

Chapters:

• (00:00) - Cold open: Pokémon cards beat the Nasdaq
• (00:53) - San Francisco to Vegas: Metallica, then Money20/20
• (01:39) - Singapore: fintech, fatherhood and F1
• (02:16) - A confession: the one thing we got wrong
• (02:56) - Meeker, Evans and a16z's 90 slides
• (03:29) - Disclaimer
• (03:49) - A weather forecast from a man selling umbrellas
• (04:37) - Five mega trends, ten takeaways, one cauldron
• (05:15) - Mega trend 1: profits are outrunning prices
• (05:55) - A snake eating its own tail: $99bn of paper gains
• (06:53) - Mega trend 2: $780 billion of capex, on credit
• (07:58) - The merry-go-round and the three witches
• (09:06) - Mega trend 3: "Atoms are so back"
• (09:48) - Linda in Loudoun County: the $14 to $37 bill
• (10:40) - The GPUs arrive on time. The grid does not
• (11:07) - Mega trend 4: Jevons paradox and cheap intelligence
• (12:09) - The AI budget is your headcount
• (12:27) - Mega trend 5: unicorns outweigh the Russell 2000
• (13:26) - Prof Z: how Meta rents its own $27bn data centre
• (14:38) - Eisman, Enron and where the risk ends up
• (15:05) - Part two: Graham in Surrey, Priya in Pune
• (15:54) - Takeaway 1: finance is the balance sheet of AI
• (16:42) - Takeaway 2: one bet in four wrappers
• (17:27) - Takeaway 3: a six-year loan on a chip
• (18:21) - London, 2000: "Nobody loses. Everybody wins."
• (20:13) - Takeaway 4: private markets, sold late to retail
• (21:03) - Takeaway 5: the fees are in power, not silicon
• (21:48) - Takeaway 6: fintech funding splits in two
• (22:29) - Takeaway 7: the bank's card ledger as macro data
• (23:04) - Takeaway 8: AI savings get competed away
• (24:18) - Takeaway 9: machines paying machines
• (25:08) - Takeaway 10: the assistant is the new shelf
• (25:45) - If the AI bubble bursts: seven days, day by day
• (27:57) - The adult in the room: correction, then hangover
• (28:34) - What happens to your pension if it bursts?
• (28:54) - Closing wagers: Anthropic or OpenAI to list first?
• (30:02) - The mic drop: the slide a16z didn't write

Referenced in this episode: a16z, State of Markets II (30 September 2026); a16z's $1.1 billion AI hardware fund (28 August 2026); Goldman Sachs on Q2 investment gains; Meta Hyperion and Blue Owl; Moody's on $662 billion of leases not yet commenced; Chicago Fed on AI tail risk for banks; JLARC Virginia on household bills; the VoxEU study of about 1,200 utility territories; the Bank of England FPC record (30 September 2026); Mary Meeker's Internet Trends; Benedict Evans's annual presentation; Macbeth, Act 4, Scene 1.

Related episodes: Why the AI Bubble Won't Burst (Even If They Say It Will); S2E28 — Revolut-ion; Could a Kitchen Table in Ohio Decide AI's Future in November?; S2E17 — SpaceX's $2 Trillion IPO vs the AI Listing Race; S2E16 — Polymarket, Kalshi and the $1B Bet on Who Knew First.

Hosted by:
Aman Narain writes at amanwhoblogs.substack.com. Zubin Vandrevala is your payments provocateur.

Enjoying A2Z Fintech? Leave a rating and review on Apple Podcasts. It is the single biggest signal to the Apple algorithm and how new listeners in our world find us.

For information and entertainment only. Not financial advice.

Transcript:

Creators and Guests

Host

Aman Narain

Boomerang Banker, Ex-Googler. Platforms, Ecosystems & Transformation junkie. Startup Investor & Advisor.

Host

Zubin Vandrevala

A2Z Fintech by Night. Spearheading revenue growth at Gr4vy by Day.

