Subscribe now

The Long Game ♟️ Perspective: An Alpha series:

On 15 July 2026, Reuters reported something rather extraordinary.

Stripe and private-equity firm Advent International had offered $60.50 a share to acquire PayPal, valuing the company at more than $53 billion. The proposal represented roughly a 28% premium to where PayPal had been trading. PayPal’s board reportedly considered the offer inadequate.

Then, on 28 August, the story changed again.

Reuters reported that Stripe and Advent had abandoned their pursuit.

PayPal had closed the previous regular trading session at $61.47. After news that the buyers had walked away, overnight trading briefly indicated roughly $53.20, although investors should treat overnight prices cautiously because liquidity is much thinner.

And strangely, I think the failed takeover has made PayPal more interesting to analyse.

Because we can remove the takeover speculation and ask the question that actually matters:

What is PayPal worth if nobody comes to rescue it?

That is where this gets interesting.

The Investment Thesis in One Paragraph

PayPal is no longer the hyper-growth fintech story investors thought they owned in 2021.

But neither is it a dying payments company.

It still has 439 million active accounts, operates across roughly 200 markets, processed $1.79 trillion of payments in 2025, generates billions in free cash flow, owns Venmo, operates Braintree, and possesses one of the largest two-sided payments networks outside the card giants.

The problem is that PayPal’s fastest-growing transaction volumes are increasingly coming from businesses with weaker economics, while its historically attractive branded online checkout business is growing at only about 2%.

That makes PayPal a very different investment from Visa or Mastercard.

The thesis is not:

PayPal will dominate digital payments again.

It is:

PayPal remains valuable enough that modest operating growth, improving monetisation and aggressive share repurchases could create surprisingly respectable per-share returns from a very undemanding valuation.

The catch is equally simple.

If branded checkout is undergoing permanent structural erosion rather than temporary stagnation, PayPal could remain cheap for a very long time.

First, Forget the PayPal You Remember

At its 2021 peak, PayPal was worth around $360 billion.

Investors valued it as though the company would remain one of the dominant gateways to digital commerce indefinitely.

That assumption turned out to be rather heroic.

Apple Pay grew.

Google Pay grew.

Shop Pay improved.

Stripe became formidable.

Merchants gained more choices.

Smartphones increasingly made the operating system, browser or merchant itself the natural owner of the checkout experience.

And PayPal went from being the convenient alternative to entering a world where convenience became commoditised.

Its market capitalisation eventually collapsed by hundreds of billions of dollars.

But falling 80% does not automatically make a company cheap.

Sometimes the market has merely corrected an earlier hallucination.

So we need to look at what remains.

What PayPal Actually Owns Today

One of new CEO Enrique Lores’s first major moves was to reorganise PayPal around three businesses.

1. Checkout Solutions & PayPal

This combines PayPal’s consumer and merchant ecosystem and includes its branded checkout proposition.

This remains strategically important because branded checkout historically generates much better economics than simple payment processing.

2. Consumer Financial Services & Venmo

PayPal wants to turn Venmo from primarily a peer-to-peer payments app into something closer to a consumer financial platform where users can send, spend, borrow and eventually manage more of their financial lives.

3. Payment Services & Crypto

This contains Braintree, SMB processing, payments infrastructure, value-added merchant services and crypto assets including PYUSD.

The reorganisation matters because management is finally admitting something the market already knew:

PayPal cannot depend on one checkout button forever.

Its future has to come from monetising the broader network.

The Numbers Are Better Than the Share Price Suggests

Start with 2025.

PayPal generated:

Revenue has also continued growing from $29.77 billion in 2023, to $31.80 billion in 2024, to $33.17 billion last year.

This is hardly a company disappearing over the horizon.

Then came Q2 2026.

PayPal processed:

$486.4 billion of TPV, up 10% year on year.

Revenue rose 5% to $8.68 billion.

Transactions increased 8%.

Transactions per active account excluding payment-service-provider activity increased 7%.

Yet active accounts were virtually unchanged at 439 million.

So the broad picture is quite clear.

PayPal does not have a usage problem.

It has a monetisation and mix problem.

And that distinction matters enormously.

The Most Important Number in the Entire PayPal Thesis

Here is the statistic I would watch before almost everything else.

In Q2:

TPV grew 9% on a currency-neutral basis.

But:

Revenue grew only 3% currency-neutral.

And:

Transaction-margin dollars grew only 1%.

Transaction-margin dollars excluding interest earned on customer balances grew 3%.

PayPal is moving substantially more money.

But each additional dollar moving through the ecosystem is not creating an equivalent increase in economic value.

You can see it in the take rate.

PayPal’s transaction take rate has fallen from around 1.68% in Q2 2025 to 1.61% in Q2 2026.

Its total take rate declined from around 1.87% to 1.78%.

That is not some accounting footnote.

That is the bear case.

The market is effectively asking whether PayPal is becoming a huge but progressively lower-yielding payments utility.

Why Is This Happening?

Part of the answer is Braintree.

PayPal’s payment-service-provider business represented roughly 45% of total TPV in Q2, and PSP volume grew 13% currency-neutral.

Braintree itself was growing in the mid-teens.

That sounds wonderful until you remember that all payment volume is not created equal.

PayPal’s own filing says Braintree transactions generally have higher expense rates than its other products and services.

In other words:

Braintree gives PayPal scale, but branded checkout gives PayPal economics.

That is why I would be careful celebrating TPV alone.

If tomorrow PayPal processes $3 trillion instead of $2 trillion but earns substantially less from every dollar processed, shareholders may not be much richer.

Volume is vanity.

Transaction-margin dollars are sanity.

Free cash flow is reality.

Which Brings Us to Branded Checkout

PayPal’s online branded checkout represented roughly 28% of TPV in Q2.

Currency-neutral volume growth was just 2%.

The good news is that this has stabilised. Growth was 1% in Q4 2025, then 2% in Q1 and 2% again in Q2.

Management has consequently raised its full-year expectation to low-single-digit growth.

That might sound uninspiring.

It is.

But for this investment thesis, uninspiring may be enough.

PayPal does not need branded checkout to suddenly grow 15%.

If it can keep that franchise growing at 2% to 4%, defend its economics and use the rest of the ecosystem to create incremental growth, the valuation starts doing a lot of the heavy lifting.

Conversely, if that 2% becomes zero, then minus 2%, then minus 5%, the thesis changes dramatically.

That is why I regard branded checkout growth as PayPal’s single most important operating health indicator.

Venmo Is Probably PayPal’s Most Interesting Asset

Now we reach the part of PayPal that looks genuinely healthy.

Venmo’s Q2 TPV grew 14%, marking its seventh consecutive quarter of double-digit growth.

Pay with Venmo volume jumped 44%.

Venmo Debit Card monthly active accounts grew more than 50%.

Debit card and tap-to-pay spending across PayPal grew more than 60%, although these businesses are still relatively small.

But the most interesting disclosure was not a growth rate.

It was monetisation.

Customers using both Venmo Debit and Pay with Venmo generated more than nine times the average revenue per account of users who only use Venmo for peer-to-peer payments.

Even more interestingly, that higher-value customer cohort has roughly doubled over the past year.

That begins to change the Venmo story.

Historically, one criticism was obvious:

Lovely app. Lots of engagement. Where’s the money?

Now PayPal has evidence that the money appears when users migrate from simply sending $20 to a friend for dinner into using Venmo as part of everyday commerce.

This is essentially a funnel:

P2P user → stored balance → debit card → merchant checkout → broader financial relationship.

If PayPal can systematically push more of Venmo’s user base down that funnel, Venmo’s economics could improve considerably without requiring heroic user growth.

That is the kind of growth I prefer.

Not more eyeballs.

More value per existing customer.

The Moat: Narrower, But Not Gone

I would not call PayPal a wide-moat company today.

But calling it moatless would also be lazy.

PayPal still has several genuine competitive advantages.

It operates a global two-sided network connecting hundreds of millions of accounts.

It sees data from both sides of many transactions.

It has extensive risk, fraud and compliance infrastructure.

It offers buyer protection, seller protection and dispute resolution.

Consumers can transact without handing financial information directly to each merchant.

It has recognised brands.

And it has regulatory licences across a huge number of jurisdictions.

These things are expensive and difficult to recreate.

That is presumably part of what Stripe saw when it reportedly considered paying more than $53 billion for the company.

But the weakness in PayPal’s moat is equally important.

PayPal increasingly does not control the moment at which the consumer decides how to pay.

Apple controls the iPhone.

Google controls Android.

Shopify controls enormous numbers of merchant storefronts.

Stripe increasingly controls merchant payments infrastructure.

Banks control the underlying funding relationship.

Card networks still sit beneath vast amounts of global commerce.

PayPal sits in between all of them.

Twenty years ago that position was revolutionary.

Today it can occasionally look crowded.

My classification would therefore be:

PayPal’s moat is narrowing, not disappearing.

That is good enough at the right price.

It is not good enough at 40 times earnings.

Management: Turnaround Number Two

This is another reason I would resist making PayPal a high-conviction investment too quickly.

Alex Chriss became CEO in 2023 and was supposed to lead PayPal’s turnaround.

Then in February 2026, the board replaced him.

The board was unusually explicit, saying that while progress had been made, the pace of change and execution had not been in line with its expectations.

Enrique Lores took over on 1 March.

Lores previously ran HP and had already served on PayPal’s board for nearly five years.

His strategy so far makes conceptual sense:

• simplify PayPal into three accountable businesses;

• invest more heavily in consumer relationships;

• accelerate Venmo monetisation;

• improve Braintree economics using value-added services;

• modernise the technology platform;

• automate aggressively using AI;

• cut unnecessary cost;

• reinvest into businesses capable of genuine growth.

PayPal expects its simplification programme to produce at least $1.5 billion of gross annualised run-rate savings over the next two to three years, with around $400 million of run-rate savings actions identified for 2026.

But there is an important catch.

Management intends to reinvest a significant proportion of those savings.

Therefore I would not make the common spreadsheet mistake of simply adding $1.5 billion to future operating profit.

The savings buy PayPal optionality.

They do not automatically belong to shareholders.

2026 Is Still an Investment Year

You can see that reinvestment in Q2.

Non-GAAP operating income fell 8%.

Non-GAAP operating margin dropped from 19.8% to 17.4%.

Technology and development spending increased.

PayPal now expects non-transaction operating expenses to grow roughly 7% to 8% in 2026.

Normally, declining margins plus low-single-digit revenue growth is hardly the recipe for excitement.

But there is a difference between:

margin deterioration because your business has weakened

and

margin deterioration because you deliberately increased investment.

PayPal currently contains elements of both.

That is precisely why this remains a turnaround rather than a clean compounder.

The Best Part of the PayPal Story May Be Capital Allocation

This is where the mathematics becomes quite powerful.

PayPal expects to repurchase roughly $6 billion of shares during 2026.

During the first six months alone, it repurchased approximately 67 million shares for $3.0 billion.

Average purchase price?

$44.99.

That looks like excellent capital allocation if the underlying business remains worth materially more than $45.

Actual shares outstanding fell from:

920 million at 31 December 2025

to

862 million at 30 June 2026.

That is roughly a 6.3% reduction in just six months.

Q2 diluted weighted-average shares were 882 million compared with 977 million a year earlier.

So unlike companies that announce enormous buybacks merely to disguise stock-based compensation, PayPal is producing meaningful net share-count reduction.

This matters tremendously.

Imagine a business whose aggregate profit barely grows.

If the company retires 5% to 7% of its equity every year at sensible prices, earnings per share can grow considerably faster than company-wide earnings.

At today’s valuation, buybacks are not merely returning capital.

They are potentially an important part of the investment thesis.

Is the Balance Sheet Safe?

At 30 June, PayPal held approximately:

$15.3 billion in cash, cash equivalents and investments

against

$13.4 billion of debt.

I do not see leverage as the central risk here.

PayPal’s funding, customer balances and credit activities make the balance sheet more complicated than a normal software company, but this is not a distressed financial structure.

Credit deserves monitoring, however, particularly because Lores wants PayPal to expand further into financial services.

PayPal’s merchant lending portfolio increased to $1.8 billion at the end of 2025.

Loans more than 90 days past due increased from 2.8% to 3.5%, while the annualised merchant net charge-off rate increased from 5.3% to 7.3%. Its allowance for merchant credit losses rose from $113 million to $170 million.

By Q2 2026, total quarterly credit losses were actually lower year on year at $74 million versus $93 million, although net charge-offs themselves had increased.

Nothing here presently frightens me.

But if PayPal wants to become more bank-like, investors must start watching more bank-like risks.

There is no free lunch, even in fintech.

Now We Come to Valuation

PayPal’s 2026 guidance currently calls for approximately:

• $15.6bn transaction-margin dollars

• $14.5bn TM$ excluding interest on customer balances

• $5.38 non-GAAP EPS

• $6bn+ adjusted free cash flow

• roughly $6bn of share repurchases

• around $1bn of capex

At the 27 August regular-session close of $61.47, PayPal trades at approximately:

11.4x 2026 non-GAAP EPS

Using the overnight indication of roughly $53.20 after the takeover withdrawal:

9.9x 2026 non-GAAP EPS

PayPal had 862 million shares outstanding at the end of June, so $53.20 would imply an equity value of roughly $46 billion.

On $6 billion of guided adjusted free cash flow, that is an adjusted free-cash-flow yield around 13%.

At $61.47, the comparable yield is still roughly 11%.

One caveat matters.

PayPal’s adjusted free-cash-flow measure removes timing differences between originating BNPL receivables and selling those receivables. Actual 2025 free cash flow was $5.56 billion versus adjusted FCF of $6.41 billion.

So I would not blindly capitalise adjusted FCF as though every dollar were conventional owner earnings.

Even after making that adjustment, however, PayPal does not look expensive.

The debate is whether it deserves to be.

A Simple Valuation Map

I do not think PayPal warrants a beautifully complicated 17-tab DCF model.

The uncertainty is not whether WACC should be 8.7% or 9.1%.

The uncertainty is whether the franchise is stabilising.

So I prefer scenario analysis.

These are deliberately broad scenarios, not price targets.

Bear case

Branded checkout returns to stagnation or decline.

Braintree keeps producing volume without attractive economics.

Margin savings disappear into reinvestment.

Venmo grows but monetises more slowly than hoped.

The company essentially uses buybacks to manufacture modest EPS growth.

In that world, PayPal could remain a $45-$60 stock for years.

Base case

Branded checkout holds low-single-digit growth.

Venmo monetisation continues improving.

Braintree grows double digits and gradually attaches higher-margin services.

The cost programme eventually offsets some investment spending.

Revenue grows modestly but share count keeps falling.

PayPal reaches $7-plus EPS without requiring a heroic turnaround.

Put a fairly ordinary 12x multiple on that, and the mathematics becomes attractive.

Bull case

Branded checkout actually reaccelerates.

Venmo becomes a meaningful financial-services platform.

Braintree successfully moves beyond commodity processing.

Cost simplification begins dropping into margins.

Per-share earnings reach the high single digits.

Then PayPal no longer deserves a single-digit or low-double-digit multiple.

I would regard this as upside optionality, not the base thesis.

The Bull Case in One Sentence

PayPal does not need much growth because the existing cash machine is buying back enormous quantities of undervalued stock.

That is the central value-investing argument.

If the company earns roughly $5 to $6 billion a year and can keep retiring equity cheaply while its underlying franchise merely remains stable, shareholders could do quite well.

You do not need PayPal to become Nvidia.

You need it to avoid becoming Yahoo.

There is a considerable difference.

The Bear Case in One Sentence

The thing PayPal is growing fastest is not necessarily the thing investors should want it to grow fastest.

PSP represents around 45% of TPV and is expanding quickly.

Branded online checkout represents roughly 28% and is growing around 2%.

If the lower-economic-value business steadily replaces the higher-economic-value business, the company’s impressive TPV headline could disguise deterioration underneath.

That is why I would never buy PayPal merely because it appears cheap on P/E.

A declining business can remain on 10x earnings forever.

Sometimes it even reaches 8x.

Then 6x.

Value investors have a technical term for this.

Pain.

What Would Make Me More Bullish?

I would watch seven things over the next four to six quarters.

1. Online branded checkout growth

I want that 2% number to hold and eventually move towards 3%, 4% or better.

If it turns negative again, I would reassess the thesis immediately.

2. Transaction-margin dollars excluding interest

This needs to accelerate from the present roughly 3% growth towards sustainable mid-single-digit growth.

3. Venmo monetisation

Forget raw Venmo user numbers.

Watch Debit Card adoption, Pay with Venmo, financial-services penetration and ARPA.

4. Braintree economics

I care much less whether Braintree grows TPV by 15% than whether PayPal can attach fraud prevention, optimisation, payouts and other higher-value services.

5. Share count

If PayPal spends another $6 billion buying stock but share count stops falling materially, something has changed.

6. Operating margin

The $1.5 billion savings programme eventually needs to produce some shareholder economics rather than disappearing entirely into perpetual reinvestment.

7. Credit quality

As PayPal expands consumer and merchant finance, delinquencies and charge-offs deserve increasingly serious attention.

Where I May Be Wrong

There are two mistakes value investors commonly make with companies like PayPal.

The first is anchoring.

The stock once traded above $300, therefore $50 must be cheap.

Nonsense.

The old valuation has precisely zero bearing on intrinsic value today.

The second is assuming cash flow creates a floor.

It does not.

A business generating enormous cash today can still destroy value if its competitive position is steadily deteriorating.

BlackBerry generated cash.

Nokia generated cash.

Kodak once generated quite a lot of cash too.

The real question is what happens to that cash-generation ability five and ten years from now.

That is the part of PayPal I am not yet prepared to declare solved.

My Verdict

Classification: Attractive Turnaround, Not Yet a High-Conviction Compounder

PayPal has become much more interesting precisely because nobody loves it.

At somewhere around 10x 2026 earnings based on the post-bid overnight indication, expectations are low.

The business still processes nearly half a trillion dollars every quarter.

Venmo is growing nicely.

Braintree has scale.

The company generates substantial cash.

The balance sheet is sound.

Share repurchases are unusually aggressive and, so far, have been executed at excellent prices.

And the new CEO has identified the right strategic problems.

Yet none of that answers the most important question:

Has PayPal stabilised its competitive position, or merely slowed the rate at which it is losing relevance?

I don’t think we know yet.

That stops me from calling this an obvious bargain.

But around the low-$50s, I think the risk/reward becomes genuinely interesting.

At roughly $45, around the price at which management itself bought enormous quantities of stock during the first half of this year, the margin of safety becomes considerably more compelling assuming the operating thesis has not deteriorated.

At $80 or $90, I would demand much more proof.

And perhaps that is the most important lesson here.

PayPal does not have to reclaim its old throne.

It does not need another pandemic.

It does not need investors to rediscover fintech.

It does not even need particularly spectacular revenue growth.

It simply needs to remain relevant enough, grow profitable parts of the ecosystem fast enough, and keep using its considerable cash flow intelligently.

At 30 or 40 times earnings, that would be a weak thesis.

At roughly 10 times?

Boring survival can be surprisingly profitable.

And that may be the real PayPal investment case.

If you found value in this article, please hold down the like ❤️ button or consider buying me a coffee ☕️ to fuel the next one!

Buy me coffee ☕️

Subscribe now

The Long Game ♟️ Perspective: A Wealth Mindset series:

He had been investing regularly in the same fund for several years.

The fund’s performance page showed something like:

3-year return: +48%

His own account showed:

Return: +22%

His question was simple:

“So where did the other 26% go?”

It sounds like something must be wrong.

Usually, nothing is.

The problem is that the fund’s return and your return are not necessarily measuring the same journey.

And once you understand that, the way you look at investment performance changes quite a bit.

Imagine getting on a train halfway 🚆

Suppose a train leaves London at 9am and reaches Edinburgh three hours later.

One passenger gets on at the first station.

Another boards halfway through.

A third joins with just 30 minutes left.

All three reach Edinburgh.

But they haven’t travelled the same distance.

Regular investing works in much the same way.

When a fund says:

“We returned 48% over three years”

it is effectively telling you what happened to money that was invested at the beginning of that three-year period and remained invested throughout.

But most ordinary investors don’t invest like that.

My client had been investing every month.

Perhaps:

January: S$1,000
February: S$1,000
March: S$1,000
April: S$1,000

and so on.

The first S$1,000 had years to grow.

The money invested one year later had less time.

The contribution made last month barely had any time at all.

So although the fund has been travelling for three years, not every dollar in his portfolio has.

That is where the confusion begins.

This is where XIRR comes in

There is a financial term called XIRR, which stands for Extended Internal Rate of Return.

The name sounds far more intimidating than the idea.

In plain English, XIRR asks:

“Taking into account when I actually invested my money, what annual return did my money earn?”

That’s it.

Go back to the train.

XIRR doesn’t pretend every passenger boarded at the first station.

It looks at when each one actually got on.

If you invested:

• S$10,000 three years ago

• S$10,000 last year

• S$10,000 last month

it would make little sense to pretend the entire S$30,000 had been invested for three years.

XIRR accounts for the timing.

And this is why it can be much more useful for anyone who invests regularly through monthly contributions, dollar-cost averaging, bonuses or occasional top-ups.

The simplest way I can put it is this:

The fund’s return tells you the history of the fund. XIRR tells you the history of your money.

Those are not always the same story.

Three numbers investors often mix up

This is where investment statements can become confusing.

We often see several numbers and assume they all answer the same question.

They don’t.

Fund return tells you how the fund itself performed over a particular period.

Portfolio P&L tells you how many dollars you have gained or lost.

XIRR tells you the annualised return earned by your actual money, based on when it went in and out.

All three are useful.

But they answer different questions.

Imagine two investors both show a return of 20%.

One achieved it in one year.

The other took five years.

Same headline number.

Very different investment outcome.

Time matters.

Does this mean DCA is worse?

This is the point where some investors may reach the wrong conclusion.

They see:

Fund: +48%

My account: +22%

and think:

“I should have invested everything on day one.”

Not necessarily.

That isn’t what XIRR is telling us.

Most people don’t wake up one morning with their entire lifetime investment capital sitting in the bank.

They earn.

They save.

They receive bonuses.

They gradually build wealth.

For many people, investing regularly is simply the practical way investing happens.

Dollar-cost averaging also removes some of the pressure of trying to guess whether today’s market price is the perfect entry point.

The lesson isn’t that DCA is bad.

The lesson is much simpler:

Don’t compare a staggered investment journey with a lump-sum return and expect the numbers to match.

So what number should you actually care about?

If I were reviewing my own portfolio, I wouldn’t stop at:

“My fund made 48%.”

I would ask:

What return did my money actually earn?

That’s where XIRR becomes useful.

But even XIRR isn’t the final answer.

Suppose your XIRR is 8% per year.

Is that good?

Maybe.

You still need to ask:

What risk did I take?

How volatile was the investment?

What did a sensible alternative return?

What fees did I pay?

Did I receive income distributions along the way?

Most importantly:

Did the investment do the job I bought it to do?

A retirement-income portfolio and a high-growth equity portfolio shouldn’t be judged by exactly the same yardstick.

The highest return is not automatically the best investment.

Sometimes a slightly lower return with considerably less risk is precisely what you wanted.

There is a bigger lesson here

What struck me after this conversation was that the same principle applies outside investing.

We tend to notice big financial decisions.

Buying a property.

Changing jobs.

Investing S$100,000.

Those feel important.

But much of our financial life is actually shaped by decisions that barely register when we make them.

S$20 spent here.

A subscription renewed there.

S$500 invested this month.

An extra S$300 saved next month.

None of them feels particularly meaningful on a random Tuesday afternoon.

Repeat them for ten years and they become your financial life.

This is why I don’t particularly like the usual advice telling people that one cup of coffee is destroying their future.

Life is meant to be enjoyed.

But the broader idea is still important.

A single decision is often insignificant.

A repeated decision becomes a pattern.

And patterns are where money starts becoming powerful.

S$500 invested once probably won’t change your life.

S$500 invested month after month for decades might.

Likewise, one unnecessary purchase doesn’t matter very much.

Hundreds of unconscious purchases can quietly reshape your finances.

The numbers are small.

The habit isn’t.

Perhaps that’s the real lesson from XIRR

When we think about money, we naturally focus on how much.

How much did I invest?

How much did I make?

How much did the fund return?

But there is another dimension we often overlook:

When did the money move?

And:

How often did that behaviour repeat?

Money is shaped by amount, timing and repetition.

Once you start seeing it that way, an investment statement becomes more than a percentage on a screen.

It becomes a record of your financial behaviour.

So the next time someone proudly tells you:

“This fund returned 50%.”

there is a useful question to ask:

“That’s the fund’s return. What was mine?”

If you found value in this article, please hold down the like ❤️ button or consider buying me a coffee ☕️ to fuel the next one!

Buy me coffee ☕️

Subscribe now

The Long Game ♟️ Perspective: A Macro & Global trend series:

Singapore just reported 5.9% GDP growth for the second quarter.

That’s a fantastic number.

The sort of number that should make you think businesses are busy, hiring is strong, consumers are spending and everyone is generally feeling a little richer.

Except...

Does it actually feel like a 5.9% economy to you?

Because when I look at what is happening closer to the ground, I struggle to reconcile the two.

Businesses are still talking about costs.

Restaurants are still disappearing.

Consumers still seem extremely price-conscious.

Companies continue to restructure.

And SMEs hardly sound like they’re partying.

This isn’t merely anecdotal.

The Singapore Business Federation’s latest business sentiment survey found that only 34% of businesses were satisfied with the current business climate.

More strikingly, 41% expected economic conditions to worsen over the next 12 months.

Among SMEs, it was 43%. (SBF)

So we have this rather strange situation:

GDP says boom.

The ground says... where?

That contradiction is what made me dig into MTI’s latest Economic Survey of Singapore, Second Quarter 2026.

And once you look underneath the headline, the answer becomes surprisingly clear.

The 5.9% is real.

But it isn’t remotely evenly distributed.

Look at those numbers.

GDP: +5.9%

Merchandise exports: +38.5%

Services exports: +10.2%

Employment: +14,000

You would be forgiven for looking at that dashboard and thinking Singapore is absolutely flying.

But GDP tells us how much value the economy produced.

It doesn’t tell us who felt it.

And that’s where this gets interesting.

Here is the number that changed how I read the entire report

Forget 5.9% for a moment.

Look at where it actually came from.

Manufacturing contributed 2.1 percentage points.

Wholesale trade contributed 1.5 points.

Finance & insurance added another 0.8 points.

That’s 4.4 percentage points out of Singapore’s entire 5.9% growth.

Roughly three quarters of the growth came from three sectors.

Now look at the sectors much closer to everyday life.

Retail?

Almost no contribution.

Accommodation?

Almost none.

F&B?

Essentially none.

And suddenly the 5.9% economy starts looking rather different.

This wasn’t everyone running faster together.

It was more like three runners sprinting so quickly that they pulled up the average for the entire team.

That’s still growth.

It’s still good news.

But it is not the same thing as broad-based prosperity.

Singapore Inc. is doing very well. Main Street Singapore? Much less obvious.

Take wholesale trade.

It grew 8.3%.

Sounds fantastic.

But what was actually booming?

Non-oil re-export volumes jumped 34.2%.

Non-oil domestic exports rose 21.3%.

MTI says machinery and equipment were major drivers.

Now put wholesale beside retail.

Wholesale trade:

+8.3%

Retail trade:

+1.0%

Non-oil re-exports:

+34.2%

Non-motor retail sales:

+1.5%

Motor-related retail sales:

+0.3%.

That’s quite a split.

And it matters because Singapore’s wholesale economy isn’t simply your neighbourhood distributor selling more cartons of drinks.

Singapore is a global trading hub.

Machinery moves.

Electronics move.

Regional transactions are booked here.

Goods are imported, exported and re-exported.

Financing, logistics and trading margins are captured here.

All of that creates genuine economic value.

But a container of semiconductor equipment being re-exported through Singapore doesn’t necessarily put another customer inside the café downstairs.

That’s not criticism.

That’s simply how our economy is structured.

And it explains why Singapore can look spectacular from 30,000 feet while feeling considerably more ordinary at street level.

Want something closer to the ground? Look at F&B.

This is where the GDP headline starts losing some of its shine.

F&B services contracted 1.5% in Q2.

Restaurants fell 0.8%.

Cafés fell 1.9%.

Other F&B establishments fell 6.7%.

Accommodation grew only 2.2%, down sharply from 6.9% in the previous quarter.

That feels rather more recognisable, doesn’t it?

And the pressure on businesses isn’t something we have to infer from restaurant chatter.

SBF’s latest survey found the overall Business Sentiment Index weakened in Q1, while the proportion of businesses expecting tougher economic conditions rose from 31% to 41% in just one quarter. SMEs accounted for much of that deterioration. Cost expectations also jumped sharply. (SBF)

Earlier SBF research found businesses consistently highlighting manpower costs, uncertain customer demand and rental costs as major challenges, with profitability under pressure. (SBF)

And yes, Singapore’s F&B closure story has been very visible.

But I wouldn’t exaggerate it either.

The Government pointed out in February that, despite high churn, the total number of F&B entities actually increased 42% between 2015 and 2025, with formations exceeding cessations in most years. (Ministry of Trade and Industry)

So this isn’t some grand collapse of Singapore F&B.

It is arguably something more interesting:

An intensely competitive market where businesses keep entering, businesses keep dying, margins stay thin, and consumers remain difficult to win.

Not exactly what most people imagine when they hear “5.9% economic growth”.

Then I found what may be the most important number in the whole report

Not GDP.

Productivity.

Wholesale trade productivity jumped 15.4%.

Information & communications: +9.4%

Real estate: +9.1%

Manufacturing: +7.6%

Finance & insurance: +5.1%

Meanwhile:

F&B: -0.5%

Construction: -0.1%

Other services: -1.6%.

But the killer statistic comes immediately afterwards.

MTI groups the economy into outward-oriented and domestically-oriented sectors.

Productivity in the outward-oriented economy rose 6.9%.

Productivity in the domestically-oriented economy fell 0.1%.

Read that again.

+6.9% versus -0.1%.

That, to me, is the real Q2 economic story.

Singapore’s globally connected economic machine is doing extraordinarily well.

Semiconductors.

AI.

Manufacturing.

Trading.

Finance.

Technology.

But the part of the economy that is much closer to everyday domestic demand is nowhere near as strong.

So perhaps we don’t really have one Singapore economy at the moment.

We have two Singapores moving at very different speeds.

And AI may actually make this gap more visible

Here is where things become even more fascinating.

Manufacturing grew 12.5%.

Electronics grew 33.8%.

Semiconductor production rose 35.7%.

Infocomms and consumer electronics surged 51.8%.

MTI directly linked much of that strength to AI-related demand.

This is excellent for Singapore.

We absolutely want to be plugged into one of the biggest global capital expenditure cycles in decades.

But there is another question nobody should ignore.

What if the industries driving our next economic boom simply don’t need that many additional people?

Wholesale trade offers an interesting clue.

Its productivity surged 15.4%.

Yet wholesale employment fell by roughly 1,900 workers quarter-on-quarter.

Those figures use different comparison periods, so we shouldn’t pretend one directly caused the other.

But the broader direction is worth watching.

If AI allows businesses to produce substantially more output with relatively fewer additional workers, then we could increasingly see something that sounds paradoxical:

Excellent GDP growth.

Excellent productivity.

Excellent corporate profits.

But only modest employment growth in some of the industries creating that output.

And if that happens, the gap between economic growth and felt prosperity becomes even more important.

The labour market tells the same mixed story

No, Singapore’s labour market isn’t weak.

Unemployment was only 2.0%.

Resident unemployment stood at 2.9%.

Employment increased by 14,000.

Those are objectively healthy numbers.

But again...

Look underneath.

Retrenchments increased from 3,830 to 4,500 in one quarter.

Most were attributed to restructuring and business reorganisation.

And while construction added around 7,100 workers and manufacturing added 2,900, several sectors lost workers during the quarter, including wholesale trade, administrative & support services, F&B, retail and information & communications.

This matters because nobody wakes up in the morning thinking:

“Wonderful. National unemployment remains 2%.”

People think:

“Is my company hiring?”

“Is my boss cutting headcount?”

“Can I find another job if this one disappears?”

“Am I getting a decent increment this year?”

That is how an economy is experienced.

Not through GDP.

Through job security, wages, margins and purchasing power.

And households? They’re not imagining the squeeze either.

Here’s another strange disconnect.

Inflation has moderated significantly.

Q2 CPI was only 1.8%.

But transport prices were up 7.3%.

Healthcare rose 3.1%.

Food increased 1.9%.

And remember:

Lower inflation does not mean things became cheaper.

It means they’re getting expensive more slowly.

That’s quite different.

A June 2026 survey of Singapore households found 81% said inflation had made it harder to meet monthly expenses.

Only 11% said they felt very financially secure.

Nearly one in four said they were drawing down savings, and roughly the same proportion reported cutting or skipping essential spending to deal with near-term budget pressures. (Sun Life)

Another April survey found 39% of Singaporeans expected Singapore to enter recession within six months, versus just 15% who expected growth. (YouGov)

Obviously they were wrong about what GDP was doing.

But that’s exactly the point.

Their perception of their own economic reality was dramatically weaker than the macroeconomic reality.

To be fair, consumer sentiment isn’t universally gloomy. UOB’s June survey found Singapore recorded one of the strongest improvements in consumer confidence in ASEAN, although concerns around household expenses, income security and long-term financial commitments persisted. (United Overseas Bank)

So I wouldn’t call Singapore consumers depressed.

I would call them:

cautious.

And a cautious consumer behaves very differently from someone living through a perceived boom.

Maybe we’re asking GDP to answer a question it was never designed to answer

This is where I eventually landed.

There is nothing inherently suspicious about Singapore’s 5.9% growth number.

The economy really did grow.

Exports really did surge.

Manufacturing really is benefiting from the AI investment cycle.

Wholesale activity genuinely increased.

Finance genuinely expanded.

The statistics are doing their job.

Perhaps we’re the ones asking the wrong question.

GDP answers:

How much more economic value did Singapore produce?

But when someone says:

“The economy doesn’t feel good.”

They’re asking:

Did my business make more money?

Did my salary grow?

Do I feel secure about my job?

Are customers spending?

Did my purchasing power improve?

Those are completely different questions.

And this quarter, the answers clearly aren’t moving together.

The real Singapore economic story isn’t 5.9%.

It’s this:

Singapore Inc. is booming.

The Singaporean on the ground is still waiting to feel more of it.

That doesn’t make the GDP number wrong.

It makes the economy uneven.

And that distinction matters enormously.

Because if Singapore’s next growth cycle becomes increasingly powered by semiconductors, AI, capital investment, global trading flows and productivity gains, GDP may continue producing impressive numbers.

But the question I would watch much more closely is:

How much of that growth eventually travels from the semiconductor factory, trading desk and balance sheet...

into salaries, SME revenues, household purchasing power and people’s sense of financial security?

Because until that happens, don’t be surprised if Singapore posts another strong GDP number...

and someone running a business downstairs looks at you and says:

“5.9%? Where?”

Perhaps they haven’t misunderstood the economy.

Perhaps they’re simply standing in the part of it that isn’t booming.

If you found value in this article, please hold down the like ❤️ button or consider buying me a coffee ☕️ to fuel the next one!

Buy me coffee ☕️

Subscribe now

The Long Game ♟️ Perspective: An occasional commentary series:

I think Jensen Huang deserves applause for this one.

Not because NVIDIA has somehow received a $500 billion cheque. It has not. The announcement is a set of memorandums of understanding, aka MoUs; the final structures are still to be agreed, and the capital is intended to come from independent third-party investors over time.

What is strategically impressive is who he has brought to the table: Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR. I would not call them literally the six largest financiers in the world because there is no sensible universal ranking for that. But they are six of the most formidable institutions in global private capital, infrastructure, credit and asset management. Jensen has effectively invited Wall Street to help solve the next constraint in AI: not whether people want compute, but whether enough customers can finance it.

That is a very Jensen move. When an industry encounters a bottleneck, he rarely treats the bottleneck as somebody else’s problem. First NVIDIA attacked graphics. Then parallel computing. Then the software ecosystem through CUDA. Then networking, systems and AI infrastructure. Now it is attacking the cost and availability of capital.

And there is a second reason this announcement deserves attention. Jensen is trying to persuade the financial system to stop seeing GPUs as expensive electronics and start seeing NVIDIA compute as productive infrastructure.

The $500 billion headline is really a collateral story.

There is another strategic dividend: this also answers the circular-financing critique

NVIDIA has spent much of 2026 under scrutiny for an increasingly intertwined AI capital loop commonly known as circular financing. The company has invested in, committed capacity to, or discussed financial support around businesses that are also major buyers or users of NVIDIA hardware. Critics worry that when a supplier helps capitalise its customers, some of the resulting purchases can look self-reinforcing: vendor money helps fund the buyer, the buyer spends on the vendor, and that spending is then cited as evidence of demand. Jensen Huang has rejected the “circular” label, but the criticism itself is real and has become part of the market debate.

Against that backdrop, this MoU is strategically elegant. NVIDIA’s own announcement repeatedly emphasises independent compute-financing platforms, third-party capital and independent underwriting by Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR. I would not claim the MoU was created to “deflect” the criticism because NVIDIA has not said that. But whether by design or not, it functions as a powerful answer to it: the company is trying to move the financing decision away from a purely vendor-supported loop and towards external institutions that must price the risk on commercial terms.

That matters because independent underwriting is a form of outside price discovery. If sophisticated long-duration capital providers are willing to finance NVIDIA-based projects after doing their own credit work, the chain becomes less self-referential. But this does not make the circularity question disappear. NVIDIA still benefits when more financed infrastructure buys NVIDIA equipment, and Reuters reported that Huang said the company has the option to backstop up to 25% of potential deals. So the right conclusion is not “circularity solved”. It is subtler: the burden of proof is being shifted from NVIDIA’s own balance sheet towards the credit judgement of outside capital.

In other words, Jensen is not merely financing demand. He is trying to get the market to independently validate that the productive asset behind the demand is financeable.

First, a correction: Jensen did not make tokens an asset class

This distinction matters. The investable asset in NVIDIA’s own announcement is compute and the full-stack AI infrastructure around it. Tokens are the output of the factory. NVIDIA increasingly uses cost per token, tokens per second and token performance per megawatt as economic measures of that output.

Think of an AI factory as a highly automated agricultural processing plant. The building is the data centre. The GPUs, networking, storage, software, cooling and power systems are the machinery. The tokens are the tonnes of processed product coming off the production line. An investor is not being asked to buy the individual “tonnes”. The investor is financing the productive machinery and the cash flows that machinery is expected to generate.

That is why Jensen’s phrase “compute is revenue” matters. It is a commercial claim that the machine is no longer merely an IT cost centre. If a deployed unit of compute can produce billable inference, serve contracted customers and be redeployed when one customer no longer needs it, financiers can begin modelling it the way they model other productive assets.

Goldman Sachs makes the ambition explicit in NVIDIA’s announcement: it sees an opportunity to create a market for credit backed by NVIDIA compute. Apollo calls modern compute a scarce, mission-critical asset class. BlackRock talks about connecting long-term capital to essential infrastructure. Those are capital-market statements, not semiconductor marketing slogans.

So how can compute actually become investable?

The exact structures have not yet been published, so it would be premature to pretend we know the final loan-to-value ratios, covenants or securitisation mechanics. But the economic building blocks are already visible: asset-backed credit, equipment or capacity leasing, infrastructure capital, and financing tied to long-duration usage or offtake cash flows. The investor earns a credit spread, lease income or infrastructure-style return, depending on the eventual vehicle. The customer gets access to expensive compute without funding every dollar upfront.

The attraction is straightforward. Compute is scarce, demand can be contracted, and NVIDIA argues that its systems have a broad user base and a long useful economic life. The risk is equally straightforward: technology can become obsolete, utilisation can disappoint, customers can fail, residual values can fall, power can be unavailable, interest rates can rise, and a competing architecture can weaken the economics. Finance does not remove those risks. It simply prices them.

This is not simply a new data-centre REIT

The distinction is easy to miss because both live inside the same physical campus. A data-centre REIT is fundamentally a real-estate vehicle. The SEC describes a REIT as a company that owns, and typically operates, income-producing real estate or real-estate-related assets. A data-centre REIT therefore gives investors exposure primarily to the property, leases, occupancy, rents, financing and residual value of the real estate.

The NVIDIA initiative is aimed at something further up the productive stack. It is intended to finance compute and full-stack AI infrastructure itself, with possible credit exposure to the equipment and usage-linked cash flows. Put differently: the REIT owns the barn, cold store or packhouse. Compute finance funds the processing line inside it and, depending on the structure, may underwrite the revenue generated by that line.

There will be overlap, of course. AI factories still need land, buildings and power, so data-centre REITs and infrastructure funds remain part of the capital stack. But the risk being underwritten is different. One is mainly property economics. The other is increasingly the economics of productive compute.

Why Jensen keeps emphasising “fungible”

Jensen explicitly described NVIDIA compute as “fungible and transferable across customers and operators”. That word deserves care. An NVIDIA GPU is not perfectly fungible in the legal or commodity sense that one Treasury bill or one standardised barrel is interchangeable with another. Different generations, cluster configurations, networking, power and locations matter.

What Jensen is claiming is operational fungibility: the same compute platform can be redirected across many models, workloads, customers and operators. If one AI lab stops using a cluster, there should be another potential user. If the software workload changes from transformers to mixture-of-experts to agentic AI, CUDA and the surrounding stack are intended to keep the hardware useful.

That is not merely a technology claim. It is a collateral claim. A lender cares deeply about what happens on the bad day, not just the good day. Can the asset be redeployed? Is there a secondary market? Are there many potential offtakers? Does software extend the productive life of the hardware? The better the answers, the lower the perceived stranded-asset risk.

Agriculture has an equivalent. A bank is more comfortable financing a widely used tractor or processing line with an established dealer network, spare parts, service expertise and resale market than a highly specialised machine that only one farm can operate. Same productive purpose, very different collateral quality.

A technology moat can become a financing moat.

The moat Jensen may be building is larger than CUDA

NVIDIA already has the familiar software moat: millions of developers, libraries, tools and workloads built around CUDA. The financing platform potentially adds four more layers. First, an offtaker moat, because a broad pool of users supports residual demand. Second, a residual-value moat, if lenders believe NVIDIA systems can stay economically useful for longer. Third, a standardisation moat, as DSX and the full-stack architecture make an AI factory easier to design, benchmark and underwrite. Fourth, a cost-of-capital moat, if NVIDIA-based infrastructure eventually receives better financing terms than less liquid alternatives.

That last one is especially powerful. If two compute platforms perform similarly, but one can be financed at a lower rate because lenders trust its residual value and secondary market, the cheaper financing effectively lowers the customer’s cost of ownership. Technology adoption improves financeability; financeability then reinforces technology adoption. That is a flywheel competitors cannot solve by adding a few more teraflops.

From fabless semiconductor company to the architect of the AI factory

There is another strategic shift worth applauding, with one factual correction. NVIDIA has not stopped being a fabless semiconductor company. Its history is built on outsourced manufacturing, including its long relationship with TSMC, and NVIDIA still explicitly lists reliance on third parties for manufacturing, assembly, packaging and testing as a business risk.

What has changed is the economic identity Jensen wants the market to assign to NVIDIA. In the financing announcement, he said, in effect, that the company began by building chips and is now helping create productive infrastructure called AI factories. In the DSX launch, he went further: NVIDIA said it is “not just shipping chips” and presented a full-stack playbook spanning silicon, systems, networking, software, facilities and partner technologies.

That is not a cosmetic rebranding. A semiconductor vendor sells a component. A platform company sells a standard. An infrastructure company helps design the whole productive system. And a company that can influence how that system is financed begins to sit even closer to the economic centre of the value chain.

The numbers already show how far NVIDIA’s centre of gravity has moved. In Q1 FY2027, Data Centre generated $75.2 billion of NVIDIA’s $81.6 billion revenue, roughly 92%. The company now describes its growth markets around Data Centre and Edge Computing rather than the old mental model of a graphics-chip company.

Jensen’s “AI factory” language is useful because a factory has inputs, machinery, throughput, utilisation, unit economics and output.

Electricity, data and models enter.

Tokens and intelligence leave.

NVIDIA wants to optimise the whole production system, not merely sell one expensive machine inside it.

And now finance is being added to that stack. Seen this way, the $500 billion MoU is not a side project. It is the financial extension of the AI factory strategy.

This is where the argument gets more interesting

‘s “Who Captures Value in AI Infrastructure?” makes a distinction that investors should tattoo somewhere near their spreadsheets: the amount of money flowing through a layer is not the same thing as the amount of economic value that layer captures.

A farmer can spend a fortune on seed, fertiliser, machinery, storage and freight. That does not mean the farmer earns the highest margin in the agricultural chain. The largest invoice and the largest profit pool are often two very different things.

AI infrastructure looks increasingly similar.

The industry can spend hundreds of billions on servers, data-centre shells, power systems, networking, memory and construction, while the durable economics concentrate in a much smaller number of choke points. Zeoli’s framework is essentially a scarcity map: where the supply is hard to replicate, margins can remain extraordinary; where capacity can be added by anybody with enough capital and a procurement team, competition eventually compresses returns.

NVIDIA is the obvious example. In its fiscal first quarter of 2027, the company reported $81.6 billion of revenue, including $75.2 billion from Data Centre, with a GAAP gross margin of 74.9%. Those figures do not prove the moat will last forever. They do show what value capture looks like when technological scarcity, software lock-in, developer adoption and demand all meet in the same place.

The financing announcement adds another layer to that scarcity map. NVIDIA is no longer only asking customers to prefer its technology. It is helping build a capital-market architecture in which its technology may become easier to finance, easier to redeploy and easier for institutional investors to underwrite. If that develops, the competitive advantage is not merely performance or software. It is a lower-friction path from capital to deployed compute.

That is a different kind of competitive advantage. The next AI moat may partly be a cost-of-capital moat.

AI may be entering its warehouse-receipt moment

This is where agriculture provides a useful analogy.