What is A2Z Fintech?

Aman Narain and Zubin Vandrevala have spent over 25 years in fintech across Banks, BigTech, and Startups. This is a podcast of them riffing on payments, fintech and everything in between.

Aman: Zubin, the most surprising chart in
a 90-slide deck from Andreessen Horowitz,

the most powerful venture firm
in the world, isn't about Nvidia.

It's about the cards
in your kid's shoebox.

Since January 2024, an index
of holographic Pokémon cards is up 110%,

Nasdaq about 72% and Microsoft about 35%.

Zubin: So Aman, nobody's
calling Pokémon a bubble.

Aman: Not while the prices keep rising.

A collectible is worth exactly
what the next person will pay.

Which is the real question about AI.

Not "is it a bubble?", but "who's holding
it when the next buyer pauses?"

Zubin: Welcome back to A to Z Fintech,
or as Aman calls it, A to Zed Fintech.

I'm Zubin, the Z, in San Francisco,

although I'm starting to feel
like a guest in my own home state.

Aman: How many trips this month, Zubin?

Zubin: Well, I just got
back from Nashville.

I've got two more trips
to Vegas, one to Austin.

The first trip to Vegas
is actually with my son,

my fellow musician
and fellow Metallica fan.

We've got tickets to see Metallica,
who's in residency at Sphere.

I'm super pumped about it.

And then I'm straight back into
Vegas for Money20/20 the week after.

And then, of course, Formula
One with my brother in Austin.

Aman: That's amazing. From Master
of Puppets to the masters of payments.

Zubin: That's right.
Same black T-shirts, more lanyards.

Aman: And I'm Aman, in Singapore.
I'm off to Dubai in two weeks.

But before that, the girls and I are
beside ourselves about F1 weekend coming

up here in Singapore.

Zubin: So that's Metallica with my son,
F1 with your daughters,

and I guess Money20/20 for the day job.

Aman: Looks like A to Z is living
the life of fintech, fatherhood and F1.

Zubin: Three Fs. Ask me about the fourth,
fatigue, maybe in November.

Aman: Yeah. And speaking of our kids,

any parents who threw out
a shoebox of Pokémon cards in 2023

should probably look away now.

Zubin: Well, I should
be checking my attic as we speak.

But before we go anywhere, a confession.

A couple of weeks ago,
on this very channel,

we published a video called
"Why the AI Bubble Won't Burst".

Aman: We did, and I stand by it.

Zubin: So why are we back?

Aman: Because we got one thing wrong.
We asked whether it would burst.

We never asked who'd be left holding it if
it did. And quietly, the answer changed.

It used to be shareholders. Now it's
your bank, your insurer and your pension.

Zubin: So I guess this is a sequel.

Aman: Yeah, the sequel nobody
asked for, but everybody needs.

Zubin: So some quick context, because
there's a whole genre out there, right?

Every year a few reports tell
you where the internet is going.

Aman: And Mary Meeker's
Internet Trends was the OG.

She's been doing it since 1995.
I think she's back.

Ben Evans's annual presentation on the
state of tech is another one everybody,

including myself, waits for.

Zubin: That's right.
And then on the 30th of September,

a16z published State of Markets:
90 slides from their Growth team,

led by David George, the general
partner who runs their Growth fund.

Aman: And I think, Zubin,
it's disclaimer time.

Zubin: Of course. Let's jump into it.

Guys, if you've been listening
to this podcast,

you know this is solely for information
and entertainment purposes.

It's two dudes riffing about fintech
and everything in between,

not investment advice. And neither of us
owns any Charizards, so stay smart.

Aman: And the most important
disclaimer is about the report.

We respect a16z enormously.

Zubin: Totally, we absolutely do.
But context matters.

On the 28th of August, a16z closed a $1.1
billion, with a B, fund for AI hardware.

And guess who was the lead among them?
George was the lead.

A month later, they published a deck
saying hardware is the winning trade

and AI isn't a bubble.