Agricultural finance has spent decades solving a simple problem: a productive asset can be economically valuable but financially awkward. A crop in the field has value, but a bank cannot easily lend against “probably a good harvest in three months”. A warehouse full of standardised grain is more useful. Once the commodity is graded, stored, verified and represented by a recognised warehouse receipt, the inventory becomes easier to finance. The receipt can serve as collateral. The farmer or trader can unlock working capital without immediately selling the crop.

FAO’s agricultural value-chain finance framework goes much further than conventional bank lending. It includes trader credit, input-supplier credit, lead-firm financing, receivables finance, factoring and warehouse receipts. The point is that financing is structured around the relationships and cash flows of the chain, not merely around the balance sheet of the farmer.

Now replace grain with compute.

A lender looking at a giant GPU cluster has a remarkably similar set of questions. Is the asset standardised? Is there a deep pool of potential users? Can it be transferred to another operator if the original borrower fails? Does it generate contracted cash flow? How quickly does it depreciate? Is there a credible secondary market? Does the asset remain economically useful after the newest generation arrives?

NVIDIA’s own argument goes straight at these questions. It describes its compute as broadly adopted, flexible across models and workloads, transferable across customers and operators, and continuously improved through CUDA. Strip away the marketing language and the financing logic is clear: NVIDIA is trying to make the lender see a GPU cluster not as a pile of electronic equipment, but as a productive machine with a wide user base and financeable cash flows.

In agriculture, the warehouse receipt does not make the grain better. It makes the grain bankable.

In AI, the interesting possibility is that a combination of standardised hardware, CUDA, a deep customer ecosystem and long-term compute contracts could play an analogous role. Not identical, of course. A tonne of maize does not become technologically obsolete because a new tonne of maize launches next year. GPUs do. But the financial principle rhymes.

The financing structures rhyme more than they first appear

The last row is the one I find most interesting.

In agricultural value chains, a lead processor, trader or buyer may provide financing, technical support or a guaranteed market to farmers. This can be genuinely beneficial. It can raise yields, reduce uncertainty and bring smaller producers into a formal commercial chain. But it also changes bargaining power. The party that organises finance, sets standards and controls market access can become more important, not less.

That is the subtlety in NVIDIA’s move. Financing can broaden access to compute while simultaneously deepening NVIDIA’s position as the standard around which that financing is built.

Finance can democratise access to a scarce input while entrenching the supplier that defines the collateral standard.

The $500 billion is not free money, and it is not one giant cheque

There is an important reality check here. NVIDIA has announced memorandums of understanding with the six financial institutions. The company says the partnerships aim to mobilise more than $500 billion over time, and the arrangements remain subject to final agreements. So it would be wrong to treat the headline as $500 billion already committed, funded and sitting in an account waiting for GPU purchase orders.

What matters is the direction of travel. The architecture of AI finance is becoming institutional.

This was already happening before the NVIDIA announcement. In June, Apollo, Blackstone and Broadcom unveiled an AI XPV financing platform with an initial $35 billion capital solution tied to more than 1 gigawatt of Anthropic compute capacity, inside a framework designed to enable more than 20 gigawatts through 2028. Apollo explicitly described AI compute as an emerging asset class with contracted cash flows and mission-critical utility.

That precedent matters because it tells us this is not simply Jensen Huang inventing a clever way to sell more GPUs. Broadcom is pursuing a similar capital-markets solution around its XPU and networking ecosystem. The competition is moving from silicon architecture into financing architecture.

In other words, the chip war is acquiring a banking layer.

Why this could reshape the economics of the entire AI value chain

For NVIDIA: financing becomes a demand accelerant

If a customer has to fund a multi-billion-dollar AI build entirely from its own balance sheet, the hurdle rate is high. Projects compete with share buy-backs, acquisitions, ordinary capital expenditure and every other corporate priority. If external capital can fund the infrastructure against contracted usage and the asset itself, the customer can build sooner and at greater scale.

That is economically similar to farm equipment finance. A farmer who cannot pay cash for a combine harvester may still operate one profitably if the financing cost is lower than the incremental cash flow the machine generates. The manufacturer sells a machine today; the farmer pays over the period in which the machine produces income.

For NVIDIA, the financial effect could be powerful. Cheaper or more available customer financing lowers the friction between demand for AI and an actual purchase order.

There is also a flywheel. If financiers increasingly understand, standardise and price NVIDIA-backed compute, that may lower financing friction further. A lender prefers collateral it knows how to value. A borrower prefers assets lenders are willing to finance. A supplier naturally prefers an ecosystem in which both happen around its products.

For AI labs and cloud operators: capital intensity becomes less suffocating

Frontier AI is an awkward business model because the infrastructure bill arrives before the revenue does. Training clusters, inference fleets, power commitments and data-centre capacity require enormous upfront capital. Structured finance can push some of that burden into long-duration vehicles and match financing more closely to the life of the asset or the life of the customer contract.

Again, agriculture has been doing this forever. A farmer’s cash-flow cycle is brutally mismatched. Seed and fertiliser are paid for before planting; labour is paid during the season; revenue arrives after harvest. Value-chain finance exists partly because businesses fail when the timing of cash flows is wrong, even if the underlying economics are sound.

AI operators have a different crop, but the timing problem is recognisable.

For Wall Street: compute becomes a new pool of yield

For Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR, this is not philanthropy and it is not a vote in a technology popularity contest. It is an origination opportunity. If compute assets can produce durable contracted cash flows, institutional capital can earn credit spreads, infrastructure returns, structuring fees and asset-management fees around them.

The really important shift is that AI infrastructure is migrating from venture-capital language into the language of pensions, insurers, private credit and infrastructure funds. Those pools are vastly larger and generally have longer duration.

Once an asset class becomes legible to insurance capital, pension capital and investment-grade credit, the cost of capital can change dramatically. And when the cost of capital changes, industrial structure changes with it.

For data-centre builders, power, cooling and networking: more finance means more throughput

This is where Zeoli’s scarcity framework becomes useful again. Financing more compute does not only benefit the chip supplier. Every financed cluster creates demand for racks, networking, transformers, switchgear, cooling, fibre, land, construction and electricity.

But the value will not be shared equally. The most attractive economics should still sit where supply is difficult to replicate. If a component can be competitively sourced from ten vendors, finance may increase volume without protecting margins. If a component sits behind a multi-year capacity constraint, qualification barrier or technical bottleneck, more financing can strengthen pricing power.

Capital does not create scarcity. It monetises scarcity.

And then we hit the physical world

This is the part of the AI boom that financial engineering cannot solve by itself.

The International Energy Agency projects global data-centre electricity consumption to reach roughly 945 TWh by 2030 in its base case, more than double the level earlier in the decade and just under 3% of global electricity consumption. Its analysis says accelerated servers, the category most closely associated with AI adoption, account for around 70% of the growth in electricity demand from servers between 2025 and 2030.

The capital markets can finance GPUs. They can finance buildings. They can finance power plants. What they cannot do is compress a seven-year grid interconnection process into seven months by issuing more debt.

This is why I suspect the most valuable AI infrastructure businesses of the next phase will not necessarily be the companies that shout “AI” the loudest. They may be the companies that quietly control the things finance cannot instantly replicate: grid connections, electrical equipment, liquid cooling expertise, advanced packaging, memory capacity, high-speed networking and permitted land with reliable power.

Agriculture has the same reality. You can finance another thousand hectares, but if there is no irrigation, no cold storage, no packhouse capacity or no export logistics, the financing merely helps you reach the bottleneck faster.

Who actually captures the value?

This brings us back to the central question.

A half-trillion-dollar capital programme sounds as though everyone in the ecosystem wins. In gross activity, that may be true. More financing can mean more chips, more data centres, more power infrastructure, more credit assets and more AI capacity.

But “the ecosystem grows” and “every participant earns attractive returns” are not the same statement.

Agriculture is full of examples. A record crop can coexist with terrible farm-gate prices. More production can improve utilisation for a processor while crushing the producer’s margin. A new financing scheme can help farmers plant more acreage while transferring more bargaining power to the buyer providing the credit. Volume and value capture can move in opposite directions.

The same could happen in AI.

If financing unleashes a flood of GPU capacity, compute rental prices could eventually fall. If the customer layer becomes commoditised, utilisation assumptions may disappoint. If alternative accelerators improve, the residual value of today’s hardware could decline faster than lenders expect. Meanwhile, NVIDIA could still capture attractive economics at the point of sale if its technology and software remain the preferred standard.

That is why I would separate the AI value chain into two questions: Who receives the capital? And who has the power to keep the margin?

They are not the same question.

The harvest can be enormous and the farmer can still be poor.

The most provocative possibility: NVIDIA is becoming part of the financial infrastructure

There is a temptation to say “NVIDIA is becoming a bank”. I think that is catchy but slightly lazy.

The company is not announcing that it will simply lend half a trillion dollars from its own balance sheet. The stated model uses independent financing platforms and third-party capital. That distinction matters, especially given concerns about circular financing in the AI ecosystem.

A better description is that NVIDIA is trying to become part of the financial grammar of compute.

Consider what that means. A technology company normally competes on performance, price, software, distribution and customer support. But if its equipment becomes the asset that lenders are comfortable underwriting, it gains another dimension of advantage: financeability.

Aircraft manufacturers understand this. Heavy-equipment makers understand this. Agriculture understands this through machinery finance, warehouse receipts and contract farming. The easier an asset is to finance, insure, redeploy and value, the larger the addressable buyer base becomes.

If NVIDIA compute becomes a preferred form of collateral because it has a deep user ecosystem, known software compatibility and perceived residual demand, then CUDA is doing more than creating developer lock-in. It may indirectly contribute to collateral quality.

That is a fascinating leap. Software could influence the credit spread on hardware.

But collateral is only as good as the assumptions behind it

This is where scepticism is useful.

A warehouse receipt works because the lender believes the commodity exists, has a recognised grade, can be stored without unacceptable deterioration, can be sold into a functioning market and will retain enough value to repay the loan if the borrower defaults.

Compute financing has to answer the same questions in a much faster technological cycle.

A GPU cluster is not merely a box that can be repossessed and auctioned. Its economics depend on power, cooling, networking, software, location, interconnection, operator competence and utilisation. Moving thousands of accelerators from one failed operator to another is not the same as moving sacks of grain from one buyer to another.

Technology obsolescence is also brutal. The lender may underwrite a five-year asset while the industry discusses the next architecture every year. A residual-value model that looks conservative in year one can look heroic in year three if performance per watt improves faster than expected or custom silicon takes share.

And there is still a circularity question. This is precisely why the announcement’s emphasis on third-party capital and independent underwriting matters: it makes the structure less self-referential than direct vendor financing. But it does not magically disinfect the ecosystem. NVIDIA still benefits from the hardware sale, may provide a residual-value backstop on some deals, and every participant still depends on sustained utilisation and credible end demand. Independent underwriting is a stronger market test, not a guarantee that the cycle cannot become overextended.

This does not mean the structure is unsound. It means investors should stop treating “financed demand” as economically identical to “unlevered end demand”. Credit can reveal demand, but it can also pull future demand into the present.

Agriculture offers one final warning

When financing enters a value chain, it often improves the chain. It can remove working-capital bottlenecks, reduce forced selling, smooth production and connect smaller participants to buyers. FAO’s work on warehouse-receipt finance, for example, highlights how better collateral structures can let producers avoid immediately selling after harvest when prices are weak.

But financing is never neutral. It changes who has optionality.

A farmer with storage and credit can choose when to sell. A farmer without them often has to accept the harvest-time price. A processor offering input finance can secure supply. A trader advancing cash can secure purchasing rights. A bank accepting warehouse receipts gains a better form of collateral. Each instrument solves a problem, but each also redistributes bargaining power.

That is exactly how I would think about AI infrastructure financing.

The important question is not simply whether $500 billion of capital arrives. It is what options that capital gives each participant, and what obligations come attached to it.

Does the AI lab gain flexibility, or merely take on more fixed commitments? Does the data-centre operator gain cheaper capital, or become more exposed to one technology cycle? Does the financier own genuinely transferable collateral, or a specialised asset whose value disappears when utilisation drops? Does cheaper finance lower the end price of compute, or is part of the benefit captured upstream by the supplier with the strongest moat?

Those are the questions that determine who harvests the economics.

What I would watch from here

The headline number will get the attention, but I would watch the plumbing.

First, watch the actual terms as the memorandums become final agreements. Debt tenor, advance rates, collateral packages, covenants, utilisation requirements, offtake commitments and residual-value assumptions will tell us far more than the $500 billion headline.

Second, watch whether financing costs begin to differ meaningfully by compute platform. If NVIDIA-backed infrastructure can borrow more cheaply than less-standardised alternatives, the capital market itself could reinforce technological market share. Conversely, if Broadcom and hyperscaler custom silicon win equally attractive financing, the “finance moat” may become an industry feature rather than an NVIDIA advantage.

Third, watch secondary-market behaviour. A real asset class needs price discovery after the initial sale. What happens to a two-year-old rack when a borrower restructures? Who buys it? At what haircut? How quickly can it be redeployed? Those answers will determine whether “compute-backed credit” behaves like infrastructure finance or like expensive equipment lending dressed in AI language.

Fourth, watch the physical bottlenecks. More capital chasing limited power, transformers, cooling capacity and construction labour can initially raise costs rather than lower them. Finance is an accelerator. It accelerates shortages too.

And finally, watch who starts offering finance next. Once customers learn that capital is part of the product, competitors cannot ignore it. The battle for AI share may increasingly include not only tokens per dollar, but basis points per dollar.

The bigger idea

Every industrial revolution eventually learns finance.

Railways learned bonds. Aviation learned aircraft leasing. Agriculture developed crop finance, warehouse receipts, receivables finance and contract farming. Renewable energy learned project finance and power-purchase agreements.

AI is now learning compute credit.

That is why I think the NVIDIA announcement matters well beyond NVIDIA. It marks a transition from “How much AI infrastructure can technology companies afford to buy?” to “How much AI infrastructure can global capital markets underwrite?”

That second pool of money is much larger.

But larger pools of capital do not abolish economics. They make the allocation of economic power more important. When money becomes abundant, scarcity elsewhere becomes more valuable.

So yes, the AI buildout may be one of the largest capital formation events of our time. Yes, structured finance could help accelerate deployment and spread the benefits across chipmakers, data-centre builders, utilities, financiers and AI customers.

But if there is one lesson agriculture has taught me, it is this: never confuse a bigger harvest with a better margin.

The question is not simply how much gets planted.

The question is who owns the seed, who controls the storage, who provides the credit, who has the buyer, and who can still say no when the crop is ready.

In AI, the equivalent question is becoming unavoidable: who writes the rules for the collateral?

Sources and references

• Research note: The linked YouTube page identifies the discussion as “NVIDIA’s Jensen Huang & Partners Today On NVIDIA $500B”.

• Chris Zeoli, “Who Captures Value in AI Infrastructure?”, Data Gravity, 13 July 2026

Data Gravity

Who Captures Value in AI Infrastructure?

In brief…

Read more

2 months ago · 83 likes · 10 comments · Chris Zeoli

• NVIDIA, compute infrastructure financing platforms announcement, 10 August 2026
https://nvidianews.nvidia.com/news/nvidia-partners-with-apollo-blackrock-blackstone-brookfield-goldman-sachs-and-kkr-to-establish-ai-compute-infrastructure-financing-platforms-to-mobilize-over-500-billion-of-third-party-capital

• NVIDIA, Q1 FY2027 financial results, 20 May 2026
https://investor.nvidia.com/news/press-release-details/2026/NVIDIA-Announces-Financial-Results-for-First-Quarter-Fiscal-2027/default.aspx

• Apollo, Broadcom AI XPV $35 billion capital solution, 9 June 2026
https://www.apollo.com/insights-news/pressreleases/2026/06/apollo-leads-35-billion-capital-solution-for-broadcom-ai-xpv-platform-in-partnership-with-blackstone-and-leading-global-banks-3308896

• FAO, Agricultural Value Chain Finance
https://www.fao.org/fileadmin/user_upload/ags/publications/AVCF_2011e3.pdf

• International Energy Agency, Energy and AI
https://www.iea.org/reports/energy-and-ai

• NVIDIA, Enterprise AI Factory Design Guide, May 2026
https://docs.nvidia.com/ai-enterprise/planning-resource/ai-factory-white-paper/latest/ai-factory-overview.html

• NVIDIA, DSX Gives Infrastructure Builders the Playbook for AI Factories, 31 May 2026
https://investor.nvidia.com/news/press-release-details/2026/NVIDIA-DSX-Gives-Infrastructure-Builders-the-Playbook-for-AI-Factories/default.aspx

• U.S. SEC, Investor Bulletin: Publicly Traded REITs
https://www.sec.gov/resources-for-investors/investor-alerts-bulletins/ib_reits

• NVIDIA, Corporate Timeline
https://www.nvidia.com/en-us/about-nvidia/corporate-timeline/

• Reuters, “Nvidia partners with Wall Street giants to raise $500 billion for AI buildout”, 10 August 2026
https://www.reuters.com/technology/wall-street-giants-partner-with-nvidia-500-billion-ai-financing-deal-ft-reports-2026-08-10/

• Reuters Breakingviews, “Jensen Huang takes wheel of $500 bln AI bandwagon”, 11 August 2026
https://www.reuters.com/commentary/breakingviews/jensen-huang-takes-wheel-500-bln-ai-bandwagon-2026-08-11/

• Financial Times, “Nvidia’s chips may be novel, but its ‘circular financing’ isn’t”, 29 July 2026
https://www.ft.com/content/33714af0-a646-4271-8078-49a87182917f

If you found value in this article, please hold down the like ❤️ button or consider buying me a coffee ☕️ to fuel the next one!

Buy me coffee ☕️

Subscribe now

The Long Game ♟️ Perspective: A Weekly Alpha series:

ServiceNow shares have fallen sharply even as subscription revenue grows above 20%. The investment case now rests on whether it becomes the control layer for enterprise AI agents, or merely another platform they learn to bypass.

ServiceNow is one of the more uncomfortable holdings in my portfolio.

The business is growing subscription revenue above 20%. Contracted future revenue continues to expand. Its AI products have crossed US$1 billion in annual contract value.

Yet the shares remain roughly 45% below their 52-week high and were down around 38% in 2026 following the latest results. ServiceNow closed at US$111.23 on 31 July, giving the company a market capitalisation of approximately US$115 billion.

Usually, when a company reports 24.5% subscription revenue growth and raises its annual guidance, investors are not busy discussing its possible extinction.

ServiceNow is different.

The question hanging over the company is simple:

What happens to workflow software when AI agents can perform the workflows themselves?

It is not a foolish question.

An AI agent can increasingly answer employee requests, generate applications, investigate incidents, retrieve company information and execute actions across software systems.

If companies can build these agents directly using OpenAI, Anthropic, Microsoft or Google, why continue paying ServiceNow?

That was broadly how I viewed the risk.

Then I began digging into a product called AI Agent Orchestrator, and the larger architecture ServiceNow has built around it.

The deeper I went, the more the investment question changed.

Perhaps AI agents will not eliminate the need for ServiceNow.

Perhaps the coming explosion of agents will create precisely the organisational chaos that makes ServiceNow more valuable.

The market believes AI will eat software

The conventional bear case is easy to understand.

Traditional software companies sell applications that help humans perform work.

AI agents threaten to automate portions of that work, reduce the number of human users and make it easier for companies to develop customised software internally.

That creates three risks for ServiceNow.

First, companies may need fewer employee licences if AI performs more tasks.

Second, AI coding tools may allow businesses to build simple workflow applications without purchasing additional ServiceNow modules.

Third, foundation-model providers may move upwards from supplying intelligence into applications, workflow orchestration and enterprise governance.

This third risk became very real on 22 July.

Hours before ServiceNow announced its second-quarter results, OpenAI introduced Presence, an enterprise product designed to deploy AI agents that can answer questions, use company systems, take approved actions and escalate cases to people.

Presence is initially available through a limited programme led by OpenAI engineers and selected systems integrators. It is not yet a self-service platform, but its direction is unmistakable. OpenAI does not intend to remain merely an intelligence supplier. It wants to participate in the enterprise workflow layer too.

TD Cowen analysts observed that Presence packages many of the capabilities promoted by software vendors, including reasoning, data access, policies and governance.

UBS had already downgraded ServiceNow earlier in 2026, arguing that AI coding tools could help customers recreate lightweight applications and place greater pressure on conventional software budgets.

The fear is not that ServiceNow disappears overnight.

The more credible fear is slow economic erosion:

1. AI reduces the time required to build workflows.

2. Customers need fewer human seats.

3. Model providers bundle basic orchestration and governance.

4. ServiceNow must lower prices or spend more heavily to differentiate itself.

5. Revenue continues growing, but margins and pricing power deteriorate.

That is the threat the stock price is attempting to discount.

But it may be underestimating a different problem.

Intelligence is becoming abundant. Coordination is not.

Imagine a large company five years from now.

Its finance department has AI agents reviewing invoices and investigating anomalies.

Human resources has agents answering employee questions and processing leave requests.

Cybersecurity has agents assessing vulnerabilities and responding to threats.

Customer service has agents resolving complaints.

Software teams have coding agents building and repairing applications.

Some agents are built using OpenAI models. Others use Anthropic, Google, Microsoft or open-source models. Still others are developed internally.

The company may have hundreds or eventually thousands of autonomous agents.

Each agent may be individually useful.

Collectively, they create a management problem.

Who decides which agent handles a request?

What information may it access?

Can it approve a payment?

Can it reset an employee’s password?

What happens when two agents recommend conflicting actions?

How does the company know what an agent changed?

Who becomes responsible when an agent makes a mistake?

How does management measure whether the agent saved money or merely consumed more computing resources?

This is the hidden bottleneck behind enterprise AI.

Companies do not merely need more intelligence.

They need permission, context, coordination, execution and accountability.

ServiceNow believes it can supply that layer.

My interpretation is that ServiceNow is trying to impose an orchestration tax on enterprise AI.

The company does not need to create the smartest model.

It needs to become the trusted platform through which intelligent models are allowed to act.

Understanding ServiceNow’s architecture

ServiceNow’s product language can quickly become a bowl of expensive alphabet soup.

The system is easier to understand if we treat it like an airport.

ServiceNow Otto is the front desk

Otto is ServiceNow’s new unified AI interface, combining Now Assist, Moveworks and its broader AI experience.

An employee can ask Otto for something in ordinary language without knowing which department, application or workflow should handle the request.

Otto interprets the request and attempts to move it towards completion across multiple systems.

It is the place where human intention enters the system.

AI Agent Orchestrator is air-traffic control

ServiceNow’s AI Agent Orchestrator coordinates collaboration between multiple agents.

A complex IT request might require one agent to diagnose the problem, another to check the affected infrastructure, another to verify permissions and another to execute the approved repair.

The orchestrator plans the workflow, routes work between agents and helps them operate as a team rather than a collection of isolated chatbots.

The CMDB is the map

ServiceNow’s Configuration Management Database, or CMDB, maps a company’s applications, infrastructure, services and their relationships.

Knowing that a server exists is not enough.

An agent needs to understand which business services depend on it, which employees will be affected, which changes require approval and what happened during similar incidents.

This operational context is one of ServiceNow’s most important assets.

A model may understand technology generally.

The CMDB helps it understand how a particular company actually works.

Action Fabric is the runway

ServiceNow’s Action Fabric lets agents take approved actions through existing enterprise workflows.

The company has also opened its platform through a Model Context Protocol server, allowing agents built using Claude, Microsoft Copilot or a company’s internal tools to interact with ServiceNow.

The strategy is deliberately model-neutral.

ServiceNow does not need customers to use only its agents. It wants external agents to execute through its infrastructure.

AI Control Tower governs the system

AI Control Tower is designed to discover, observe, secure and measure agents running across the enterprise.

ServiceNow says the platform can discover AI deployments across more than 30 enterprise integrations, monitor agent behaviour, enforce risk controls, govern identities and track AI spending.

The Control Tower is supposed to answer questions such as:

• What agents are active?

• Who owns them?

• Which systems can they access?

• Are they complying with company policies?

• What did they do?

• What value did they create?

• Should one be suspended?

Put together:

Otto receives the request. The Orchestrator coordinates the agents. The CMDB supplies the context. Action Fabric executes the work. The Control Tower watches everything.

That is the investment thesis in product form.

Why ServiceNow may be better positioned than it appears

ServiceNow’s advantage is not artificial intelligence itself.

Frontier models will continue improving, and the cost of intelligence will probably decline.

Its advantage is the accumulated knowledge of how work moves through large organisations.

ServiceNow already sits inside IT operations, employee services, customer support, security, risk and application development for many large enterprises.

Its platform contains:

• incident histories;

• approval chains;

• asset relationships;

• employee roles;

• workflow rules;

• service catalogues;

• access permissions;

• audit records;

• and integrations with older systems.

That information is not glamorous.

It is also difficult to replace.

An AI agent capable of reasoning is not automatically authorised to change a production system, approve a refund or access confidential employee information.

ServiceNow’s opportunity lies between knowing what should be done and being trusted to do it.

This is why management increasingly describes the company as the “AI control tower” or the “agent of agents”.

The language is promotional, but the underlying position is credible.

Models supply intelligence.

ServiceNow wants to supply enterprise memory, permission and execution.

Read more

Subscribe now

The Long Game ♟️ Perspective: An occasional Wealth Mindset series:

When people explain compound interest, they usually begin with a calculator.

Invest a certain amount every month.

Assume a reasonable return.

Wait for several decades.

Watch the number grow.

The maths is easy enough to understand.

The difficult part is living through those decades without destroying the calculation.

A spreadsheet does not lose its job.

It does not have ageing parents.

It does not receive a medical bill.

It does not panic during a market crash.

It does not suddenly decide that everyone else seems to be living better.

Human beings do.

That is why compounding is less about mathematics than behaviour.

The formula may fit on one line.

Living it can take a lifetime.

Your Real Capital Is Not Just Money

When you are young, it may feel as though you have very little.

A small salary.

A modest account balance.

Perhaps some debt.

You look at people with large portfolios and assume that capital is what separates you from them.

But the greatest asset a young person possesses is often invisible.

It is not the first S$10,000.

It is the next 30 or 40 years.

Money can be lost and earned again.

Time cannot.

A modest amount invested for decades may grow into something meaningful.

A larger amount invested too late has less time to work.

This is not because starting young guarantees wealth. It does not.

Returns vary.

Life interrupts.

Plans change.

But an early start gives you something valuable: room for imperfect decisions.

You can survive a bad year.

Recover from a mistake.

Change careers.

Learn through a market cycle.

Begin again without the clock immediately becoming your enemy.

Time does not remove risk.

It gives you more opportunities to recover from it.

Buffett’s Real Advantage Was Endurance

Warren Buffett is rightly admired for his ability to evaluate businesses.

But his extraordinary fortune was not created by investment intelligence alone.

He began young and remained invested for an unusually long time.

His returns mattered.

The duration mattered more than many people realise.

A spectacular return maintained for three years may produce an impressive story.

A sensible return maintained for several decades may produce a fortune.

That is the part people find boring.

We are attracted to the annual return because it looks like the engine.

But time is the road.

A powerful engine is not very useful if the journey ends after two kilometres.

The search for higher returns often causes investors to sacrifice the very thing compounding needs most: continuity.

They jump between funds.

Chase what has already risen.

Take risks they cannot emotionally withstand.

Sell when frightened.

Return after prices recover.

Then wonder why their results never resemble the beautiful charts.

They did not lack opportunities.

They lacked a system they could remain inside.

Wealth Is What You Do Not Need to Display

A new car is visible.

A watch is visible.

A holiday is visible.

Savings are usually not.

That makes wealth psychologically difficult.

Consumption provides immediate evidence.

You can drive it, wear it, photograph it and let other people admire it.

Financial resilience has no logo.

Nobody sees the emergency reserve that allowed you to leave a terrible job.

Nobody applauds the mortgage you chose not to maximise.

Nobody notices the investment contribution made instead of upgrading a perfectly functional car.

Morgan Housel makes a simple but uncomfortable distinction in The Psychology of Money:

Richness is often visible.

Wealth is usually hidden.

The car shows money that has been spent.

The portfolio shows money that has been kept.

Neither is automatically right or wrong.

Money exists partly to make life enjoyable.

A life spent refusing every pleasure in order to admire a growing account would be financially disciplined and spiritually rather bleak.

The question is not whether you spend.

It is whether your spending quietly removes your future choices.

Savings are not only unused money.

They are stored flexibility.

They allow you to wait.

Say no.

Change direction.

Help somebody.

Survive a period when life stops behaving according to plan.

You are not merely saving dollars.

You are purchasing room to breathe.

The Point Is Not Never Touching Your Capital

Some investors treat their principal as sacred.

They believe success means never withdrawing a dollar.

But capital is supposed to support life, not become a museum exhibit.

The real danger is not using money.

It is being forced to use it at the worst possible time.

Imagine putting every available dollar into the market.

It appears efficient.

Then a job loss arrives during a market decline.

The portfolio is down.

The rent is due.

The investor has no emergency reserve.

The choice is no longer whether to sell.

It is what to sell first.

That is how compounding ends.

Not because the investment was necessarily wrong.

Because the household had no margin for error.

An emergency reserve may appear unproductive.

Cash earns less than long-term assets.

It can feel like a lazy employee sitting in the corner.

But that cash has a job.

Its job is to prevent a temporary crisis from becoming a permanent financial wound.

The investment portfolio is designed to grow.

The emergency fund is designed to protect the portfolio’s time.

One is the engine.

The other is the spare tyre.

You may resent carrying the spare tyre until the day the road is dark and the wheel gives way.

Survival Comes Before Return

The most important rule of compounding is not maximising performance.

It is avoiding permanent interruption.

That means basic life must come first.

Housing.

Food.

Insurance against catastrophic risks.

High-interest debt.

Emergency liquidity.

Only money that is unlikely to be needed soon should be exposed to long-term market risk.

This order can feel slow.

People want to invest immediately because investing feels like progress.

Paying off expensive debt feels defensive.

Keeping cash feels unambitious.

But compounding is not impressed by appearances.

A credit card balance growing at a high interest rate compounds too.

Just in the wrong direction.

A person earning 8% on an investment while paying 20% on debt is not investing intelligently.

They are filling a bucket while somebody drills holes in the bottom.

The glamorous part of wealth is investing.

The essential part is building a financial life sturdy enough to let investments remain invested.

The Best Plan Is the One You Can Continue

Theoretical returns attract attention.

Behaviour determines what you actually receive.

A volatile strategy may offer higher expected returns.

But if it causes you to lose sleep, check prices constantly and sell during downturns, it is not the right strategy for you.

A calmer, diversified approach that you can hold for decades may produce the better real-world result.

This can feel unsophisticated.

The financial world often rewards complexity with impressive language.

But your portfolio does not care how intelligent the strategy sounds.

It cares whether you stay.

A plan should survive:

• difficult markets

• career changes

• family responsibilities

• periods of lower income

• boredom

• fear

• your own occasional foolishness

If it works only when life is calm, it does not work.

At the Beginning, Invest in the Person Too

A twenty-year-old with a small portfolio should invest.

But the greatest return may not come from the portfolio yet.

It may come from increasing the ability to earn.

A useful qualification.

A technical skill.

The ability to communicate.

The ability to sell.

Industry knowledge.

A piece of work that can be reused.

A reputation for solving difficult problems.

An additional S$500 of monthly earning power may matter far more in the early years than squeezing another percentage point from a small account.

This does not mean endlessly buying courses.

A course is not automatically an asset because somebody called it an investment in yourself.

The test is practical.

Will this knowledge improve your judgement?

Increase your usefulness?

Raise your earning potential?

Help you create something that continues producing value?

If not, it may simply be consumption wearing academic clothing.

The same scepticism should apply to everything.

A home is not automatically a liability.

A car is not always wasteful.

An investment product is not automatically an asset.

The right question is:

What role does this play in the life I am trying to build?

Compounding Is a Behavioural Skill

Compound interest is usually described as money earning money.

But the financial compounding we see at the end rests on several quieter forms of compounding.

The habit of saving.

The ability to delay gratification.

The judgement to avoid obvious traps.

The confidence to remain calm.

The skill that gradually increases income.

The reputation built through repeated reliability.

The health that allows you to keep working, thinking and caring for others.

Money compounds best when the person managing it is also becoming more capable.

That is why financial success cannot be reduced to picking the right product.

Two people can own the same fund and experience completely different outcomes.

One remains invested.

The other panics.

One increases contributions as income rises.

The other increases lifestyle.

One keeps an emergency reserve.

The other is forced to sell.

The investment was identical.

The behaviour was not.

Three Things You Can Do Now

You do not need to redesign your entire life this weekend.

Begin with three ordinary actions.

First, track one month of income and spending.

Not to punish yourself.

To see what is actually happening.

Most financial problems remain powerful because they are vague.

Numbers make them less mysterious.

Second, separate emergency money from long-term money.

The emergency fund protects the present.

The investment portfolio serves the future.

Do not ask one pool of money to perform both jobs.

Third, automate what matters.

When income arrives, move a sustainable portion towards savings and long-term investment before the rest is absorbed by life.

Automation is useful because motivation is unreliable.

Your future should not depend on whether you feel disciplined on payday.

The Goal Is Not a Perfect Beginning

Starting at twenty is helpful.

Starting at thirty is still useful.

Starting at forty is better than continuing to wait.

People lose more time waiting for the perfect moment than they lose by beginning imperfectly.

Markets may seem expensive.

Income may feel insufficient.

You may believe you need to understand everything first.

You do not.

You need a sensible starting point and a system that can improve.

The greater danger is repeatedly resetting yourself.

Chasing fast returns.

Taking risks you cannot afford.

Interrupting the plan whenever life becomes uncomfortable.

Beginning again every few years with a new strategy and less confidence.

Compounding does not demand brilliance.

It demands that you stop destroying your own progress.

You do not need one heroic investment to change your family’s future.

You need to preserve part of today.

Some money.

Some skill.

Some health.

Some attention.

Then give those things enough time to become more useful.

The greatest gift of understanding compounding is not knowing how much S$1 may become.

It is recognising that small, sensible choices do not remain small when they are protected for long enough.

At twenty-two, your greatest asset may not be your money.

It may be the forty years in front of it.

But time is not automatically on your side.

You must build a life capable of staying beside it.

If you found value in this article, please hold down the like ❤️ button or consider buying me a coffee ☕️ to fuel the next one!

Buy me coffee ☕️

Subscribe now

The Long Game ♟️ Perspective: An occasional Alpha series:

Tesla just delivered its strongest second quarter in history.

Revenue rose 26%. Vehicle deliveries increased 25%. Active Full Self-Driving subscriptions grew 56%. Energy-storage deployments reached 13.5 GWh.

Yet operating income fell 57%. Operating margin collapsed to 1.4%. Capital expenditure reached US$5.79 billion, and free cash flow turned negative.

Tesla’s problem is no longer growth.

It is the economics of growth.

The bear case is not that Tesla has no future. The bull case is not proved merely because that future has become more visible.

The real question is whether Tesla’s automotive business can finance an extraordinary collection of future businesses without destroying too much value along the way.

Tesla shareholders are being asked to pay for the bridge before anyone knows how much traffic will eventually cross it.

The quarter in one table

Tesla ended June with US$43.52 billion of cash and short-term investments. This is not a liquidity crisis. It is a returns-on-capital problem.

Source: Tesla Q2 2026 shareholder deck

Growth returned. Profit did not.

Tesla’s top line looked like a growth company.

Its operating line looked like a construction site.

Total revenue reached US$28.24 billion, up from US$22.50 billion a year earlier. But operating expenses rose 47% to US$4.35 billion, while operating income fell to just US$398 million. Adjusted EBITDA declined 4%, despite the substantial revenue increase.

Source: Tesla Q2 2026 shareholder deck

Chart 1: Revenue returned to growth, but very little reached operating profit.

Source: Tesla Q2 2026 shareholder deck

Reported net income also deserves care.

Tesla recognised a US$1.005 billion pre-tax unrealised gain on its SpaceX investment during the quarter. Tesla correctly excluded the gain from adjusted EBITDA and non-GAAP earnings because an investment revaluation is not evidence that the operating business became more profitable.

Source: Tesla Q2 2026 shareholder deck

Tesla generated substantially more revenue while producing less operating profit.

That does not prove the investment programme is failing. Growth investments often depress current margins before producing future returns.

But “we are investing” is not a complete capital-allocation argument.

The eventual cash flows must justify the capital consumed.

Cars still pay for tomorrow

Tesla increasingly describes itself through AI, robotics, autonomy and energy.

Economically, cars still fund the transformation.

Automotive revenue was US$20.52 billion in Q2, representing nearly 73% of total revenue. Energy contributed US$3.14 billion, while services and other revenue reached US$4.58 billion.

Source: Tesla Q2 2026 Form 10-Q

That makes the health of the automotive business more important, not less.

Tesla delivered 480,126 vehicles while producing 451,758. Deliveries exceeded production by more than 28,000 vehicles, helping reduce global inventory from 27 days of supply in Q1 to 15 days in Q2. Selling previously produced inventory is not inherently negative, but it means part of the delivery surge cannot be repeated indefinitely without fresh production and demand.

Source: Tesla Q2 production, deliveries and deployments

The more uncomfortable number is margin.

Automotive gross margin excluding regulatory-credit sales fell from 19.2% in Q1 to 16.3% in Q2. Regulatory-credit revenue also declined from US$439 million a year earlier to US$146 million.

Source: Tesla Q2 2026 shareholder deck

Chart 2: Record Q2 deliveries did not produce better automotive economics.

Source: Tesla Q2 2026 shareholder deck

This is the tension at the centre of the quarter.

Tesla sold far more cars, but it retained less gross profit from each dollar of automotive revenue.

The demand-collapse thesis has weakened. The profitability problem has not.

Tesla has proved that it can recover volume.

It has not yet proved that recovering volume will restore attractive automotive margins.

Tesla’s strongest moat may be the installed base

The most compelling bull argument does not begin with this quarter’s vehicle sales.

It begins with the millions of Tesla products already on the road.

Cumulative deliveries reached approximately 9.7 million. Active FSD subscriptions increased to 1.48 million from 950,000 a year earlier. More than 55% of new North American deliveries included an FSD subscription during the quarter.

Source: Tesla Q2 2026 shareholder deck

That installed base gives Tesla several opportunities to earn revenue after the original vehicle sale:

· FSD subscriptions

· Charging

· Maintenance and collision repair

· Insurance

· Connectivity

· Software upgrades

· Future autonomous-transport services

This is closer to a connected-device platform than a conventional one-off vehicle transaction.

Services and other revenue grew 50% to US$4.58 billion. The segment produced a record US$648 million of gross profit, with a gross margin of approximately 14%. The segment is not pure software, but its growth shows that Tesla is becoming better at monetising customers after they purchase a vehicle.

Source: Tesla Q2 2026 shareholder deck

The weakness is disclosure.

Tesla reports subscriber counts and aggregate deferred automotive software revenue, but it does not separately disclose FSD revenue, gross margin, churn or customer-acquisition economics. Deferred revenue associated with FSD, connectivity, software updates and free Supercharging totalled US$4.05 billion, but these components remain combined.

Source: Tesla Q2 2026 Form 10-Q

That leaves investors with evidence of adoption, but incomplete evidence of profitability.

An installed base is a potential moat. A stream of high-return cash flow is an economic moat.

Tesla has clearly built the first.

The second is still partly hidden.

Robotaxi is real. Its economics are not.

It is no longer fair to describe Tesla’s autonomy ambitions as nothing more than presentation slides.

Cybercab production has begun at Gigafactory Texas. Tesla expanded unsupervised operations in Austin, launched unsupervised rides in Miami, Orlando and Tampa, and reported that its robotaxi service was live in seven major metropolitan areas.

Source: Tesla Q2 2026 shareholder deck

That is genuine progress.

It is not yet enough to value the business confidently.

Tesla does not separately disclose:

· Robotaxi revenue

· Active commercial fleet size

· Rides completed

· Revenue per mile

· Vehicle utilisation

· Insurance expenses

· Remote-assistance costs

· Contribution margin

· Capital required to enter each new city

A map with more cities is operational evidence. It is not yet an income statement.

Tesla may be approaching the point where autonomy scales non-linearly. Reliability improvements and regulatory approval could allow a small network to expand rapidly.

But until the company discloses fleet economics, investors are being asked to assign a large value to a business without knowing how much each ride earns, how much each vehicle costs to operate or how much capital each city requires.

The correct valuation is not zero.

It is also not whatever number is needed to justify the share price.

Energy is too large to ignore, but too volatile to carry the thesis

Tesla’s energy-storage business is the strongest evidence that the company can build a major platform outside automobiles.

Storage deployments reached 13.5 GWh, up 41% year on year. Energy revenue increased 13% to US$3.14 billion. Tesla also reported US$10.05 billion of unsatisfied or partially unsatisfied long-term energy performance obligations, of which US$4.56 billion is expected to be recognised over the next 12 months.

Source: Tesla Q2 2026 shareholder deck

Those are meaningful figures.

The margin story was less impressive.

Energy gross margin fell from 30.3% to 20.4%, with Tesla citing sales mix and unfavourable warranty adjustments.

Source: Tesla Q2 2026 Form 10-Q

The energy business therefore presents the same contradiction as the company as a whole:

· Strong demand

· Growing capacity

· Significant long-term opportunity

· Uncertain margin durability

Megapack and Powerwall are still physical products. Their economics are influenced by batteries, commodities, warranties, tariffs, manufacturing yields and competition.

Energy could eventually deserve a premium valuation if Tesla demonstrates stable margins, strong cash conversion and attractive returns on each new factory.

For now, energy strengthens the business case.

It does not solve the valuation case.

Tesla is entering its most capital-intensive era

This was the most important development in Q2.

Tesla spent US$5.79 billion on capital expenditure during the quarter, compared with US$2.49 billion in Q1. Operating cash flow was US$4.70 billion, leaving free cash flow at negative US$1.09 billion.

Source: Tesla Q2 2026 shareholder deck

Chart 3: Capital expenditure exceeded operating cash flow in Q2.

Source: Tesla Q2 2026 shareholder deck

Tesla now expects 2026 capital expenditure to exceed US$25 billion. The programme includes AI compute infrastructure, data centres, manufacturing and research facilities, company-operated autonomous assets, retail, servicing and charging infrastructure. Tesla also acknowledges that periods of elevated spending may require financing beyond operating cash flow.

Source: Tesla Q2 2026 Form 10-Q

The company is simultaneously investing in vehicles, Cybercab, Optimus, Tesla Semi, batteries, lithium refining, energy storage, solar manufacturing, AI infrastructure and semiconductor development.

That is not one expansion plan.

It is several industrial revolutions sharing one balance sheet.

There is also a shareholder cost.

Tesla’s reported common shares outstanding increased from approximately 3.751 billion at the end of 2025 to 3.949 billion at the end of June 2026, an increase of roughly 5.3%. The filing attributes the issuance primarily to equity incentives and acquisitions. Stock-based compensation reached US$1.15 billion during Q2 and US$3.80 billion over the trailing 12 months.

Source: Tesla Q2 2026 Form 10-Q

Again, none of this proves the investment is misguided.

It establishes the hurdle.

The future businesses must eventually create enough value to compensate shareholders for the cash spent, shares issued, depreciation incurred and projects that do not succeed.

Capital expenditure is not automatically bullish.

It becomes bullish only when the cash generated by the assets exceeds the cash consumed to build them.

The market is pricing several successes at once

At Tesla’s 24 July closing price of US$313.03 and using the 3.9495 billion common shares reported as outstanding, the implied equity value is approximately US$1.24 trillion.

After adjusting for reported cash, short-term investments and debt, the implied enterprise value is roughly US$1.20 trillion.

Source: Tesla Q2 2026 Form 10-Q

Valuation note: This calculation uses Tesla’s reported common shares outstanding, rather than the lower quarterly diluted weighted-average share count.

Over the trailing 12 months, Tesla generated approximately:

· US$103.6 billion of revenue

· US$15.3 billion of adjusted EBITDA

· US$5.76 billion of free cash flow

· US$3.80 billion of GAAP net income

Source: Tesla Q2 2026 shareholder deck

Present multiples are distorted because Tesla is in a heavy investment cycle.

That makes reverse valuation more useful than arguing about whether Tesla deserves 100, 150 or 300 times present earnings.

What must Tesla eventually earn?

These are not forecasts.

They are the cash-flow hurdle embedded in the present valuation.

Even at a generous mature multiple of 30 times free cash flow, Tesla would eventually need to generate approximately US$40 billion annually.

That probably requires several things to go right:

· Automotive margins must recover.

· Energy must become larger and consistently profitable.

· FSD must produce high-margin recurring revenue.

· Robotaxi must achieve meaningful scale and attractive unit economics.

· Optimus or another new platform must contribute real profit.

· Capital expenditure must eventually normalise.

· Dilution must remain controlled.

The market is not valuing Tesla as though one of these outcomes will succeed.

It is valuing Tesla as though several will.

That is not a margin of safety.

It is a tightly packed itinerary.

The strongest bull case

Tesla owns a combination of assets that few companies can reproduce:

A recognised global brand. Nearly 10 million cumulative deliveries. A growing FSD subscriber base. Direct consumer distribution. Manufacturing expertise. A large charging network. Real-world AI data. Expanding energy-storage capacity. And more than US$43 billion of cash and short-term investments.

Source: Tesla Q2 2026 shareholder deck

Tesla does not need every project to dominate its industry.

A sufficiently successful combination of FSD, services, energy and robotaxi could transform the earnings mix.

Software and fleet income could improve margins even if vehicle manufacturing remains competitive. A small number of high-return platforms may eventually justify today’s spending.

The bull thesis is therefore serious:

Tesla is deliberately depressing today’s profit to build several much larger, higher-margin businesses.

That thesis cannot be dismissed merely because current free cash flow is weak.

But it must be demonstrated, not assumed.

The strongest bear case

The bear case is not that Tesla has stopped growing.

It is that the quality of growth remains weak relative to the valuation.

Revenue rose 26%, but operating income fell 57%.

Deliveries increased 25%, but automotive margin excluding regulatory credits declined.

Energy deployments grew 41%, but energy gross margin fell sharply.

FSD subscriptions increased, but Tesla still does not disclose their standalone economics.

Capital expenditure rose 142%, free cash flow turned negative and the reported share count increased.