It's a weather forecast from a man
who's selling umbrellas.

Aman: A very good forecaster who
sells exceedingly good umbrellas.

Zubin: That's right.

Aman: And the closing
section leans heavily on data

from their own portfolio companies.

Zubin: Which, we must admit,
doesn't make it wrong.

Aman: It doesn't. So here's how today
works. a16z doesn't hand you a neat list.

It's 90 slides, a cauldron of ingredients.

We've boiled them down to five mega
trends for the internet and ten takeaways

for financial services and fintech.

Zubin: Right.
Bubble, bubble, toil and trouble.

Aman: Which, for the record,
isn't the line, Zubin.

Zubin: All right.
Since you threw me there,

I'll go back to my twelfth-grade
English class.

I know, it's Macbeth, Act Four:
"Double, double, toil and trouble."

Let's throw it in there: "Fire burn and
cauldron bubble." But in this market...

Aman: It's bubble, bubble. Part one.

Zubin: All right.

Aman: The five mega trends.
Zubin has agreed to defend a16z.

Zubin: Under protest, I must admit.

Aman argues the other side, and we score
every one. All right, so let's jump in.

Mega trend number one,
and I'll start strong.

Tech delivered about 76% of the S&P 500's
earnings growth this year. That's crazy.

76% of all of the S&P 500.

Aman: Unbelievable.

Zubin: Forward earnings
are up about 35% in a year,

and the P/E ratio has actually
fallen from about 22.5 to under 20.

A bubble is when prices
run ahead of profits.

Here, profits are running ahead of prices.

Aman: Yeah, it's a lovely chart.
Now look inside the profits.

In Q2, Alphabet booked about $99 billion
of gains on its stakes in Anthropic

and SpaceX. Amazon booked
$53 billion on Anthropic.

Zubin: So, paper gains.

Aman: Paper gains that Goldman reckons
were about 12% of the S&P 500's earnings

last quarter. 12%. So a private equity
round inflates a public company's profits,

which lowers its P/E,
which proves it isn't a bubble.

Zubin: Right, like a snake eating
its own tail and calling it a diet.

Aman: That's an interesting
mental picture.

Zubin: Yes. But strip away those gains
and earnings still grew about, what, 29%?

So that's not really a snake. That's,
I would argue, a pretty healthy animal.

Aman: It is, and it's real growth.

Zubin: So I'll take the point.

Aman: Point to you, and I guess
it's concession number one.

Zubin: All right, I'll take that.
Let's jump into mega trend number two.

The cloud giants will spend about $780
billion this year. $780 billion.

That's insane. But they're also the most
profitable companies in history.

Goldman reckons that they
could borrow another $1.3,

$1.4 trillion and still sit comfortably.

Borrowing to build assets that earn money
is exactly what companies are supposed

to do. And by the way, we covered
this in our AI episode as well.

Aman: Yeah, except that free cash flow
turns negative this year and next.

So they're borrowing: about $320 billion
of bonds and off-balance-sheet vehicles

by August, against under $50
billion in a normal year.

Zubin: But markets are happy to lend it.

Aman: Less happy than it was.

In February, their bond deals were five
times oversubscribed. By July, under two.

Oracle has been cut
to one notch above junk.

Zubin: But this is still describing
borrowing, right? This is not a bubble.

Aman: Then follow the money
around in a circle.

A cloud giant invests in an AI lab.

Zubin: Amazon, Anthropic.

Aman: The lab promises to buy its cloud
capacity. That's Anthropic's $518 billion.

The promise becomes backlog,
now about $1.7 trillion.

The backlog justifies the borrowing.

Zubin: I see the loop, right?

So the borrowing becomes,
basically, Nvidia's profit.

Aman: And the lab raises again at a higher
price. And the cloud giant books its gain.

Every step is real business.
Together, it's a merry-go-round.

And the banks are selling the tickets.

Zubin: Yeah.
I don't have a good answer to that.

And if I had to quote Macbeth once again,

it's the three witches around one
cauldron: the cloud giants, the labs,

the lenders, each stirring
the other's pot, almost.

Aman: My literature professor, Mr Guzder,
is going to be delighted by this.

Zubin: You know the crazy part?

I bumped into a school
friend from 30 years ago,

and we were talking about Aspi Kapadia,
who was our literature teacher,

literally yesterday over coffee.
So this is really appropriate.

Serendipity. Let's move on.

Aman: Serendipity. So, point to me.

Zubin: Point to you. OK, I'll give
that to you. Mega trend number three.

And I think this is the best
slide title in the deck,

and we were laughing about this.

Aman: I love it.

Zubin: Which is: "Atoms are so back."

The money has moved from software to
physical things: chips, memory, cooling,

power. Real factories, real jobs.

About 300,000 extra construction
jobs tied to data centres.

And a16z even shows research saying data
centres might lower your electricity bill.

Aman: On jobs, agreed.

On the bill, and we covered this as well,
you're the only one living in America.

Why don't you introduce
our first character?

Zubin: Yeah, let's do that.
And this is one we had some fun with.

So let's call her Linda. She's
a composite, she's not a real person.

Let's assume she lives
in Loudoun County, Virginia,

which is the data-centre
capital of the world.

Aman: Didn't know that.

Zubin: Virginia's own projections suggest
households there would pay an extra $14,

and up to almost $37, a month by 2040.

Aman: Yeah, and a study
published a week before the deck,

covering about 1,200 utilities, found one
large data centre adds about 4% to local

household prices.
70% of Americans don't want one next door,

and New York has even got a moratorium.

Zubin: Yeah. And of course, since I'm the
American, and the capitalist American,

in fairness to a16z, spreading
the grid's fixed costs over more

customers is real economics.

Aman: And I think it is, early on.
Concession number two.

But here's the twist:
the GPUs arrive on time.

It feels, Zubin, everything else is late.

Zubin: Yeah, like a transformer
takes two to five years.

Order a gas turbine today
and it turns up, what, 2031?

Aman: So the bottleneck, I would argue,
has moved from Nvidia to the grid and,

arguably, the ballot box come November.

Zubin: Yeah, let's call this a draw.

Aman: OK, I'll take a draw.

Zubin: All right. So what are we? One-one,
one draw. Yeah, I think we're doing good.

Mega trend four. Intelligence is getting
cheaper and usage is exploding.

On one platform, OpenRouter, token
volume is up 27% [27 times] in a year.

It's Jevons paradox. Make something
cheaper and people use far more of it.

Take coal in the 1860s, compute today.

Aman: There's some high intellectual
stuff going on here, Zubin.

You know, Macbeth, Jevons paradox.

Zubin: I prep well, my friend.

Aman: Three buts. First, 86%
of the tokens AI agents use are cached,

which are billed at a steep discount,
so volume kind of flatters revenue.

Zubin: Fair point. What about second?

Aman: The top 1% of companies spend
about 600 times the median per employee.

And third, about 2% of US
households pay for AI. 2% only.

Zubin: So if I had to imagine this,

it's a party where about a hundred people
are dancing and you, my friend Aman,

have bought the entire bar.

Aman: And the chart every
bank COO should frame:

only about a third of AI
budgets are new money.

And this is important: 18% comes
straight out of labour budgets.

Zubin: So the AI budget...

Aman: ...is your headcount. Point to me.

Zubin: I'll give that to you.
I have no rebuttal.

Let's talk about mega trend five,
and this is the last one.

American unicorns are now
worth, what, $5.3 trillion.

And the entire Russell 2000
is about $3.5 trillion.

Aman: Wild.

Zubin: The value creation is real.
It's happening in private markets.

And that's the uniqueness of this.
That's where the growth is now.

Aman: Yeah, and when they finally list?

SpaceX floated in June, as we covered,
at about $1.75 trillion, 94 times sales,

with about 5% of its shares.
It has gone roughly nowhere since day one.