Source: Tesla Q2 2026 shareholder deck

Meanwhile, the businesses carrying most of the valuation remain early.

Robotaxi lacks disclosed unit economics. Cybercab is in initial production and testing. Optimus has not yet produced meaningful external revenue. Semiconductor manufacturing remains an investment project rather than a demonstrated profit centre.

The bear thesis is straightforward:

Tesla is funding speculative, capital-intensive businesses with an automotive operation whose margins are weakening, while investors are already paying for those future businesses as though they have succeeded.

That is not an argument that Tesla will fail.

It is an argument that ordinary success may not be enough.

Where I may be wrong

A fair thesis must explain how it could fail.

FSD economics may be far better than Tesla’s current disclosure suggests. If subscriptions carry software-like margins and churn continues falling, the installed base could become a powerful recurring-revenue engine.

Robotaxi may scale suddenly rather than gradually. A small network today may tell us little about deployment after regulatory approval and reliability cross the required threshold.

Energy may become Tesla’s second major platform. The backlog, deployment growth and factory expansion could produce far larger cash flows than the present revenue mix suggests.

Tesla may also retain a premium valuation for far longer than conventional analysis expects. Markets are not obliged to give patient investors a tidy moment of convergence.

Finally, current capital expenditure may prove to be genuinely high-return growth investment. The bearish thesis weakens considerably if Tesla shows that today’s spending produces attractive incremental returns rather than merely defending its competitive position.

These are serious possibilities.

They are also why I would not classify Tesla as a value trap.

A value trap looks cheap while its economics deteriorate.

Tesla presents the opposite problem: potentially exceptional assets offered at a price that already assumes exceptional execution.

Final classification: Too expensive now

Tesla is not merely a car company.

Its installed base, FSD adoption, energy operation, charging network and manufacturing capabilities deserve meaningful value beyond conventional automotive earnings.

But option value is not infinite value.

At roughly US$1.20 trillion of enterprise value, investors are paying today for several future businesses whose margins, capital requirements and commercial timelines remain insufficiently clear.

Q2 strengthened the evidence that Tesla can grow.

It weakened the evidence that this growth currently produces attractive returns on capital.

The decisive question is therefore not whether Tesla has an extraordinary future.

It is whether that future can become large enough, profitable enough and soon enough to justify the price already attached to it.

For a value investor, the burden of proof remains with the valuation.

Primary sources

· Tesla Q2 2026 shareholder deck

· Tesla Q2 2026 Form 10-Q

· Tesla Q2 production, deliveries and deployments

· Tesla Investor Relations

This article is for research and educational purposes. It is not personal investment advice.

If you found value in this article, please hold down the like ❤️ button or consider buying me a coffee ☕️ to fuel the next one!

Buy me coffee ☕️

Subscribe now

The Long Game ♟️ Perspective: An occasional Wealth Strategy series:

For the past few months, I have been closely following the capital expenditure of the major AI hyperscalers.

Microsoft, Alphabet, Amazon and Meta are committing extraordinary sums to data centres, advanced chips, power infrastructure and computing capacity.

The figures are enormous.

But as investors, we do not look only at how much these companies are spending.

We ask harder questions.

How is the expenditure being financed?

Can operating cash flow support it?

What return might the investment eventually generate?

What happens if AI demand takes longer than expected to materialise?

How much financial flexibility remains if management’s assumptions prove too optimistic?

While thinking about these questions, I had a personal-finance epiphany.

Most households will never build a data centre.

But many of us will make one enormous capital-allocation decision of our own.

We buy a home.

For most families, a property will be the largest asset purchased, the largest debt assumed and the longest financial commitment undertaken.

Yet we often analyse it with less discipline than an investor would apply to a company’s quarterly capital expenditure.

We consider the location.

We admire the layout.

We calculate the downpayment.

We ask how much the bank will lend us.

Then we look at the monthly instalment and conclude that the property is affordable.

A value investor would never assess corporate debt that way.

No serious investor would say:

“The bank agreed to lend the company money, so the debt must be safe.”

Why should a household be any less rigorous?

A home is not literally corporate CapEx

Before property professionals sharpen their keyboards, an important distinction is necessary.

An owner-occupied home is not corporate capital expenditure in the strict accounting sense.

A company invests in a factory, data centre or piece of machinery because it expects that asset to increase revenue, reduce costs or improve productivity.

Your home may not produce income.

It provides shelter, privacy, stability, convenience and emotional value. These are real benefits, but they are not corporate earnings.

The more accurate comparison is this:

A home is often the household’s largest capital-allocation decision, while the mortgage is the financing structure attached to it.

That distinction matters.

We should not pretend that a family is a listed company.

But we can borrow the discipline behind corporate debt analysis:

• separate dependable income from optimistic income

• measure debt against the cash flow available to service it

• distinguish asset value from liquidity

• recognise that assets require maintenance and replacement

• stress-test the borrower under difficult conditions

• consider what other opportunities are sacrificed when capital is committed

What we learn from corporate finance is learnable and applicable to personal finance.

The comparison simply needs to be made like for like.

Start with household operating surplus, not gross income

When analysing a company, EBITDA is not the same as revenue.

Likewise, gross household income should not automatically be treated as money available for a mortgage.

A more useful starting point is what we might call household operating surplus:

Recurring spendable household resources
minus essential non-housing expenses
minus recurring property operating costs

Recurring household resources might include:

• dependable take-home income

• conservatively assessed business income

• sustainable rental income

• housing contributions genuinely available from retirement accounts

Essential expenses might include:

• food

• transport

• insurance

• childcare

• parental support

• essential healthcare

• recurring education costs

Property operating costs might include:

• maintenance fees

• property tax

• home insurance

• routine repairs

This produces a far more realistic picture than simply dividing the mortgage instalment by gross salary.

Gross income tells you how much money enters the household.

Operating surplus tells you how much financial capacity remains after life has taken its share.

That is the number the mortgage must ultimately compete for.

EBITDA interest coverage: the generous view

For a company:

EBITDA interest coverage = EBITDA ÷ Interest expense

The closest household equivalent might be:

Household operating surplus before major replacement provisions ÷ Mortgage interest

Suppose a household has S$8,000 remaining each month after essential expenses and recurring property costs.

The interest portion of the mortgage is S$2,000.

Its household EBITDA-style interest coverage would be:

S$8,000 ÷ S$2,000 = 4 times

That sounds comfortable.

But it is also the most forgiving interpretation.

Like corporate EBITDA, it ignores the cost of assets eventually wearing out.

Renovations age.

Air-conditioning systems fail.

Appliances need replacing.

Vehicles deteriorate.

Medical and family commitments often rise over time.

EBITDA is useful because it answers a narrow question.

It becomes dangerous only when a narrow answer is mistaken for the entire picture.

The same applies to household affordability.

EBIT interest coverage: recognising that things deteriorate

Corporate EBIT deducts depreciation and amortisation.

Households do not prepare depreciation schedules for refrigerators, but the economic reality still exists.

Things wear out.

A prudent household should therefore reserve money for predictable maintenance and replacement costs.

An EBIT-style household ratio might be:

Household operating surplus after replacement provisions ÷ Mortgage interest

This is not literally accounting EBIT.

It performs a similar analytical function.

It stops us pretending that a home and the assets inside it can be maintained indefinitely without further capital.

A family that can service its mortgage only because it saves nothing for repairs, healthcare, retirement or future replacement is not as financially strong as the monthly instalment suggests.

The mortgage may be current.

The household balance sheet may still be deteriorating.

Cash interest coverage: check the bank account, not the budget

A company can report accounting profits while generating weak cash flow.

Households can do the same.

On paper, a family may believe it should save S$5,000 every month.

Its bank account may show that only S$1,500 actually remains.

The second number is the one that matters.

A household cash interest coverage ratio might therefore be:

Actual recurring cash surplus before debt service ÷ Actual mortgage interest paid

Not an expected promotion.

Not an uncertain bonus.

Not projected investment gains.

Not the amount the family intends to save after becoming more disciplined next month.

Actual recurring cash.

Debt does not care about financial intentions.

It is serviced with money that arrives.

The most important household ratio is not interest coverage

This is where the corporate analogy must be adapted carefully.

Interest coverage tells us whether income or cash flow can cover interest expense.

But most residential mortgages require the household to repay both interest and principal.

Principal repayment is not an economic expense in the same way as interest.

Interest is a financing cost.

Principal repayment reduces debt and builds home equity.

That is a genuine financial benefit.

But principal remains a compulsory cash outflow.

A homeowner cannot tell the bank:

“This portion is technically balance-sheet accumulation, so I have decided not to pay it this month.”

For households, the more important measure is therefore debt-service coverage.

Corporate lenders often examine:

Cash available for debt service ÷ Principal and interest payments

The household equivalent would be:

Household operating surplus after prudent provisions ÷ Mortgage principal and interest

Return to the earlier example.

The household has S$8,000 of monthly operating surplus.

Its mortgage interest is S$2,000.

The interest coverage ratio is four times.

But assume the total instalment, including principal, is S$5,000.

The household debt-service coverage becomes:

S$8,000 ÷ S$5,000 = 1.6 times

Still manageable.

But much less comfortable than the four-times interest-coverage figure initially suggested.

Now suppose dependable household income falls and the available surplus declines to S$6,400.

The ratio falls to:

S$6,400 ÷ S$5,000 = 1.28 times

The mortgage can still be serviced, but the margin for error has narrowed sharply.

If available surplus falls below S$5,000, the family must:

• reduce other spending

• stop saving or investing

• draw down reserves

• sell assets

• refinance

• or eventually miss payments

Interest coverage measures sensitivity to financing costs.

Debt-service coverage measures whether the household can actually carry the mortgage.

That is the like-for-like comparison.

“But property appreciates over time”

It may.

A well-selected property can appreciate significantly.

Leverage can enhance the homeowner’s return on equity.

Inflation may also reduce the real burden of fixed debt as salaries rise.

These are valid arguments.

But appreciation and debt coverage answer two different questions.

Property appreciation asks:

What might the asset eventually be worth?

Debt coverage asks:

Can the household meet its obligations while waiting?

An appreciating property cannot directly pay next month’s instalment unless it is sold, refinanced or monetised.

Those options take time, involve costs and may become less attractive precisely when the owner is under financial stress.

A favourable long-term investment thesis does not eliminate short-term liquidity risk.

Companies can own valuable assets and still fail because they run out of cash.

Households are not exempt from that reality.

“But our salaries will increase”

They may.

A fixed mortgage can become easier to service as incomes rise and inflation reduces the real value of the debt.

But prudent underwriting distinguishes between:

• income already received

• income reasonably dependable

• income merely expected

A future salary increment should improve future affordability.

It should not be required to justify today’s purchase.

There is nothing wrong with optimism.

There is something wrong with constructing a compulsory repayment schedule around it.

“But the bank has already tested affordability”

The bank has assessed whether the loan represents an acceptable credit risk.

That is not the same as deciding whether the mortgage supports your wider life.

The lender does not necessarily know:

• when you want to retire

• whether one spouse may stop working

• whether you intend to fund overseas education

• whether you support ageing parents

• whether your income is volatile

• whether you plan to start a business

• how much liquidity gives you peace of mind

• what opportunities may emerge in the future

The bank asks:

Is this an acceptable loan?

The household should ask:

After servicing this loan, can we still build the life we want?

The maximum mortgage available is the bank’s risk limit.

It is not necessarily your ideal budget.

“But rent is wasted money”

Rent pays for the use of a home.

Mortgage interest pays for the use of borrowed capital.

Property tax, maintenance, insurance, transaction costs and renovation depreciation do not build equity either.

Principal repayment does build equity.

So the honest comparison is not simply rent versus mortgage instalment.

It is rent and flexibility versus interest, ownership costs, principal accumulation, leverage and exposure to future property values.

Buying may be the better decision.

Renting may also be rational when flexibility matters, transaction costs are high or the intended holding period is uncertain.

“Rent is wasted money” is not analysis.

It is a slogan that happens to close transactions.

The real risk is not always foreclosure

When people discuss mortgage risk, they tend to imagine default or forced selling.

Those are serious risks, but they may not be the most common consequence of carrying too much debt.

The quieter cost is the loss of optionality.

You remain in a job because the mortgage is large.

You avoid starting a business because income uncertainty becomes intolerable.

You delay retirement.

You stop investing whenever family expenses rise.

You keep too much of your net worth concentrated in one illiquid asset.

You become conservative everywhere else because the house has already consumed most of your risk capacity.

The household survives.

But its choices shrink.

This is similar to a highly leveraged company that remains operational but no longer has the capacity to invest, innovate or withstand a downturn.

Debt does not only claim cash flow. It can also claim future choices.

That may be the most important household debt ratio of all.

Not because it appears neatly on a spreadsheet.

But because it tells you how much of your future has already been committed.

Your home is competing for capital

This may be the most important lesson I took from studying AI capital expenditure.

Every dollar committed to a data centre cannot simultaneously be used for acquisitions, dividends, debt reduction or alternative investments.

The same is true for households.

Every additional dollar committed to a more expensive home cannot simultaneously fund:

• retirement

• education

• entrepreneurship

• investment portfolios

• emergency liquidity

• family support

• future opportunities

This does not mean the money should never be spent on the home.

It means the opportunity cost should be acknowledged.

Capital allocation is rarely about choosing between something good and something bad.

It is usually about choosing between several good things while pretending that no trade-off exists.

A larger home may be desirable.

A stronger investment portfolio may also be desirable.

Earlier retirement may be desirable.

Greater liquidity may be desirable.

More freedom to change careers may be desirable.

The property decision decides, quietly, how much remains available for everything else.

A home should provide security, not consume it

This is not an argument against property.

It is not an argument against mortgages or responsible leverage.

A sensibly structured mortgage can preserve liquidity, finance a valuable long-term asset and provide genuine family stability.

The lesson from corporate finance is not that debt is bad.

The lesson is that debt should be supported by dependable cash flow, sufficient coverage and a margin of safety.

When analysing AI hyperscaler spending, investors do not ask only whether the company can afford the expenditure today.

They ask what returns it might produce, how it affects free cash flow and what happens if the anticipated future arrives later than expected.

We should ask equally demanding questions about our largest personal capital commitment.

Not merely:

Can we pay the mortgage instalment today?

But:

Can we service the full mortgage, preserve liquidity, continue investing and survive a difficult period without surrendering our future choices?

The bank calculates how much it is prepared to lend.

The property agent helps you decide what you might purchase.

Only you can decide how much of your future capital and freedom you are prepared to commit.

That may be the most important capital-allocation decision you ever make.

How do you assess mortgage affordability: by the instalment you can pay today, or by the choices that remain after paying it?

If you found value in this article, please hold down the like ❤️ button or consider buying me a coffee ☕️ to fuel the next one!

Buy me coffee ☕️

Subscribe now

The Long Game ♟️ Perspective: An occasional Wealth Mindset series:

When people suffer a major setback, they usually make the same mistake.

They try to recover everything at once.

An investor who has lost a large amount starts looking for the one stock that could make it all back.

A business owner who has missed several months of sales begins chasing an enormous contract.

A professional who feels behind decides to completely redesign their career, routine and life by Monday morning.

The size of the response begins to mirror the size of the pain.

It feels ambitious. It can even look courageous.

But much of the time, it is simply panic wearing a suit.

A setback changes more than your results

A significant loss does not only reduce your money, revenue or progress.

It also changes the way you see the next decision.

When confidence is healthy, we notice possibilities. We are patient enough to wait for the right opportunity. We can accept uncertainty without treating every unknown as a threat.

When confidence collapses, the same environment looks completely different.

Opportunities become traps.

Temporary setbacks feel permanent.

Every decision carries the emotional weight of the previous failure.

This is why recovery is rarely just a mathematical problem.

It is also a psychological one.

The investor is not merely trying to recover the money. He is trying to recover the feeling that he knows what he is doing.

The business owner is not only trying to restore revenue. She is trying to prove that the business still works.

The professional is not simply trying to improve. He is trying to escape the fear that he may have fallen permanently behind.

Under those conditions, asking someone to take a huge leap is usually the wrong advice.

The bigger the setback, the smaller the next target should become.

Small does not mean unambitious

A smaller target is not the final destination.

It is the first stable step back towards good judgement.

After a major setback, the immediate objective should not be to recover everything. It should be to rebuild the ability to make sound decisions without desperation.

That may mean:

• reducing position sizes after an investment loss;

• focusing on one realistic client opportunity rather than ten speculative ones;

• completing one meaningful task instead of designing an elaborate productivity system;

• rebuilding one reliable habit before attempting a total personal transformation.

The goal is not to stay small.

The goal is to create evidence that you can move forward again.

Confidence is strengthened less by encouragement than by experience. We begin to trust ourselves when we repeatedly see ourselves doing the right things.

A small, well-earned success provides that evidence.

One success becomes two.

Two become a pattern.

The pattern eventually becomes momentum.

The four-step recovery process

A useful way to think about recovery is:

1. Stabilise

Stop making decisions whose main purpose is to erase the emotional discomfort of the setback.

Do not add risk merely because you feel behind.

Do not chase a deal simply because you need a win.

Do not make a permanent decision while responding to a temporary loss of confidence.

Create enough breathing room to think clearly.

2. Shrink the target

Choose the smallest meaningful objective that remains within your control.

Not “recover my entire portfolio”.

Instead: follow my investment process correctly on the next decision.

Not “save the business this month”.

Instead: have five serious conversations with qualified prospects.

Not “become disciplined”.

Instead: complete the important task before opening social media tomorrow morning.

The target should be small enough to execute, but meaningful enough to rebuild trust in yourself.

3. Stack evidence

Do not dismiss small progress because it looks unimpressive.

Record it.

Study what worked.

Repeat it.

People often celebrate dramatic outcomes but ignore repeatable behaviours. Yet sustainable recovery is built from behaviours long before it appears in the results.

The question is not merely, “Did I win?”

4. Scale gradually

Once judgement and confidence begin returning, increase the size of the challenge.

Take more responsibility.

Deploy more capital.

Pursue the larger opportunity.

But let the scale of your actions grow from evidence, not impatience.

The purpose of the early small wins is to make the later larger moves rational again.

The investment lesson

This principle matters particularly in investing.

After a large loss, investors often feel compelled to recover quickly. They increase position sizes, shorten their time horizon or move into investments they do not properly understand.

They stop asking, “Is this a good investment?”

They start asking, “Can this make back what I lost?”

Those are entirely different questions.

The second question allows the past to control the next decision.

A more disciplined recovery begins by separating the previous loss from the next opportunity.

The market does not know what price you paid.

It does not care how much you need to recover.

Your next investment should stand on its own merits.

This does not mean blindly holding or adding to a losing position. A broken investment thesis should be acknowledged. Small wins are not an excuse for denial.

It means returning to position-size discipline, independent thinking and a process that can survive being wrong.

The objective is not to feel fearless.

It is to prevent fear from becoming the portfolio manager.

Recovery is usually boring

The way back is rarely dramatic.

That is precisely why the ego dislikes it.

The ego wants the heroic comeback, the single breakthrough and the decision that proves everyone wrong.

Real recovery often looks much less impressive.

One sensible investment.

One productive meeting.

One completed task.

One promise to yourself that you actually keep.

But there is enormous power in becoming reliable to yourself again.

A person who can repeatedly take the next sound step is eventually capable of travelling a very long distance.

When the setback is large, do not ask:

“How can I recover everything immediately?”

Ask:

“What is the smallest meaningful win that would help me trust my judgement again?”

The comeback does not begin when the entire loss has been recovered.

It begins with the first decision that is no longer controlled by the loss.

If you found value in this article, please hold down the like ❤️ button or consider buying me a coffee ☕️ to fuel the next one!

Buy me coffee ☕️

Subscribe now

The Long Game ♟️ Perspective: An occasional Alpha series:

There is a lazy way to analyse Copart.

Revenue growth has slowed.

The share price has fallen.

The chief executive is being replaced.

Therefore, something must be wrong with the business.

I think that conclusion is too convenient.

Copart’s latest numbers do reveal a genuine slowdown. They also reveal something more interesting: the company’s US business has stagnated while its international operations continue to expand, margins remain unusually strong, and the balance sheet is overflowing with liquidity.

This is not a broken business.

It may be a maturing one.

That distinction matters because the market is no longer asking investors to pay an extravagant price for Copart’s historical quality.

At roughly US$27.52 per share at the time of writing, Copart trades at around 17 times my calculation of trailing diluted earnings. That is a dramatically less demanding starting point than the valuation investors were previously asked to accept.

The central question is therefore not:

Is Copart still a good company?

It clearly is.

The real question is:

Is Copart still a compounder, or has it become a mature cash machine whose historical valuation no longer applies?

My argument is that the market may be confusing slower growth with a weaker moat.

But investors should resist the opposite mistake too.

A strong moat does not automatically guarantee attractive future returns.

The price paid still matters.

The Investment Thesis

Copart operates a global online marketplace for damaged, salvage, wholesale and used vehicles.

But describing it as an online auction company misses the point.

Copart’s real business is the ownership and coordination of an unusually difficult physical and digital network:

• storage yards near major population centres

• towing and vehicle-processing infrastructure

• title-processing expertise

• relationships with insurance carriers

• a global buyer network

• online auction technology

• data accumulated from millions of transactions

The company connects approximately one million members across more than 185 countries. It operates at more than 250 locations in 11 countries and sold more than four million vehicles in the most recent year.

The business looks industrial.

The economics look closer to a marketplace.

Copart does not need to manufacture vehicles.

It does not need to own most of the vehicles it sells.

Its core service revenues come from transaction fees, transportation, storage, title processing, vehicle loading and other auction-related services.

This produces an unusual combination:

physical scarcity plus digital scalability.

Most online marketplaces can be copied with enough software and marketing expenditure.

Copart cannot.

A competitor would need land.

It would need the correct zoning.

It would need transport capacity.

It would need insurance relationships.

It would need title-processing expertise.

It would need enough buyers to produce competitive auction prices.

And it would need enough vehicles to keep those buyers interested.

That circular relationship is the moat.

More sellers attract more buyers.

More buyers create better auction liquidity.

Better auction liquidity produces higher returns for sellers.

Higher seller returns attract more volume.

The loop strengthens itself.

The Moat Is Not The Website

Copart’s website is visible.

Its moat is not.

The real advantage sits beneath the screen.

Copart’s 2025 annual report explicitly identifies zoning requirements as a barrier to acquiring and developing new storage facilities. It also describes title-processing expertise as a competitive advantage.

That matters because salvage auctions are not purely digital.

A damaged vehicle has to be collected.

It has to be stored.

It has to be inspected.

Its title has to be processed.

It has to be presented to buyers.

It may need to be loaded for domestic transport or export.

A new entrant cannot simply build an app and declare itself a competitor.

It must replicate a regulated logistics network.

Copart’s buyer network then adds a second layer of protection.

International buyers purchase approximately 40% of the vehicles sold through Copart’s US auctions and account for nearly half of gross transaction value. International bidders participate in more than 90% of Copart auctions involving vehicles that are drivable or expected to be rebuilt.

This is economically important.

The winning bidder sets the sale price.

But the losing bidders create it.

Auction liquidity is valuable because every additional credible participant increases the probability that the seller receives the vehicle’s highest available value.

That is why insurance companies care about Copart’s buyer network.