And the median venture fund from 2021
has returned no cash at all.

Zubin: Aman, even I can't defend this one.

The public gets invited to the party
when they're stacking the chairs, right?

Aman: Good, I'm running away with this.
Point to me. What's the final score?

Zubin: All right, I'll give this to you.
Aman, my lawyers will be in touch.

Aman: And what's the verdict?

Zubin: All right.
a16z is right about the companies.

It's just not asking
who's paying for them.

Aman: Prof Z, before part two, take us
through how Meta built its Hyperion data

centre in Louisiana without putting
$27 billion on its balance sheet.

Zubin: A bit of research, and I'm glad
we're calling it out separately.

Three moves. One: they set up a separate
company, a special purpose vehicle.

Two: outside investors own most of it.

Funds run by Blue Owl hold almost 80%,
and Meta only holds about 20%.

And the vehicle borrows roughly
$27 billion in the bond market.

And then the third point is Meta rents
the finished building for up to 20 years,

and promises to cover the downside if the
building isn't worth what it should be.

Aman: That's fascinating.
So Meta is renting its own building.

Zubin: From people who borrowed to buy it,
using Meta's promise as their comfort.

Aman: So why not just borrow directly?

Zubin: That's an interesting one.

And we wanted to call this out:
this is rather unique,

and something that is worth
us paying attention to.

It keeps the debt off the headline
balance sheet, protects the credit rating,

and shares the risk.
And it's everywhere, right?

Moody's counts about $662 billion of
leases like this that haven't even started

yet across the cloud giants.

Aman: Steve Eisman, of The Big
Short, has compared these

structures to Enron's. Is that fair?

Zubin: Yeah, but I don't know.

Auditors flag it as a critical audit
matter, which tells you it's not simple.

The real point is where the risk ends up.

That means pension funds, that means
insurers and wealthy families.

Aman: Da Vinci said that simplicity
is the ultimate sophistication.

That sounded very
sophisticated, Professor.

Zubin: I'll keep adding those
stars to my trophy board.

Aman: OK, part two. Ten takeaways for the
financial services and fintech industry,

which is what you guys dial in here for,
in three groups: who's holding it,

who's making the money,
and who's being disrupted.

And we'll follow two more people.

Zubin: Yeah, we talked through this one,
so let's take the first one.

This is someone, again a composite,
let's call him Graham.

He's 64, lives in Surrey, and his pension
sits in a nice, sensible balanced fund.

Aman: And let's talk about Priya.

Let's call her 31, lives in the IT
services town of Pune, where I grew up,

and has a US tech fund in her monthly SIP.
Both composites.

Zubin: Of course.
And let's not forget Linda,

still paying her electricity
bills in Virginia.

Aman: Poor Linda.

Zubin: Right, takeaway one.
Aman, why don't you take this one?

Aman: OK, so finance has become
the balance sheet of AI.

Hyperion is the template.

The risk that Meta's shareholders used to
carry now sits with Blue Owl's investors.

Zubin: And banks lend to all of them.

Aman: Yeah, in fact, really important:

the Chicago Fed estimated that large
US banks have drawn AI-related loans

of about 9% of their core capital, but
with committed lines of 25%. A quarter.

I know you know what 25% is,
but I just thought I'd say that.

Zubin: Yeah, but I think
that's the thing, right?

Aman, the gap between 9 and 25 is the bit
nobody sees until it's a bad week.

All right, let me take
takeaway number two.

And this one is about Graham,
the composite we talked about,

the 64-year-old guy living in Surrey.
He thinks he's spread out, right?

Equities, bonds, a bit of private
credit, maybe an infrastructure fund.

Aman: And it's one bet in,
essentially, four wrappers.

AI-linked companies are about half
of the S&P 500 by value.

AI borrowers were about a quarter of this
year's investment-grade bond issuance.

Zubin: A third of private
credit deals last year were AI.

And what about his infrastructure fund?

Aman: Data centres. Graham is long
the same capex cycle four times over.

Zubin: Do you know how much of yours
rides on five companies' spending plans?

Tell us in the comments. All right, let's
jump to takeaway number three. Collateral.

Aman: OK, so chip-backed loans
are now investment grade.

CoreWeave borrowed at an A3 rating
on loans that run up to 2032.

But a16z's own charts show
the older A100 chip losing about 37%

of its value in 17 months.

Zubin: So it's a six-year car loan on a
car that loses a third of its value before

its second service. And the funny
part: we've covered Kalshi before,

but Kalshi now runs a market on the hourly
price of renting an Nvidia B200.

So compute is turning into a commodity.

Aman: Yeah, I love the way you put it.

And these loans are being
packaged up and sold on.

JPMorgan expects $30 to $40 billion
a year of data-centre securitisations,

which takes me straight back to my first
year at Standard Chartered Bank.

So let me tell you a bit of a story.

I was a management trainee in the
corporate banking team in London,

surrounded by relationship managers and
credit heads who'd been doing their jobs

longer than I'd been alive.
Intimidated doesn't begin to cover it.

The hottest new thing
when I joined banking was the CDO.

Someone from credit was explaining
how these were packaged, Zubin,

and I put my hand up and I asked, "Sir..."

I didn't say sir, but I thought I'd
go with the Shakespearean thing.

"Do these contain bad
loans as well as good ones?

And if they do, who loses money?"
The room went silent.

The two most senior people looked at me
as if I'd asked something ridiculous.

"Nobody loses. Everybody wins." And the
room, mostly men, chuckled knowingly.

My fellow management trainees,
still friends, smiled nervously at me.

And I knew all of them were making
a mental note never to ask

a stupid question like that again.

Zubin: Yeah, I mean, you're keeping
with the Shakespearean theme.

"Nobody loses, everyone wins." That's
almost straight out of Macbeth, right?

The witches tell him no man
born of a woman can harm him.

Sounds like a guarantee,
right up until it isn't.

Aman: It took a global financial crisis
for me to realise that stupid question

was probably the most valuable
one in that room on that day.

I've never stopped asking stupid
questions, as you know, since then.

Sometimes it's the stupid
question that saves a bank.

Zubin: So what is today's
version of that question?

Aman: Yeah, I've thought about it,
and it is this, which is the nub

of this episode: who's holding it?

Zubin: Let that one sit, and let's come
back to that. So let's shift.

Here's takeaway four,
the wealth piece of this.

Aman: Private markets are being
sold to retail investors late,

and with less liquidity
than the label suggests.

This year, Cliffwater received
redemption requests for 14%

of its fund against a 7% cap.

Zubin: And what about in your part
of the world, Aman?

Aman: Yeah, every relationship manager
here in Singapore that I know of is being

asked for pre-IPO access
to Anthropic, OpenAI.

The honest question for their client is:

how much cash has this asset class
actually returned? Not what is the IRR.

Zubin: Yeah, it's a bit like:
the IRR is the photo on your dating app,

while cash returned
is who actually turns up.

Aman: You are on a roll today, Zubin.

Zubin: I am.
It's Friday evening and I'm feeling good.

Aman: I'll do the boring end of this.

Zubin: All right, takeaway five.
Something cheerful, please. Come on, Aman.

Aman: All right, I'll try.
I've been Mr Doom. Here's one.

The lasting fees are in power,
grid and buildings, not silicon.

Project finance is up 54% this year,

and Morgan Stanley's capital-markets fees
jumped from $1.4 billion to $2.3 billion

in the first half of this year.

Zubin: And Aman, for Asia, the deck's list
of scarce components is basically a client

list, right? TSMC, SK hynix,
Samsung, Hitachi Energy.

Aman: Yeah, a chip, I guess,
lasts three or four years.

A substation lasts 40. Bankers like 40.

Zubin: That's right.
All right, takeaway six.

And I'll take this one
because it's really from the Valley.

Fintech capital markets have split in two.