They are not simply outsourcing disposal.

They are trying to maximise salvage recovery.

The company is therefore embedded in a mission-critical part of an insurer’s claims process.

Copart’s Title Express platform now processes well over one million titles annually. The company is also expanding into earlier total-loss decisioning and other workflow tools used by insurance carriers.

The deeper Copart moves into insurer workflows, the less useful it becomes to describe it merely as an auction site.

It is becoming part marketplace, part logistics network, and part claims infrastructure.

That is a formidable business.

But formidable businesses can still stop growing.

The Uncomfortable Numbers

Copart’s third-quarter fiscal 2026 results were not disastrous.

They were simply uninspiring.

For the quarter ended 30 April 2026:

• revenue rose 2.1% to US$1.237 billion

• gross profit rose 3.7% to US$572.6 million

• net income attributable to Copart fell 1.0% to US$402.4 million

• diluted earnings per share rose 2.4% to US$0.43

For the first nine months of fiscal 2026, revenue declined 0.2%, gross profit rose 0.8%, net income was almost flat and diluted EPS increased 1.7%.

This chart captures the central tension.

The top line has stalled.

Profitability has not collapsed.

That tells us the market is not dealing with a business in free fall.

It is dealing with a business whose growth engine has temporarily lost momentum.

The distinction matters.

If Copart’s revenues were falling while costs rose and margins collapsed, the moat would deserve serious doubt.

That is not what the latest results show.

Gross profit grew faster than revenue.

Cost of vehicle sales fell 5.6% in the quarter despite vehicle sales revenue increasing 2.3%.

Copart is still operating efficiently.

The question is whether efficiency can substitute for growth indefinitely.

It cannot.

Cost discipline can protect earnings.

It cannot manufacture demand.

Read more

Subscribe now

The Long Game ♟️ Perspective: An occasional macro series:

Central thesis

Artificial intelligence has crossed from being a technology-sector theme into a macroeconomic transmission mechanism, supporting exports, tax revenues, electricity investment and national growth. The principal risk is therefore no longer merely overvaluation, but the world economy’s growing dependence on an unusually concentrated capital-spending cycle.

Five companies are expected to spend more than US$1 trillion on AI-related capital expenditure across 2025 and 2026.

During the same week that this figure began to look less like a technology statistic and more like a macroeconomic one, China reported a 27 per cent rise in June exports and South Korea raised its 2026 growth forecast from 2 per cent to 3 per cent, largely because of the semiconductor boom.

The International Monetary Fund now says AI-driven demand is helping to offset the economic damage caused by war and energy disruption.

This changes the central question surrounding artificial intelligence.

The usual debate asks whether AI is a genuine technological revolution or an enormous financial bubble.

That may be the wrong distinction.

AI can be transformative and overbuilt. Its infrastructure can become socially valuable while producing disappointing returns for many of the companies financing it. The technology can succeed even as parts of the investment cycle fail.

The more urgent question is this:

How much of the global economy’s apparent resilience now depends on a handful of companies continuing to spend at extraordinary rates?

The world may be acquiring what we might call an AI put.

Like a financial put option, it appears to place a floor beneath economic weakness. When consumer demand softens, property markets struggle, or geopolitical shocks interrupt trade, AI infrastructure spending continues to support factories, utilities, construction companies, semiconductor exporters and government tax revenues.

But unlike a central-bank backstop, the AI put is private, concentrated and revocable.

It exists only for as long as several corporate boards believe the next hundred billion dollars will earn an acceptable return.

The boom has escaped the screen

AI is still discussed as though it were primarily a software phenomenon: chatbots, coding assistants, autonomous agents and increasingly capable models.

Economically, however, the AI expansion looks more like an industrial mobilisation.

The International Energy Agency estimates that five large technology companies spent more than US$400 billion on capital expenditure in 2025. It expects their expenditure to rise by another 75 per cent in 2026.

Data-centre electricity consumption increased by 17 per cent in 2025, compared with approximately 3 per cent growth in total global electricity demand.

Money spent on AI does not remain inside Silicon Valley.

It travels through a physical chain:

Hyperscalers purchase processors and memory chips.

Chipmakers order fabrication equipment and build factories.

Data-centre developers buy land, transformers, cooling systems, gas turbines and backup batteries.

Utilities expand generation and transmission.

Construction companies hire workers.

Suppliers earn profits.

Governments collect taxes.

Some of those governments then recycle the revenue into further AI infrastructure.

This is the AI capital-spending multiplier.

The immediate economic impact does not depend upon AI already transforming every office or producing enormous end-user revenue. The act of building the infrastructure creates demand before the infrastructure proves its ultimate value.

That is why the IMF can project global growth of 3 per cent in 2026 and 3.4 per cent in 2027 despite war-related energy shocks. The expansion is highly uneven, but countries integrated into the technology supply chain are receiving a powerful external lift.

The AI boom is no longer merely anticipating future productivity.

It is producing present-day GDP.

China’s export cushion hides domestic weakness

China offers the clearest example of how this mechanism works.

Its exports increased by 27 per cent year on year in June, exceeding economists’ expectations of 18.2 per cent. Imports rose 36 per cent, their fastest pace in five years, while the monthly trade surplus reached US$125.6 billion.

China exported more than 32 billion integrated circuits during the month. Imports from South Korea rose 85 per cent, and imports from Taiwan increased 41.1 per cent, partly because semiconductor prices and AI-related demand were rising rapidly.

This was not solely an AI story. Automobiles and other manufactured products also contributed.

But the broader mechanism is revealing.

China’s domestic economy remains constrained by weak property activity, subdued investment and cautious household spending. Foreign demand for technology hardware is compensating for weakness at home.

According to data cited by Reuters, exports reached 24 per cent of total Chinese manufacturing sales during the first four months of 2026. That was up from 18.3 per cent in 2019 and represented the highest proportion since China entered the World Trade Organisation in 2001.

This creates a peculiar form of stability.

AI investment allows Chinese factories to keep producing despite weak domestic demand. Yet the more production China sends abroad, the greater the risk of tariffs, quotas and political retaliation.

If the AI infrastructure boom weakens, China loses part of its external cushion.

If it continues at full force, trade tensions could intensify.

The same force supporting growth is also exporting industrial pressure into other economies.

This is the first sign that AI has become more than a technology cycle. It is beginning to reshape trade balances, industrial policy and international relations.

South Korea’s semiconductor feedback loop

South Korea provides an even cleaner illustration.

On 14 July, the government raised its 2026 growth forecast from 2 per cent to 3 per cent, the strongest expected expansion since 2021 and well above the 1.1 per cent growth recorded last year.

The government attributed much of the improvement to semiconductor exports and announced plans to accelerate major projects involving chips, AI data centres and physical AI.

It is also planning to increase 2027 government spending by at least 10 per cent, taking the budget above 800 trillion won. Stronger tax receipts from semiconductor companies are helping to finance this expansion.

Consider the feedback loop:

1. American technology companies increase AI capital expenditure.

2. Demand rises for high-bandwidth memory and advanced chips.

3. South Korean semiconductor companies generate higher sales and profits.

4. The government receives more tax revenue.

5. Higher revenue supports public investment in AI infrastructure.

6. That investment strengthens domestic demand and potential growth.

Private AI expenditure in one country becomes fiscal capacity in another.

SK Hynix’s July listing in the United States demonstrated the scale of financial enthusiasm surrounding this chain. Its shares opened 14 per cent above the offer price following a US$26.5 billion share sale that was reportedly more than seven times oversubscribed. Its Korean shares were still approximately 630 per cent higher than one year earlier, despite having fallen from their recent peak.

The important point is not whether those valuations are justified.

It is that semiconductor profits, equity-market confidence, government revenue and national growth expectations are increasingly moving together.

AI has become part of South Korea’s business cycle.

The market has quietly priced an “AI put”

Investors appear to understand that AI spending is supporting economic activity.

They may not fully appreciate how dependent their optimism has become on its continuation.

Bank of America’s July fund-manager survey found that investor cash allocations had fallen from 4.1 per cent to just 3.6 per cent, a level low enough to trigger the bank’s contrarian sell signal.

A record 54 per cent of respondents expected a global economic scenario with no recession and no meaningful slowdown.

At the same time:

• 82 per cent described global semiconductor shares as the market’s most crowded trade.

• 61 per cent believed hyperscalers were unlikely to reduce capital expenditure during 2026.

• 45 per cent identified an AI bubble as the largest tail risk facing markets.

This is a remarkable contradiction.

Investors believe AI is both the greatest risk and the reason the economy will avoid recession.

They fear the boom, but rely upon it.

That is the psychological foundation of the AI put.

As long as hyperscaler spending remains strong, earnings flow through the semiconductor supply chain, economic forecasts remain resilient, and markets can dismiss other weaknesses as temporary.

But this confidence makes the eventual consequences of a slowdown more significant.

A normal corporate capital-expenditure cycle affects a company and its suppliers.

This one increasingly affects national exports, fiscal revenue, power generation, construction activity, credit markets and global equity valuations.

A genuine technology can still produce poor investments

The history of technological progress is filled with innovations that changed society while disappointing the people who originally financed them.

The telecommunications boom of the late 1990s is an especially useful comparison.

Annual US investment in communications equipment increased from approximately US$62 billion in early 1996 to more than US$135 billion by late 2000.

By the fourth quarter of 2001, it had fallen below US$93 billion.

The Nasdaq index of telecommunications shares rose from 198 in April 1997 to 1,230 in March 2000. By May 2003, it had fallen to 136.

Yet telecommunications technology was not fraudulent.

Internet usage continued expanding. Fibre-optic networks remained useful. Digital connectivity eventually became fundamental to modern commerce.

The problem was that capital had been invested faster than profitable demand developed.

The technology succeeded.

Many owners of the capital did not.

The OECD estimated that information and communications technology contributed between 0.3 and 0.9 percentage points a year to economic growth in several countries during the second half of the 1990s.

That was a genuine economic contribution, even while telecom investors suffered enormous losses.

This distinction is essential for long-term investors:

Technological importance does not guarantee attractive investment returns.

Returns depend on the price paid, the intensity of competition, the durability of margins, the useful life of assets and who ultimately captures the productivity benefits.

Consumers may capture them through cheaper services.

Employees may capture them through higher productivity.

New businesses may capture them through lower operating costs.

The companies that paid for the infrastructure may discover that competition has transferred much of the value elsewhere.

The Bank for International Settlements has warned that hyperscaler investment is already outpacing the earnings and free cash flow of the companies involved, prompting some firms to raise additional debt.

It argues that competition for technological dominance may encourage every participant to overinvest. Even when the industry creates substantial gross value, total returns after investment costs could disappoint.

A business can be right about the future and still spend too much reaching it.

The hidden variable is physical capacity

The most popular AI discussions focus on model intelligence.

The more immediate constraints may be electricity, transformers, grid connections and financing.

The IEA expects data-centre electricity consumption to rise from approximately 485 terawatt-hours in 2025 to 950 terawatt-hours in 2030. Power demand from AI-focused facilities is projected to triple during that period.

The technology sector accounted for around 40 per cent of corporate renewable-power agreements signed in 2025.

Conditional agreements between data-centre operators and small modular nuclear-reactor projects have expanded from 25 gigawatts at the end of 2024 to approximately 45 gigawatts.

At the same time, shortages of transformers, gas turbines, advanced chips and grid capacity are delaying projects and encouraging developers to secure supply through long-term contracts.

This creates another feedback loop.

Scarcity encourages companies to reserve more capacity than they may eventually need.

Those reservations signal strong demand.

Suppliers expand production in response.

The additional supply may arrive after the original shortage has eased.

What initially looked like prudent capacity planning can become collective overbuilding.

This is a familiar feature of capital-intensive booms. Each company behaves rationally from its own perspective, but the industry produces an irrational aggregate result.

No hyperscaler wants to discover that a competitor secured all the available power, chips or data-centre land.

So everyone reserves more.

The shortage itself can plant the seeds of the future surplus.

The consequence cascade

The first-order effect of continuing AI expenditure is straightforward.

Chipmakers, data-centre operators, utilities and infrastructure companies benefit.

The second-order effects are broader.

Export economies receive stronger external demand. Governments collect higher corporate taxes. Power markets attract new investment. Construction activity rises. Equity wealth supports consumer confidence.

The third-order consequences are where the fragility appears.

Government budgets become more reliant on semiconductor profits.

Export forecasts become more dependent on hyperscaler orders.

Utilities build capacity around demand from a small number of customers.

Suppliers borrow against assumptions of continuing expansion.

Investors own what appear to be diversified assets that are actually exposed to the same underlying factor.

A portfolio containing American technology platforms, Asian semiconductor manufacturers, data-centre property companies, utilities, and industrial equipment suppliers may appear diversified across industries and geographies.

Economically, it may still be one large AI capital-expenditure position.

The BIS notes that US shares now represent approximately 64 per cent of the MSCI Global Index. A meaningful repricing of the largest technology companies could therefore transmit through global portfolios, household wealth and consumption.

This is how a sectoral slowdown could become a macroeconomic event.

A reduction in AI spending would not simply lower orders for processors.

It could simultaneously weaken equipment investment, Asian exports, electricity projects, construction contracts, tax collections, corporate credit and equity-market confidence.

The global economy is not necessarily there yet.

But it is moving in that direction.

The strongest argument against this thesis

There is a persuasive alternative interpretation.

Perhaps the capital spending is not excessive at all.

Perhaps companies are correctly anticipating an enormous expansion in demand as AI agents, robotics, scientific discovery and automated business processes become widespread.

At the task level, studies cited by the BIS have found time savings of between 20 and 50 per cent from AI tools. If those gains spread through entire organisations, the eventual economic benefit could justify today’s investment.

Computing costs may also decline faster than expected.

Cheaper intelligence could stimulate new applications that are currently impossible to predict, just as cheaper bandwidth enabled streaming, cloud computing and mobile commerce.

From this perspective, underinvesting could be more dangerous than overinvesting. A company that hesitates may permanently lose technological capability, data, talent and market position.

This argument deserves to be taken seriously.

The difficulty is that task-level efficiency does not automatically become company-level profit or economy-wide productivity.

The same BIS assessment notes that aggregate productivity estimates remain below 1 per cent over a long horizon, reflecting the difficulty of redesigning organisations and integrating new technology into real production systems.

The difference between those two numbers, 20 to 50 per cent task-level savings versus less than 1 per cent estimated aggregate productivity improvement, is the heart of the uncertainty.

It is possible that adoption will close the gap.

It is also possible that markets are capitalising the benefits years before businesses learn how to capture them.

What this means for long-term decision-makers

The conclusion is not that investors should abandon AI-related companies.

Nor is it that the expansion must end soon.

Calling the exact peak of an investment cycle is usually an efficient method of becoming confidently wrong in public.

The more useful approach is to separate three questions:

Is AI an important technology?

Almost certainly.

Will AI infrastructure spending continue growing indefinitely at its present rate?

Almost certainly not.

Which businesses will retain attractive returns after competition, depreciation and financing costs?

That remains unresolved.

Long-term investors should therefore look beyond revenue growth and order backlogs.

The more important indicators are whether operating cash flow is keeping pace with capital expenditure, whether debt is replacing internal financing, whether returns on invested capital are improving and whether AI demand is spreading beyond a small number of hyperscalers.

Business owners face a parallel challenge.

They should adopt AI where it produces measurable operational value, but should not confuse the amount vendors are spending with the amount customers are earning.

Infrastructure investment is evidence of conviction.

It is not proof of return.

For policymakers, the task is even harder.

Countries benefiting from semiconductor exports and data-centre investment should use the windfall to improve broader productivity, rather than allowing their economies to become permanently dependent on one volatile industry.

Temporary prosperity is most valuable when it finances permanent adaptability.

The economy is learning to need the boom

The strongest evidence that AI is not merely a speculative fiction is no longer found in demonstrations of clever models.

It is visible in China’s export figures, South Korea’s growth forecasts, electricity demand, construction plans and national budgets.

AI is doing real economic work.

That is precisely why the risks are growing.

A harmless bubble can burst without damaging much beyond the people who bought it.

A useful boom is different. Companies reorganise around it. Governments budget around it. Investors rely on it. Entire countries begin to treat its continuation as part of the economic baseline.

The most dangerous stage of a boom is not when it is obviously fictional.

It is when it becomes useful enough for everyone to depend on it.

The question is no longer whether AI will change the economy.

It already is.

The question is whether the economy can eventually stand without the extraordinary spending required to build it.

The Long Game ♟️ examines the forces reshaping business, markets and wealth before they become obvious. Subscribe for one evidence-backed essay each week.

Essential sources

• Bank for International Settlements, Annual Economic Report 2026: Progress and Peril.

• International Energy Agency, Key Questions on Energy and AI and April 2026 update.

• International Monetary Fund, World Economic Outlook Update, July 2026.

• Reuters, China’s June 2026 trade results.

• Reuters, South Korea’s revised 2026 growth strategy.

• Reuters, SK Hynix’s US listing and share sale.

• Reuters, Bank of America Global Fund Manager Survey, July 2026.

• Federal Reserve Bank of Richmond, Boom and Bust in Telecommunications.

• OECD, ICT Investment and Economic Growth in the 1990s.

If you found value in this article, please hold down the like ❤️ button or consider buying me a coffee ☕️ to fuel the next one!

Buy me coffee ☕️

Subscribe now

The Long Game ♟️ Perspective: An occasional Alpha series:

The lazy bear case on Alphabet was simple.

AI will kill Search.

That argument sounded persuasive because it was clean. Google was built on links, keywords and advertising. AI would answer questions directly. Fewer clicks. Less traffic. Lower ad revenue. The castle falls.

But investing is rarely that tidy.

Alphabet’s latest numbers do not look like the numbers of a company being quietly destroyed. Search is still growing. Google Cloud is accelerating. AI is being inserted into Search, Ads, YouTube, Cloud, Workspace, Gemini, Android, Chrome and infrastructure.

The better question is no longer:

“Will AI kill Google?”

The better question is:

“Can Alphabet spend aggressively on AI infrastructure and still preserve the quality of its free cash flow?”

That is a more serious debate.

It moves the discussion away from disruption theatre and into capital allocation.

And capital allocation is where the real investment work begins.

Executive Summary

Alphabet remains one of the most economically powerful businesses in the world.

In Q1 2026, Alphabet reported revenue of US$109.9 billion, up 22% year-on-year. Operating income rose 30% to US$39.7 billion, with operating margin expanding to 36.1%. Google Search & other revenue grew 19% to US$60.4 billion. Google Cloud revenue grew 63% to US$20.0 billion, while Cloud operating income increased to US$6.6 billion.

Those are not broken-business numbers.

But the investment case has changed.

Alphabet is no longer just a capital-light digital advertising compounder. It is becoming one of the world’s largest AI infrastructure companies.

Management now expects 2026 capital expenditure of US$180 billion to US$190 billion, largely for technical infrastructure. Alphabet has also highlighted a very large Google Cloud backlog, with more than half expected to convert into revenue over the next 24 months.

That is the tension.

The business quality is still high.

The reinvestment requirement is now much heavier.

The old Google was easier to analyse. It was an advertising toll road with extraordinary margins.

The new Alphabet is harder.

It is still that advertising toll road, but it is also building the roads, bridges, power stations and compute factories for the AI economy.

That could become a deeper moat.

It could also become a very expensive arms race.

The Market Debate Has Moved

The original AI bear case was that generative AI would weaken Google Search.

That concern was understandable.

Search is Alphabet’s crown jewel. It captures user intent close to the point of action. If users move from Google Search to AI assistants, Alphabet could lose query volume, ad inventory, publisher relationships and pricing power.

But the latest evidence complicates the bear case.

Search & other revenue grew 19% year on year in Q1 2026. Alphabet has also said AI experiences are increasing usage.

That does not prove Search is immune.

But it does show that the simple collapse narrative was too lazy.

The debate has moved from Search survival to Search monetisation density.

In plain English:

The question is not whether people still search.

They do.

The question is whether AI-powered Search can monetise longer, more complex queries without damaging margins.

That distinction matters.

Traditional Search monetises short, high-intent queries very well.

AI-powered Search may create new monetisation surfaces: deeper queries, product discovery, comparison, recommendations, shopping assistance and eventually agentic commerce.

But it may also cost more to serve.

So the Search equation becomes:

More useful answers plus better intent understanding minus higher compute cost.

If the first two outweigh the third, Alphabet strengthens.

If the third overwhelms the first two, margins weaken.

That is the real Search debate.

Search Is Still The Core Asset

Search remains Alphabet’s most important business because it captures intent.

A social media platform knows what entertains you.

A streaming platform knows what keeps you watching.

A marketplace knows what you bought.

Search often knows what you are about to do.

That is incredibly valuable.

A person searching for insurance, wealth planning, enterprise software, a hotel, a mortgage, a tax product or a business loan is often not merely consuming content.

They are moving towards a decision.

This is why Google’s advertising business has been so profitable.

The risk is not that intent disappears.

The risk is that the format of intent changes.

AI may reduce the need for users to click through to websites.

It may compress the traditional results page.

It may change publisher economics.

It may invite regulatory attention.

It may increase cost per query.

But it may also increase the number of useful queries Google can monetise.

If AI helps Google understand longer and more complex searches, even a modest improvement in Search monetisation can produce large profit dollars because Search is already so large.

That is why the bear case cannot simply say “Search will be disrupted”.

It must answer a harder question:

Will AI reduce the economic value of Google’s intent engine, or will it increase the monetisation density of that intent?

We do not know yet.

But the early numbers do not support a collapse narrative.

Google Cloud Is The Hardest Evidence In The Bull Case

Google Cloud used to be a strategic expense.

Now it is becoming a serious profit engine.

In Q1 2026, Google Cloud revenue increased 63% year on year to US$20.0 billion. Operating income increased to US$6.6 billion, giving the segment an operating margin of roughly 33%.

That matters.

If Alphabet were only defending Search, the AI story would be narrower.

But Cloud gives Alphabet a second payback path.

Enterprises need compute.

They need models.

They need infrastructure.

They need data tools.

They need AI deployment support.

They need security, storage, orchestration and integration.

Google Cloud is one of Alphabet’s main ways to sell into that demand.

This is important because AI demand is increasingly constrained by infrastructure.

If customers want more compute than the industry can supply, capex may be justified.

If demand later weakens, capex becomes dangerous.

This is why Cloud is the cleanest evidence to watch.

Search shows whether Alphabet can defend and enhance its old profit engine.

Cloud shows whether it can build the new one.

TPUs: The Underappreciated Layer

Investors often discuss AI infrastructure through Nvidia GPUs.

That is understandable.

But Alphabet has a different advantage: TPUs.

TPUs have historically helped Google optimise internal AI workloads. Now they are also part of Google Cloud’s external offering.

This matters because AI infrastructure is not just about demand.

It is about cost curves.

If two companies sell similar AI capabilities, the one with lower serving cost has more strategic flexibility.

It can price more aggressively.

It can protect margins.

It can scale more efficiently.

It can decide whether to rent, sell or internalise compute.

Alphabet’s full-stack position matters here.

It has models, custom silicon, data centres, cloud distribution, consumer surfaces, advertising tools and enterprise relationships.

Most companies have only parts of that chain.

Alphabet has more of the chain than almost anyone else.

That does not guarantee superior returns.