Three of the ten most valuable private
companies in the deck are fintechs.

Stripe at, what, $159 billion?

Revolut, we just covered this,
at $115 billion, and Ramp at $44 billion.

Everyone else is on rations.

Fintech deal count in Q2 was the
lowest in more than four years.

Aman: Our Revolut episode has aged rather
well, Zubin. So the bargain bin is open.

Zubin: Right.
And the banks have the trolley, right?

So it's the best time since 2022 to buy
capabilities, teams and licences.

All right, takeaway seven, mine again.
Financial data has become macro data.

And the deck's best chart that we could
see was on whether households pay for AI.

And that comes from PNC, a bank.

And Stripe's data shows
software revenue recovering

while software shares were crashing.

Aman: Yeah, so the most valuable dataset
in the AI economy is sitting, arguably,

in a bank's card ledger.

Zubin: And the sad part is most
banks never publish it, right?

The ones that do build a franchise, and
the ones that lend with it lend better.

Aman: So, takeaway number eight.
This is Priya's.

AI efficiency is real,
and it gets competed away.

Lemonade's claims-handling
cost is 5% of its premium,

against about 9% for incumbents.

Chime, the US neobank, has cut its cost
to serve 10% a year for four years.

Zubin: But Aman, Lemonade's shares also
still fell, what, 15% on its results?

I think Jamie Dimon put it best:
you don't uniquely benefit from AI.

The savings go to the customer.

Aman: True. Hard to argue with Mr Dimon.
Now, let's connect it to Priya.

And remember, 18% of AI
budgets come out of labour.

A big slice of the world's
outsourced labour budget is spent

in Bengaluru, Pune, Chennai.

Zubin: I mean, that's a question,
not a forecast.

Aman: It is a question, and the right one.

About 63,000 bank job cuts
have been announced this year.

And in fact, here in Singapore,
the Deputy Prime Minister has asked banks

the question: what new
jobs are you creating?

Zubin: And the winners turn savings into
lower cost and wider lending, right?

So margins get competed away.

Takeaway nine, and I'm taking this one
to Money20/20 with me. Agentic payments.

Consumer agentic shopping has,
let's just call it, stalled.

OpenAI shut its Instant Checkout in March,

and Walmart saw conversions
three times lower in chat.

Aman: So Zubin, is it dead?

Zubin: No, we're looking
in the wrong place, right?

Agents on one platform alone burn about,
what, 50 trillion tokens a week,

and every call is a metered purchase.

The first big agentic payments
market is machines paying machines.

Visa is live with more
than 30 European issuers.

Mastercard launched its
agent-identification scoring at the end

of September. And I guess US
stablecoin laws kick in in January.

Aman: So Zubin, the question
is, can the agent pay?

Zubin: It's who's liable when it
buys the wrong thing. All right,

takeaway ten. Mr BankBazaar.

Aman: Search traffic fell 31% in a year,

and referrals from AI
tools are close to zero.

Priya doesn't Google
credit cards any more.

She asks the assistant,
and she stops there.

Zubin: So every comparison site and every
neobank buying search ads is fishing in a,

I guess, shrinking pond.

Aman: The assistant is the new shelf.

If you're a bank, your product
has to be one of its picks.

Zubin: SEO is dead. Long live AEO!

Aman: Please don't make that a Short.

Zubin: It's definitely a Short. All right.

Zubin: Let's look into
the cauldron, right?

The question people actually search for:
what happens if the AI bubble bursts?

Aman: So let's walk it
day by day, and split it.

I'll take the first four days,
and you can take the rest.

To be clear, this is hypothetical,
but every link exists today. Day one.

A frontier lab's next funding
round comes in below the last one.

Its cloud-giant backers
mark their stakes down,

and the $99 billion of gains starts
running backwards through their earnings.

Day two. The lab asks to stretch
its compute commitments.

Most can't be cancelled.

So this stops being a sales conversation
and becomes a credit conversation.

Analysts start subtracting
from that $1.7 trillion backlog.

Day three comes along,
and so do the neoclouds.

CoreWeave's interest bill is already close
to half its earnings before depreciation,

and in July, its credit default swaps
briefly priced roughly even odds

of default over five years. Lenders reread
their covenants. Day four. Collateral.

Chip-backed loans and data-centre bonds
get re-marked. Remember that A100.

Zubin: Yeah, let me take the next few.
So day five. Private credit.

Funds with AI-heavy books mark down,
investors ask for their money back,

and semi-liquid funds hit their gates.

We're already seeing this dress rehearsal:

14% of redemption requests against
a 7% cap. Day six. The banks.

Borrowers draw their committed lines,

and that gap between 9% and 25% of core
capital that we talked about earlier stops

being theoretical. And day seven. Retail.

Leveraged chip ETFs, I guess,
what, about $43 billion of them,

turn a 10% fall into 30%. Margin calls
land on $1.45 trillion of margin debt.

Graham opens his quarterly statement.

Priya's SIP buys the dip,
whether she likes it or not.

Aman: That is one grim
week we just painted there.

Zubin: Now Aman, let me
be the adult in the room.

Of course, there's no other adult here.
Banks have far more capital than in 2008.

The cloud giants have more than a
trillion dollars of borrowing room.

And the Bank of England itself
said correction, not a crisis.

Aman: All true. And the likelier path
isn't one bad week. It's a plateau.

Consensus has capex flat at $1.1
to $1.2 trillion from 2027.

Chip growth stalls, cash flow stays
negative, refinancing gets dearer.

Zubin: Not a crash.
A long, expensive hangover.

Aman: So, the search question: what
happens to your pension and your life

insurance if the AI bubble bursts?

Zubin: Mostly a markdown, not a full wipe.

Aman: And the danger isn't
the size of any one loss.

It's that they all
arrive in the same week.

Zubin: Yeah, definitely a grim week,
but a thoughtful, balanced view of this,

I must say. All right, Aman,
let's do some predictions here.

Two each, specific, dated,
and we'll check them on the 6th

of October 2027. You first.

Aman: Pressure. All right.

One: hyperscaler capex for 2027 comes
in below today's consensus of about $1.1

trillion. Analysts have under-forecast
four years running.

I think the fifth year
breaks that pattern.

Zubin: Bold. Mine's going to be simple:
Anthropic lists before OpenAI.

Aman: I think you'll
find that out very soon.

Two: by the end of 2027, one
global bank publishes a single,

all-in number for its AI exposure across
loans, bonds, leases and investments.

Zubin: Interesting. And my second:
I'm going to say by the end of 2027,

a major exchange lists a futures contract
on the price of renting AI compute.

And if it trades like a commodity,
someone will hedge it like one.

Aman: I think that's the boldest one so
far. So, written down. Now, the open loop.

Zubin: All right, Aman,
you've been itching for this.

I feel like there's a mic drop.

Aman: Yeah, let's go back to slide 88.

Pokémon cards have beaten
the Nasdaq since 2024,

and nobody calls it a bubble
when the prices keep rising,

because a collectible is worth whatever
the next person is willing to pay.

The AI boom is real. The revenue is real.

But the question was never
whether the cards are valuable.

It's who's holding them
when the next buyer pauses.

Two years ago, that was the shareholder.

Today, it's Linda's electricity bill,
Graham's pension and Priya's monthly SIP.

a16z has written a brilliant deck
about who's winning the AI boom.

Our job is to read
the slide they didn't write,

and to keep asking the stupid question.

In a London meeting room
in 2000, it was "who loses?"

Today, it's "who's holding it?"

Zubin: Boom. If this was useful,
subscribe, vote in the polls,

and send it to that one colleague
who still says it's all a bubble.

And by the way, if you are at
Money20/20, come and find me.

I might be that one guy
in the Metallica T-shirt.

Aman: Stay curious.

Zubin: And stay purposeful,
my friends, and of course, stay liquid.