But it gives Alphabet a better chance of controlling the economics.

The Capex Problem

Now for the uncomfortable part.

Alphabet’s capital expenditure is enormous.

Management expects 2026 capex of US$180 billion to US$190 billion, compared with about US$91 billion in 2025 and roughly US$31 billion in 2022.

That is a massive change.

This is not a small adjustment in the business model.

It is a fundamental transformation.

Alphabet is moving from an asset-light software and advertising business into an infrastructure-heavy AI business.

High capex is not automatically bad.

For a truly high-return business, reinvestment is a gift.

The best compounders can absorb capital and earn attractive returns on it.

But high capex becomes dangerous when the returns are uncertain.

Alphabet generated US$45.8 billion of operating cash flow in Q1 2026 and US$10.1 billion of free cash flow after heavy property and equipment investment.

That is still huge.

But relative to Alphabet’s market capitalisation and capex ambitions, free cash flow is now the pressure point.

The old Google converted enormous revenue into cash.

The new Alphabet is converting enormous operating cash flow into infrastructure first, and free cash flow second.

That can still work.

But investors need evidence.

The Payback Framework

This is how I would frame Alphabet from here.

AI capex must pay back through measurable channels.

Not vibes.

Not conference demos.

Not headlines.

Cash flow.

Alphabet has several possible AI payback paths:

1. Higher Search monetisation

2. More complex commercial queries

3. Better ad targeting and conversion

4. Cloud backlog conversion

5. Cloud operating leverage

6. TPU monetisation

7. Workspace and Gemini subscriptions

8. Agentic commerce

9. Waymo and other long-dated optionality

But not all payback paths are equal.

Search and Cloud are core.

TPUs are important but still developing.

Agentic commerce is interesting, but still optionality.

Waymo may be valuable, but remains harder to value with precision.

That is why I would not build the investment case on the most exciting parts.

I would build it on the measurable parts.

Search monetisation.

Cloud revenue conversion.

Cloud operating margin.

Capex intensity.

Free cash flow.

The boring metrics will tell the truth before the exciting narratives do.

Valuation: Strong Business, High Expectations

Alphabet’s valuation is not irrational.

But it is not obviously cheap either.

The company still has several characteristics that deserve a premium:

High revenue growth.

Very strong operating margins.

A dominant Search franchise.

A fast-growing Cloud business.

Deep AI infrastructure.

A strong balance sheet.

Large recurring user behaviour across its ecosystem.

But free cash flow gives a more demanding picture.

If we use trailing free cash flow, Alphabet trades at a demanding free cash flow multiple because capex is now absorbing a much larger share of operating cash flow.

This creates the valuation fork.

If current capex is growth investment that produces high future operating income, then today’s free cash flow understates Alphabet’s long-term earning power.

If high capex is the permanent cost of competing in AI, then free cash flow quality may be lower than the market assumes.

That is the investment question.

Not whether Alphabet is a good company.

It clearly is.

The question is whether today’s price already assumes too much of the AI payback.

Wonderful businesses can still become demanding investments when expectations are high.

This is where long-term investors must be careful.

The market does not punish bad stories only.

It also punishes good stories that were already fully priced.

Bull Case

The bull case is straightforward.

Alphabet is not being disrupted by AI.

It is absorbing AI.

Search remains resilient.

Cloud is accelerating.

Cloud margins are expanding.

AI is improving advertising tools.

TPUs may reduce cost and create a new external revenue stream.

Gemini can be distributed across existing Google surfaces.

Cloud backlog gives visibility.

The company has a strong balance sheet and massive operating cash flow.

The strongest bull argument is that Alphabet owns several layers of the AI value chain.

It has consumer distribution through Search, Android, Chrome, YouTube, Gmail and Maps.

It has enterprise distribution through Google Cloud and Workspace.

It has infrastructure through data centres, Cloud regions and networking.

It has silicon through TPUs.

It has models through Gemini and related AI systems.

It has monetisation through advertising, subscriptions, enterprise contracts and potentially commerce.

Most AI companies are strong in one or two layers.

Alphabet is strong across many.

That is rare.

If AI increases the value of Search, accelerates Cloud demand, expands subscription revenue and improves advertising returns, Alphabet could remain one of the strongest compounders in the market.

The bull case is not that AI saves Google.

Google did not need saving.

The bull case is that AI expands the number of places where Google can monetise intent, compute and workflow.

Bear Case

The bear case is not that Alphabet disappears.

That is too dramatic.

The bear case is that Alphabet survives AI but becomes a lower-return business.

Search may continue to grow, but cost per query may rise.

AI answers may reduce click-through behaviour and create friction with publishers.

Cloud may grow quickly, but require ever-larger capex.

Depreciation may rise.

Power and data centre costs may rise.

AI compute pricing may compress as supply expands.

Competition from Microsoft, Amazon, Meta, OpenAI, Anthropic and other infrastructure players may pressure returns.

Regulatory scrutiny may remain heavy.

Capital markets may become less forgiving if free cash flow does not recover.

That is the risk.

Not death.

Dilution of economics.

A company can win the technology race and still disappoint investors if winning becomes too expensive.

That is the kind of risk long-term investors should respect.

Key Numbers To Watch

For Alphabet, I would watch five numbers.

First, Search revenue growth.

If Search continues growing while AI features expand, the disruption case weakens.

Second, Cloud revenue and Cloud operating margin.

This is the cleanest evidence of AI infrastructure demand converting into profit.

Third, capex as a percentage of revenue.

This tells us how far Alphabet is moving away from the old asset-light model.

Fourth, free cash flow margin.

Operating income can look healthy while cash conversion weakens.

Fifth, incremental return on invested capital.

This is the real test.

Not whether Alphabet can spend.

It clearly can.

The question is whether the spending earns enough.

Where I May Be Wrong

I may be too cautious.

It is possible that Alphabet’s capex cycle is exactly what great companies should do when the opportunity set is unusually large.

If Cloud backlog converts, Search monetisation improves, TPUs become a meaningful external revenue stream, and infrastructure utilisation remains high, today’s free cash flow pressure may look temporary in hindsight.

I may also be underestimating Waymo.

Alphabet said in Q1 2026 that Waymo surpassed 500,000 fully autonomous rides per week.

That is no longer a science project.

It is a real operating asset with long-term optionality.

But I may also be too generous.

AI could change user behaviour faster than current Search numbers suggest.

Cloud backlog may not convert at expected margins.

Depreciation may rise faster than revenue.

AI infrastructure could become a capital-intensive commodity layer.

The market may decide that Alphabet deserves a lower free cash flow multiple if the business becomes structurally heavier.

That is why this is not a simple “buy the dip” story.

It is a “watch the payback” story.

Final Classification

Attractive for Further Study

Alphabet is not a simple cheap stock.

It is also not an obvious value trap.

It is a high-quality business facing a genuine capital allocation test.

The early AI bear case was too simplistic.

Search has not collapsed.

Cloud is no longer a side project.

Alphabet has one of the deepest AI stacks in the world.

But the bullish case also needs discipline.

The market is already giving Alphabet credit for AI leadership, Cloud scale, infrastructure advantage and future monetisation.

The company must now prove that this enormous capex cycle converts into operating income, free cash flow and acceptable returns on capital.

My classification is therefore:

Attractive for Further Study, but not a clean margin-of-safety case at current expectations.

The investment question is no longer whether Alphabet is an AI winner.

The investment question is whether Alphabet is an AI winner at an attractive return on capital.

That is the difference between a great company and a great investment.

Capital has a memory.

Every dollar Alphabet spends on chips, data centres, power, networking and AI infrastructure will eventually ask the same question:

“What did I earn?”

That is the question worth following.

Not the headline.

Not the hype.

The return.

If you found value in this article, please hold down the like ❤️ button or consider buying me a coffee ☕️ to fuel the next one!

Buy me coffee ☕️

Subscribe now

The Long Game ♟️ Perspective: An occasional macro series:

The “free tokens” and “forward-deployed engineers” offered by frontier AI labs are usually described as partnerships.

That is the friendly version.

The value investing version asks a colder question:

Who captures the economics over time?

That question matters more than the demo. It matters more than the press release. It matters more than the phrase “co-innovation”, which increasingly sounds like something invented by a consultant who charges by the syllable.

From a value investor’s point of view, the central issue is not whether AI is useful. It clearly is. The issue is whether enterprises are using AI to strengthen their moats, or whether they are quietly handing their operating intelligence to the next platform monopoly.

A good business is not just one that earns high margins today. It is one that can defend those margins tomorrow. That defence usually comes from some combination of customer relationships, proprietary data, workflow knowledge, switching costs, distribution, brand, regulation, scale, and accumulated operational judgement.

These assets often do not sit neatly on a balance sheet. But they are frequently the real source of durable returns on capital.

A bank’s underwriting process is a moat.
A logistics company’s routing intelligence is a moat.
A manufacturer’s production data is a moat.
A law firm’s precedent library is a moat.
A software company’s user feedback loop is a moat.

AI does not merely automate these assets. It turns them into computable patterns.

That is why the ownership question matters.

To be fair, the crude version of this argument is wrong. It is not accurate to say that every enterprise prompt or dataset is automatically being used to train frontier models. OpenAI says that, by default, it does not train on inputs or outputs from ChatGPT Business, ChatGPT Enterprise, ChatGPT Edu, ChatGPT for Healthcare, ChatGPT for Teachers, or its API platform. Anthropic similarly says it does not use inputs or outputs from commercial products such as Claude for Work, the Anthropic API, and Claude Gov to train models by default. (OpenAI)

That matters. Precision matters.

But that does not eliminate the strategic risk.

The deeper risk is not only model training. The deeper risk is that the AI provider learns where the value is.

It sees which workflows enterprises are willing to pay to automate. It sees which use cases move from pilot to production. It sees the integration problems, the compliance frictions, the buying behaviour, the organisational bottlenecks, and the vertical-specific pain points. None of that has to be “training data” in the narrow technical sense to be commercially valuable.

This is not necessarily espionage in the cloak-and-dagger sense. It is something more ordinary and perhaps more dangerous: platform economics.

The platform sits close to the customer, watches where the profit pool forms, then moves up the stack.

The Old Playbook: Microsoft, Google, Amazon

The AI platform risk sounds new because the technology is new.

The economic pattern is not new at all.

Technology platforms rarely remain neutral infrastructure forever. Once a company controls the standard, interface, distribution layer, or default user behaviour, it gains the ability to tax, bundle, prioritise, or absorb the profit pools around it.

Microsoft is the classic case.

In the 1990s, Windows became the control point for personal computing. In the U.S. government’s Microsoft antitrust case, the Department of Justice argued that Microsoft’s operating systems had “well over 90%” share of the market for Intel-compatible PC operating systems. The same findings described the “applications barrier to entry”: users wanted Windows because the applications were there, and developers wrote for Windows because the users were there. That feedback loop protected Microsoft’s position. (justice.gov)

Once Windows controlled the operating environment, Microsoft could extend deeper into applications. Word and Excel did not win only because they were good products. They won because Microsoft understood bundling, file formats, enterprise standardisation, and distribution.

WordPerfect and Lotus 1-2-3 were once serious business-software leaders. But as the world moved to Windows and then to Microsoft Office, the centre of gravity shifted from standalone applications to the integrated productivity suite.

Decades later, the economics are visible in Microsoft’s financial statements. In fiscal 2025, Microsoft’s Productivity and Business Processes segment generated $120.8 billion in revenue and $69.8 billion in operating income. Microsoft 365 Commercial products and cloud services alone generated $87.8 billion, while Microsoft 365 Consumer reached 89 million subscribers. (Microsoft)

That is what durable standard-setting looks like when it matures into cash flow.

The lesson is not simply that Microsoft “copied” competitors. The more important lesson is that once a company owns the productivity layer, it can keep expanding its economics around the user.

The Microsoft Teams case shows the same logic in a more recent form. Teams benefited from distribution through Microsoft 365 and Office 365. The European Commission later accepted Microsoft’s commitments to address competition concerns around Teams being tied to Microsoft’s productivity suites, after complaints from rivals including Slack and alfaview. (Reuters)

Google tells a similar story in search.

Google began as a better search engine. Then search became the default gateway to the internet. As of June 2026, StatCounter showed Google with 91.27% global search-engine market share. (StatCounter Global Stats)

That dominance changed the economics of the internet. Publishers, advertisers, merchants, local businesses, comparison sites, travel platforms, and entire content industries had to optimise around Google’s rules. The search box became a toll road for attention.

Regulators recognised the same pattern. In 2024, the U.S. District Court concluded that Google was a monopolist and had acted as one to maintain its monopoly in general search. In Europe, the Commission fined Google €2.42 billion in 2017 for abusing its search dominance by giving an illegal advantage to its own comparison-shopping service. (justice.gov)

That Google Shopping case is directly relevant to AI. It shows what can happen when the owner of discovery also competes downstream. The platform does not need to block competitors outright. It can simply favour its own product, bundle more tightly, or change the rules of visibility.

Amazon’s marketplace offers another warning. The European Commission raised concerns over Amazon’s use of non-public marketplace seller data and later accepted commitments designed to prevent Amazon from using such seller data for its own retail operations. (European Commission)

Again, the lesson is not moral outrage. The lesson is economic.

When a platform sees enough transaction flow, it learns where the margin is.

AI could become the most powerful version of this playbook because this time the platform is not merely organising information, distributing software, or hosting sellers.

It is learning how work itself gets done.

AI Compresses the Platform Cycle

Historically, platforms took years to move from infrastructure to applications.

AI compresses that cycle.

At first, the AI lab sells access to a model. Then it provides engineering help. Then it observes which enterprise workflows are worth automating. Then it packages those workflows into repeatable products. Eventually, it may become a competitor to the application layer, the systems integrator, or even parts of the enterprise’s own operating model.

This is already happening.

AWS announced in June 2026 that it would invest $1 billion to create a Forward Deployed Engineering organisation, embedding thousands of experts with customers to co-develop and deploy agentic AI solutions. Reuters reported that these engineers would work directly with customer teams for 45-day periods and that demand for such roles grew 42-fold from 2023 to 2025. (Amazon News)

That is not a side project. That is a major strategic shift.

The question is not whether these engineers are helpful. They probably are. The question is what the vendor learns while helping.

A forward-deployed engineer sees the workflow from the inside. They see the messy bits that never appear in a software requirements document. They see where the process breaks, where the customer gets stuck, where the regulatory exception lives, where the human judgement still matters, and where the margin hides.

That is a valuable signal.

A value investor should immediately ask: who owns the learning loop?

If the enterprise owns it, AI strengthens the moat.
If the vendor owns it, AI may weaken the moat.
If nobody is sure who owns it, the lawyers should be invited before the champagne.

The Cursor and Figma Signal

The application layer is already showing how quickly value can be created and contested.

Cursor is one of the clearest examples. In June 2025, Cursor announced a $900 million raise at a $9.9 billion valuation, with more than $500 million in annual recurring revenue and usage by more than half of the Fortune 500. By November 2025, Reuters reported that Cursor had crossed $1 billion in annualised revenue. (Cursor)

That is extraordinary business formation.

But it also reveals the tension. If AI coding workflows can produce billions of dollars of annualised revenue at the application layer, why would model providers remain content selling raw tokens underneath?

OpenAI launched Codex in 2025 as a cloud-based software engineering agent that can write features, answer questions about a codebase, fix bugs, and propose pull requests, with each task running in a cloud sandbox preloaded with the user’s repository. (OpenAI)

That is not just a model. That is a workflow product.

The frontier lab is moving up the stack.

Figma shows the same pressure from the other direction. Figma is a strong software business because it owns a valuable workflow: design collaboration. Its S-1 disclosed net dollar retention of 134% as of December 2024 and 132% as of March 2025. That is exactly the kind of retention profile value investors love because it implies expansion within existing customers. (SEC)

But Figma is not standing still. It has launched Figma Make, an AI tool that can turn prompts into working prototypes, and Figma says Make currently uses Anthropic’s Claude 3.7 Sonnet. (Figma)

This is rational. Figma understands that if design moves towards prompt-to-prototype and prototype-to-code workflows, the company must move with the workflow or risk being abstracted away by someone else’s agent.

That is the point.

AI is not just adding features to software. It is redrawing the boundaries between models, applications, and enterprise workflows.

The AI Value Chain Is Being Repriced

To understand the investment implications, we need to look at the entire AI value chain.

At the bottom are chips, networking, power, data centres, and cooling. This layer is capturing enormous value because compute is scarce. NVIDIA reported Q1 fiscal 2027 revenue of $81.6 billion, up 85% year on year, with data centre revenue of $75.2 billion and GAAP gross margin of 74.9%. (NVIDIA Newsroom)

Above that sit the hyperscalers.

Their capital spending tells us that AI is no longer a normal software cycle. Amazon’s 2025 annual report showed cash capital expenditures rising from $77.7 billion in 2024 to $128.3 billion in 2025, primarily reflecting technology infrastructure investment, with the majority supporting AWS growth. Microsoft’s fiscal 2025 annual report noted a $20.1 billion increase in additions to property and equipment, while also reporting that Microsoft Cloud gross margin fell to 69%, partly due to scaling AI infrastructure. (SEC)

These are industrial numbers.

AI is making parts of software look more like energy, telecoms, or semiconductors: capital intensive, scale-driven, and dependent on utilisation. That matters for valuation. A high-growth AI company is not automatically a high-return business if it must constantly feed the furnace with GPUs, power contracts, and depreciation.

Then come the model labs.

This layer currently gets the most attention, but its long-term economics are not obvious. Frontier models may remain valuable, but customers are already learning to route tasks across multiple models. Open-weight models are improving. Inference costs are falling. Many enterprise tasks do not require the absolute frontier. They require secure, reliable, domain-adapted intelligence at the right cost.

DeepSeek was the uncomfortable data point. Reuters reported that DeepSeek disclosed a training cost of $294,000 for its R1 reasoning model, using 512 Nvidia H800 chips, based on details published in Nature. There are valid debates over what is included in that figure, and it should not be treated as a full all-in corporate R&D cost. But directionally, it challenged the assumption that useful reasoning models must always require frontier-lab levels of spending. (Reuters)

Above the models sit the application companies.

This is where the economic fight gets brutal. Many AI application companies look like high-growth SaaS, but some are essentially workflow wrappers over rented intelligence. If the underlying model provider moves up, or the enterprise builds internally, the application company can be squeezed from both sides.

The most defensible AI application companies will not be the ones with the prettiest demo. They will be the ones that own workflow distribution, proprietary data, deep integrations, switching costs, evaluation systems, and customer trust.

Finally, at the top of the value chain sits the enterprise itself.

This is the most underappreciated layer.

Enterprises already own the customer relationship, domain knowledge, transaction history, regulatory context, workflow reality, and proprietary data. In theory, incumbents should be the biggest beneficiaries of AI.

But only if they do not give away the learning loop.

The Learning Loop Is the New Moat

The new moat is not simply data.

It is the loop between data, model, workflow, user feedback, evaluation, and operational improvement.

A company that owns this loop can compound.
A company that rents this loop may commoditise itself.

That distinction matters.

Use closed frontier models for generic productivity. Fine. Use them for meeting summaries, first drafts, internal search, research assistance, customer-service triage, and experimentation. There is no need to build a nuclear reactor to toast bread.

But where the workflow is proprietary, margin-rich, and strategically central, the enterprise should think harder.

The board-level questions are straightforward:

Who owns the workflow telemetry?
Who owns the evaluation data?
Who controls model adaptation?
Who sees the usage patterns?
Who benefits from fine-tuning?
Who controls cost as usage scales?
Can the vendor become a competitor?
Does this strengthen or weaken our moat?

These are not IT questions. They are capital allocation questions.

A value investor should treat AI dependency the same way they would treat supplier concentration, customer concentration, weakening pricing power, or a declining return on invested capital.

It may not hurt earnings this quarter. In fact, it may improve them. The first phase of AI adoption may look fantastic: lower costs, faster workflows, fewer manual tasks, happier analysts, and a board deck with just enough sparkle to dazzle the non-technical directors.

But over time, dependency has a habit of showing up as margin compression.

First, the vendor gives free credits.
Then the workflow becomes embedded.
Then usage scales.
Then procurement discovers the token bill.
Then the vendor bundles more products.
Then the vendor launches a native workflow application.
Then the enterprise realises that part of its operating intelligence has become rented infrastructure.

That is how value leaks out of a business.

Quietly first. Then suddenly.

The Value Chain Consequence

AI will not distribute value evenly.

The chip layer captures value when compute is scarce. NVIDIA is the obvious winner today, but investors must watch supply growth, customer concentration, export restrictions, custom silicon, and GPU depreciation cycles.

The hyperscaler layer captures value when AI workloads standardise around cloud infrastructure. But the capital burden is enormous, and returns will depend on utilisation, pricing, power availability, and customer lock-in.

The model layer captures value when intelligence is scarce and differentiated. But if open-weight models and specialised models continue improving, generic model access may become less defensible over time.

The application layer captures value when it owns workflow, distribution, and customer context. Thin wrappers are vulnerable. Durable software companies need proprietary data, embedded process ownership, switching costs, and a credible strategy for the model layer underneath.

The enterprise layer captures value only if AI deepens its moat. If every competitor rents the same intelligence from the same supplier, the whole industry may become more efficient but less profitable.

That is the value investing paradox of AI:

Productivity can rise while moats shrink.

An entire industry can become more efficient and still destroy shareholder value if the productivity gains are competed away or captured by suppliers.

Airlines became more efficient over decades, but efficiency alone did not guarantee attractive returns. Retailers adopted e-commerce, but Amazon captured a large share of the economics. Publishers went online, but Google and Meta captured much of the advertising value.

AI could do the same to knowledge work.

The question is not whether AI improves output. It will.

The question is who keeps the economics.

The Investment Conclusion

“We use AI” is not an investment thesis.

Everyone uses AI now. That sentence is becoming as useful as “we use email”.

The better questions are:

Is AI expanding the company’s moat or compressing it?
Is it improving returns on invested capital, or creating dependency on expensive suppliers?
Is it reducing costs in a way competitors cannot easily copy, or simply giving the whole industry the same tool?
Is the company building proprietary intelligence, or training the market for someone else’s future product?
Does management understand the difference between access and ownership?

The winners will not necessarily be the loudest AI adopters. They will be the companies that know exactly where their edge comes from and refuse to outsource it blindly.

A sensible enterprise AI strategy should be simple in principle:

Rent intelligence for commodity work.
Own intelligence where the margin is earned.
Use closed models selectively.
Use open-weight and self-hosted models where control, cost, privacy, or customisation matter.
Own the workflow data.
Own the evaluation layer.
Own the customer relationship.
Own the learning loop.

This does not mean every company should become an AI lab. That would be capital misallocation with a GPU invoice attached.

But it does mean the old SaaS procurement mindset is not enough.

AI is not just another software subscription. It is a new intelligence layer being inserted into the value chain. Wherever that layer sits, economics will move towards it.

For value investors, the central question is no longer:

“Which company has AI?”

The better question is:

When AI enters the value chain, who gets to keep the value?

Because free tokens are not always free.

Sometimes they are the cheapest way for a platform to learn where the next profit pool is.

If you found value in this article, please hold down the like ❤️ button or consider buying me a coffee ☕️ to fuel the next one!

Buy me coffee ☕️

Subscribe now

The Long Game ♟️Perspective An occasional Alpha series:

Alibaba looks cheap.

But that is the easy observation.

The harder question is whether it deserves to be cheap.

That is where the Alibaba case becomes interesting.

This is not a simple story of a forgotten business trading at a low multiple. Alibaba is cheap because the market is not only valuing its e-commerce business, cloud division, AI ambitions, and cash pile.

It is also valuing China risk.

It is valuing regulation.

It is valuing geopolitics.

It is valuing years of broken investor trust.

And now, it is also valuing the uncomfortable reality that Alibaba’s AI and cloud ambitions are no longer cheap to fund.

In FY2026, Alibaba generated RMB1.024 trillion in revenue. On the surface, revenue grew only 3%. But excluding disposed businesses such as Sun Art and Intime, like-for-like revenue grew 11%.

That is not explosive growth.

But it is also not a dying business.

The cloud business looks more interesting. In the March quarter, Alibaba’s Cloud Intelligence Group revenue grew 38% year-on-year to RMB41.6 billion. External cloud revenue grew 40%. AI-related products accounted for roughly 30% of external cloud revenue.

That is real progress.

But here is the catch.

Alibaba’s free cash flow swung from an inflow of RMB73.9 billion in FY2025 to an outflow of RMB46.6 billion in FY2026.

That is not a small footnote.

That is the whole debate.

Alibaba is trying to convince the market that it is becoming China’s AI infrastructure platform.

The market is replying with a colder question:

Fine. But how much cash must you spend to get there?

That is why I do not think the key number for Alibaba is P/E.

The key number is free cash flow.

More precisely, whether cloud and AI growth can eventually become durable free cash flow per share.

A low multiple can make a stock look safe.

But if the business needs to keep spending heavily on cloud infrastructure, quick commerce, AI models, logistics, and competitive defence, the margin of safety becomes less obvious.

Alibaba may be cheap.

But cheap is not the same as safe.

The market is not just valuing the business

When investors look at Alibaba, they are not only valuing Taobao, Tmall, Alibaba Cloud, AliExpress, Cainiao, Qwen, or its broader digital ecosystem.

They are also valuing the country.

That is the awkward part.

A spreadsheet can handle revenue growth.

It can handle margins.

It can handle cash flow.

It struggles with trust.

Alibaba still has enormous scale. It remains one of China’s most important digital commerce platforms. Its cloud business is accelerating. Its Qwen model family gives it a credible AI platform. Its balance sheet remains substantial.

But investors have been trained by experience to be sceptical.

Regulation has hurt sentiment before.

China’s macro recovery has been uneven.

U.S.-China tension remains a live risk.

Foreign investors do not fully trust the policy environment.

The recent U.S. defence blacklist issue adds another layer of discomfort.

This is why Alibaba’s discount is not random.

The market is not being stupid.

It may be too pessimistic.

But it is not being stupid.

That distinction matters.

A low valuation can mean the market is wrong. It can also mean the market is charging you for risks that do not fit neatly into a valuation model.

Alibaba sits exactly there.

The question is not whether the company is optically cheap.

It is whether the discount is too harsh.

The bull case is getting more interesting

The bull case is no longer just:

Alibaba is cheap.

That is too lazy.

The stronger bull case is that Alibaba is still being valued largely like a politically damaged Chinese e-commerce company, while the business is trying to reposition itself as a full-stack AI and cloud infrastructure platform.

That matters.

Cloud revenue grew 38% in the March quarter. External cloud revenue grew 40%. AI-related products made up around 30% of external cloud revenue. Cloud adjusted EBITA grew 57% to RMB3.8 billion.

That puts cloud adjusted EBITA margin at roughly 9.1% for the quarter.

Not amazing.

But not irrelevant either.

The important point is not that Alibaba Cloud is suddenly AWS.

It is that cloud is becoming more meaningful inside the Alibaba story.

For years, Alibaba was mostly seen through the lens of Chinese e-commerce. That made sense. Commerce was the core engine. It funded the ecosystem. It gave Alibaba scale, data, customers, merchants, payments exposure, logistics relevance, and strategic reach.

But the next phase of the company may be different.

Qwen gives Alibaba a foundation-model layer.

Alibaba Cloud gives it the infrastructure layer.

Its commerce ecosystem gives it use cases.

Its enterprise relationships give it distribution.

Its logistics and consumer platforms give it real-world operating data.

If these pieces connect, Alibaba is not just defending an old business.

It is trying to build the infrastructure for a new one.

That is the upside case.

The market may eventually stop asking:

Is Alibaba still a wounded e-commerce stock?

And start asking:

What is China’s leading cloud and AI platform worth?

Those are very different questions.

The balance sheet gives Alibaba time

Alibaba is not a weak company trying to buy relevance with money it does not have.

As of 31 March 2026, Alibaba had RMB520.8 billion in cash and other liquid investments.

That is roughly 51% of FY2026 revenue.

This matters because AI and cloud are capital-intensive games.

Model training is expensive.

Computing infrastructure is expensive.

Enterprise cloud expansion is expensive.

Quick commerce is expensive.

User acquisition is expensive.

A weaker company would be forced to choose between survival and ambition.

Alibaba still has room to invest.

That does not guarantee success.

But it gives the company time.

And in technology, time matters.

Especially when the market is impatient.

This is the part that makes Alibaba difficult to dismiss. The company is not merely hoping that AI will save it. It has the balance sheet, infrastructure, and domestic scale to make a serious attempt.

But having the resources to fight is not the same as winning the fight.

Which brings us to the uncomfortable side of the story.

Growth is not free

The bear case starts with one number.

Negative RMB46.6 billion.

That was Alibaba’s FY2026 free cash flow.

A year earlier, free cash flow was positive RMB73.9 billion.

That is a swing of more than RMB120 billion.

On FY2026 revenue of RMB1.024 trillion, Alibaba’s free cash flow margin was roughly negative 4.6%.

This is the number that keeps the thesis honest.

Yes, cloud is growing.

Yes, AI-related revenue is becoming meaningful.

Yes, the balance sheet is still strong.

But the current investment cycle is expensive.

Alibaba is spending heavily on quick commerce, user acquisition, cloud infrastructure, AI development, and technology businesses.

In the March quarter, sales and marketing expenses rose to 21.9% of revenue, compared with 15.3% a year earlier.

Product development expenses, excluding share-based compensation, increased to 7.3% of revenue, compared with 5.7% a year earlier.

This is not a quiet compounder at the moment.

This is a company investing aggressively while profitability is under pressure.

FY2026 income from operations fell 64%.

Adjusted EBITA fell 56%.

Non-GAAP net income fell 62%.

That is not noise.

That is the price of the current strategy.

The company may be right to spend.

But shareholders still need to ask what they are paying for.

There is a big difference between investing through a temporary cycle and entering a permanently more capital-intensive business model.

That distinction will decide whether Alibaba is cheap or merely optically cheap.

The key number is not P/E

Many investors will naturally start with the valuation multiple.

That is understandable.

Alibaba often looks cheap compared with large U.S. technology companies.

But I do not think P/E is the right starting point here.

Not for this type of situation.

The key number is free cash flow.

More specifically:

Can Alibaba turn cloud and AI growth into durable free cash flow per share after funding the infrastructure required to compete?

That is the question.

A low P/E can make a stock look cheap.

A large cash pile can make it look safer.

An AI story can make it feel exciting.

But the real test is whether the business can generate cash after reinvestment.

Owner earnings matter.

Not headlines.

Not adjusted narratives.

Not “AI exposure”.

Actual cash that can be retained, reinvested, or returned to shareholders.

That is why Alibaba’s numbers are so mixed.

The cloud business is growing quickly.

AI-related cloud revenue is becoming more visible.

The balance sheet is strong.

But free cash flow has gone negative, margins are under pressure, and spending intensity has increased.

That is not a broken thesis.

It is a complicated thesis.

And complicated theses require a larger margin of safety.

The numbers that matter from here

For Alibaba, the investment case does not need more slogans.

It needs evidence.

These are the numbers I would watch.

Cloud revenue growth.

If cloud growth stays strong, Alibaba’s AI infrastructure case becomes more credible.

AI-related cloud revenue share.

If AI-related products continue taking a larger share of external cloud revenue, the AI story becomes more than marketing.

Cloud adjusted EBITA margin.

In the March quarter, cloud adjusted EBITA margin was roughly 9.1%. That needs to improve over time if cloud is to become a serious profit engine.

Group free cash flow margin.

FY2026 free cash flow margin was roughly negative 4.6%. This needs to recover.

Capex intensity.

If cloud and AI require consistently high reinvestment, the valuation deserves to remain lower.

Sales and marketing as a percentage of revenue.

At 21.9% of revenue in the March quarter, this shows how expensive competition and quick commerce are becoming.

Product development intensity.

At 7.3% of revenue, excluding share-based compensation, this reflects the cost of technology investment.

China commerce margin trend.

The core business still matters. If the cash engine weakens while the growth engine requires heavy reinvestment, the thesis becomes much harder.

Buybacks versus free cash flow.

Buybacks only create value if they are funded by durable economics, not by balance-sheet shrinkage alone.

This is the real scoreboard.

Not whether Alibaba announces more AI products.

Not whether the stock gets another short-term sentiment bounce.

The scoreboard is whether growth eventually becomes cash.

The China discount may not disappear

The hardest part of Alibaba is that even good execution may not be enough.

That sounds unfair.

It may still be true.

Alibaba could continue growing cloud.

It could build one of China’s strongest AI platforms.

It could improve cloud margins.

It could buy back shares.

It could stabilise commerce profitability.

And the market may still refuse to value it like a Western technology company.

Because the discount is not only about business quality.

It is about trust.

Investors are asking whether shareholder value will be protected.

They are asking whether policy risk can suddenly change the economics.

They are asking whether geopolitical tension can restrict customers, reduce foreign appetite for the shares, or keep valuation multiples structurally depressed.

That is why Alibaba is not a clean compounder.

A clean compounder gives investors the luxury of focusing mostly on business quality, reinvestment runway, management, and valuation.

Alibaba does not give that luxury.

With Alibaba, investors must also underwrite the jurisdiction.

That deserves a discount.

The only debate is how large the discount should be.

This is where many Alibaba discussions become too simplistic.

The bulls say the stock is cheap.

The bears say China risk makes it uninvestable.

Both sides may be too comfortable.

The more honest view is that Alibaba has real business progress and real non-business risk at the same time.

That is what makes it difficult.

And difficulty is exactly why the opportunity exists, if it exists at all.

Where I may be wrong

There are two ways to be wrong on Alibaba.

The first is being too cautious.

Alibaba’s AI and cloud business may scale faster than expected. AI-related revenue may become the main growth driver of cloud sooner than the market assumes. Qwen may become more important across enterprise, commerce, agentic applications, and robotics-related use cases.

Cloud margins may improve as utilisation rises.

Buybacks at depressed valuations may quietly increase intrinsic value per share.

China sentiment may recover.

The geopolitical discount may narrow.

If that happens, Alibaba can re-rate quickly.

Hated stocks do not need perfection to move.

They only need the story to become less bad.

The second way to be wrong is being too optimistic.

The China discount may be permanent.

AI capex may stay elevated for longer than expected.

Quick commerce may keep pressuring margins.

Cloud growth may not convert into enough free cash flow.

Geopolitical tension may keep foreign investors away.

The business may look cheap for years while shareholders wait for a re-rating that never comes.

That is the danger.

Alibaba may be cheap for a reason.

Sometimes the reason fades.

Sometimes it becomes the investment case.

Final view

Alibaba is one of the more interesting large-cap value situations in technology.

But it is not a comfortable one.

The company has real scale, real assets, real AI and cloud momentum, and enough balance-sheet strength to keep investing through a difficult cycle.

At the same time, free cash flow has weakened, profitability is under pressure, and the geopolitical discount is not imaginary.

My classification:

Attractive for further study, but high complexity.

I would not describe Alibaba as a clean compounder.

I would not describe it as an obvious bargain.

I would describe it as a discounted complexity bet.

The upside depends on three things.

Cloud and AI must keep growing.

That growth must eventually convert into durable free cash flow.

And the China discount must not permanently overwhelm the business progress.

That is a high bar.

But it is also why the stock is interesting.

The market may be too pessimistic.

But it is not being stupid.

The Long Game takeaway

Cheap stocks are seductive because they appear to offer mathematical comfort.

Low multiple.

Big cash pile.

Famous company.

Strong narrative.

But value investing is not just buying what looks cheap.

It is asking whether the market is wrong about the future, or simply right about the risk.

Alibaba sits in that uncomfortable middle.

The business is improving in places that matter.

The risks are real in places that cannot be ignored.

That is the kind of situation where investors must slow down.

Not because there is no opportunity.

But because the opportunity is not the same as the story.

Alibaba may turn out to be mispriced.

But it is not mispriced for no reason.

And that is exactly what makes the case worth studying.

If you found value in this article, please hold down the like ❤️ button or consider buying me a coffee ☕️ to fuel the next one!

Buy me coffee ☕️

Subscribe now

The Long Game ♟️Perspective An occasional macro trend series:

The Miracle That Cannot Turn a Profit

Sit down with any venture capitalist or portfolio manager here in Singapore today, and you will hear a familiar, breathless narrative. We are deep into 2026, and the capabilities of generative artificial intelligence are genuinely miraculous. The ability to converse with a machine, to have it synthesise complex datasets in seconds, or to generate functional code without a human engineer is astonishing. The utility is real. The disruption is vast.

Yet, beneath this effervescent enthusiasm lies a cognitive dissonance that the financial media seems entirely unwilling to address. There is a quiet, creeping panic setting in across the boardrooms of the largest technology firms regarding the sheer physics of their ambition. The capital expenditures required to train and run frontier generative models are staggering, and the return on that investment remains dangerously opaque.

We are witnessing a profound conflation in the modern markets. Investors have convinced themselves that technological utility automatically translates into economic value capture. The boosters point to undeniable productivity gains. The sceptics warn of technological plateaus. Both factions are missing the third, and historically most common, scenario. You can have a revolutionary, world-changing technology that alters the trajectory of human civilisation while simultaneously acting as a wealth-destroying furnace for the investors who fund its early stages.

The mathematics are refusing to resolve. If a foundational model provider decides to raise their prices tenfold tomorrow to cover these spiralling infrastructure costs, a boutique healthcare start-up running a highly specific application might pay the premium. However, a multinational corporation employing ten thousand engineers cannot absorb a massive multiplier on its software overhead without fundamentally breaking its operating margins. The current business model for widespread, frictionless artificial intelligence is structurally flawed.

The Ghost of Kitty Hawk and the Capex Reality

To understand the trap that modern capital is walking into, we must discard the recent past and look to the deeper archives of industrial history. The venture capital community has spent the last two decades operating under the paradigm of Software as a Service. In the software paradigm, the marginal cost of production approaches zero. Once a developer writes a piece of code, distributing it to one user or one million users costs roughly the same amount. This economic reality generated the highest profit margins and the fastest wealth creation in modern history.

Investors are now violently misapplying that exact same zero-marginal-cost mental model to artificial intelligence. This is a profound categorisation error. Generative computing is not software. It is heavy industry disguised as software.

Every single query typed into a foundational model requires a physical asset to compute it. It requires sophisticated cooling systems to prevent the hardware from melting. It requires massive baseload electricity. The marginal cost does not drop neatly as user adoption scales. It remains stubbornly high. The infrastructure required to sustain this illusion of seamless digital magic is brutally physical. It involves pouring thousands of tonnes of concrete, laying miles of copper wiring, and hiring legions of electricians and plumbers. The ultimate constraint on technological progress is no longer the brilliance of the software engineer in Silicon Valley, but the availability of the tradesperson and the capacity of the local power grid.

History is replete with exactly this dynamic. Consider the commercial aviation industry. No one can deny that flight fundamentally changed human existence. It collapsed geography and built the modern global economy. Yet, Warren Buffett famously observed that the airline industry was such a reliable machine for incinerating capital that an enterprising capitalist standing at Kitty Hawk in 1903 would have done future investors a massive favour by shooting Orville Wright down. The societal benefit was immense, but the capital intensity and total lack of pricing power meant that airlines destroyed investor wealth for decades.

A more recent parallel is the biotechnology revolution. The ability to manipulate genetic material has saved countless lives. However, comprehensive research by Harvard Business School academic Gary Pisano demonstrated a sobering reality. For the three decades between 1975 and 2004, the biotechnology sector as a whole generated almost zero aggregate economic profit. It was a furnace for capital. Investors poured billions into research and development, navigated extreme biological uncertainty, and endured long delays. Society reaped the reward of life-saving drugs, but the early equity investors were largely diluted into irrelevance by the time the cash flows materialised.

Today’s technology giants face an even steeper climb. A pharmaceutical company, once it navigates the clinical trial process, is rewarded with a patent. That legally enforced monopoly is a genuine economic moat. Artificial intelligence models have almost no defensible moat. Recent benchmarking shows that open-source models (including highly powerful variants emerging from Chinese laboratories) are currently lagging the frontier models by a matter of mere months. If your multi-billion-dollar proprietary advantage can be replicated and offered for free by a competitor in less than a year, you do not have a moat. You have a very expensive, very temporary head start.

Fear, Neuroeconomics, and the Herd Mentality

If the historical precedents are so clear, and the capital expenditure mathematics are so daunting, why are the world’s smartest capital allocators still funnelling billions into the sector? The answer lies not in spreadsheets, but in biological hardwiring.

Markets are entirely composed of human decisions. In neuroeconomic terms, the pain of social exclusion registers in the exact same region of the human brain as physical pain. This is the physiological root of herd mentality. For a portfolio manager, sitting out the current infrastructure rally carries acute career risk. If you participate and it turns out to be a bubble, you fail conventionally alongside your peers (which is always forgiven by the financial industry). If you sit out and the technology achieves escape velocity, you are ostracised and rendered obsolete. Therefore, the rational choice for the individual agent seeking career preservation is to join the irrational herd.

There is another psychological dynamic at play beneath the surface, one that directly impacts corporate behaviour. This is the weaponisation of fear within the labour market. Some optimistic macroeconomic analysts have noted that algorithms have not yet caused mass unemployment, taking this as a sign that the technology is purely additive. This is a remarkably shallow reading.

The true impact is not immediate replacement, but pervasive ambient anxiety. Even if an employee is not actively being replaced today, the credible threat that they could be replaced tomorrow fundamentally shifts the balance of power. During the recent pandemic, leverage belonged entirely to highly skilled labour. Workers demanded remote flexibility and aggressive wage increases. Today, the mere existence of sophisticated algorithms serves to discipline the workforce. Employees are far less likely to demand a raise when they are terrified of rendering themselves obsolete. For the chief executive officer facing crushing infrastructure bills, this suppressed wage growth offers a convenient, hidden subsidy.

Four Metrics to Separate the Signal from the Furnace

For the sophisticated investor managing generational wealth, the task is to navigate this treacherous landscape without falling prey to blind optimism or total cynicism. The technological utility is real, and the financial bubble is real. Both realities coexist.

To separate the fundamental truth from the speculative frenzy, you must look past the press releases and track specific, pragmatic metrics.

1. Free Cash Flow Over Adjusted Earnings

   Ignore adjusted earnings completely. Ignore top-line revenue growth if it is entirely consumed by hardware acquisition costs. The only metric that matters for an infrastructure-heavy business is Free Cash Flow. If a company is structurally incapable of generating positive cash flow after accounting for the constant capital expenditures required to stay relevant, it is a speculative science experiment.

2. Return on Invested Capital Against Cost of Capital

   Every dollar poured into data centres and advanced processors must eventually yield a return that exceeds the firm’s cost of capital. Currently, the industry is operating on a blind faith philosophy. You must scrutinise whether the incremental revenue generated by a new algorithmic feature actually justifies the astronomical cost of the compute power required to deliver it.

3. The Open-Source Decay Rate

   Watch the performance gap between proprietary, closed-source models and open-source alternatives. If an enterprise client can achieve 95% of the performance of a premium model using a free open-weights alternative, the multi-trillion-dollar valuations of the market leaders will face a brutal reckoning. This decay rate is the single most important variable in the valuation equation.

4. Energy Capacity and Regulatory Pushback

   Because this technology is a physical phenomenon, its ultimate constraint is energy and real estate. Data centres are facing increasing political resistance globally. Track the energy procurement contracts and local regulatory approvals carefully. If a company cannot secure the physical grid capacity to power its servers, its software ambitions are irrelevant.

Positioning for the Second-Order Reality

What does this mean for the long-term wealth and portfolio strategy of a Singapore-based reader?

Singapore offers a highly instructive lens through which to view this crisis. The city-state is the ultimate physical constraint environment. Land is strictly finite, and energy is largely imported. The government here understands the physical limitations of the digital world better than most, having previously imposed a strict moratorium on new data centres to manage the severe strain on the national grid. Even as that moratorium has been cautiously lifted in recent years, the focus is strictly on highly efficient, power-conscious infrastructure that must meet rigorous sustainability criteria. Singapore does not have the luxury of pretending that digital actions have no physical consequences.

The mainstream investment view is to aggressively buy the companies designing the software or selling direct consumer applications. This is precisely where the highest risk of capital incineration lies. The second-order thinker looks at the physical bottlenecks.

If this new paradigm is the equivalent of the nineteenth-century railroad boom, you do not want to invest in the railroad operators who are engaged in a vicious price war. You want to own the land rights, the steel mills, and the coal mines. In the modern context, this translates to hard, constrained assets. It means looking deeply into the industrial companies that manufacture the advanced liquid cooling systems required for high-density server racks. It means exploring the utility companies and grid infrastructure providers that act as the unavoidable tollbooths for this extraordinarily energy-hungry technology. It means considering real estate investment trusts that hold legacy data centres which (due to tightening zoning restrictions globally) can no longer be replicated or replaced by new market entrants.

Do not allow the spectacular nature of the technology to distract you from the dull, unchanging mathematics of value investing. A bad business model wrapped in a miraculous technology is still a bad business model.

Ensure that your portfolio is anchored in businesses that possess actual pricing power, defendable moats, and structural profitability. Let the venture capitalists fund the high-risk science experiments. Your mandate is to preserve and compound capital in the real world.

The Question: If the open-source community manages to match the capabilities of proprietary artificial intelligence models within the next twelve months, what is the true intrinsic value of the software giants currently dominating your portfolio?

Sources

1. Bloomberg Television (Interview with Setha Macunda on AI Costs and Infrastructure Constraints, May 2026).

2. Harvard Business School (Gary Pisano’s historical research on biotechnology sector profitability).

3. Monetary Authority of Singapore (Guidelines on sustainable infrastructure and data centre investments).

If you found value in this article, please hold down the like ❤️ button or consider buying me a coffee ☕️ to fuel the next one!

Buy me coffee ☕️

Subscribe now

The Long Game ♟️Perspective An occasional macro trend series:

Read more

Subscribe now

The Long Game ♟️Perspective An occasional macro trend series:

Read more

Subscribe now

The Long Game ♟️Perspective An occasional Alpha Library series:

Read more

Subscribe now

The Long Game ♟️Perspective An occasional macro trend series:

Read more

Subscribe now

The Long Game ♟️Perspective An occasional wealth mindset series:

Read more