AI Spending Is Moving From Cash to Debt. Why Investors Should Care

A few weeks ago, I posted on my Instagram about the more important question surrounding AI.

Not:

Will AI work?

But:

Will the returns eventually justify the hundreds of billions of dollars being spent to build it?

I still think that’s the right question.

But something else is beginning to happen that makes the discussion even more interesting.

The AI build-out is no longer purely a story about extremely profitable technology companies taking excess cash from their businesses and reinvesting it into data centres.

Increasingly, debt is entering the picture.

And that matters.

Because there is a big difference between investing money you already have...

And borrowing against the assumption that future demand will justify what you’re building today.

 

The First Phase of the AI Boom Had an Important Advantage: Cash

One reason I’ve been relatively cautious about comparing today’s AI boom directly with the dot-com bubble is that the companies spending the most money today are very different from many of the companies that drove investment 25 years ago.

Amazon.

Alphabet.

Microsoft.

Meta.

These are not start-ups with speculative business models and no profits.

They are some of the most cash-generative businesses ever created.

S&P Global estimates that five major hyperscalers — Amazon, Alphabet, Microsoft, Meta and Oracle — have already spent around US$1.1 trillion in capital expenditure over the past five years.

And based on Visible Alpha consensus estimates, another US$5.3 trillion could be spent by 2030.

That’s extraordinary.

But S&P makes an equally important point.

Despite the unprecedented investment intensity, this cycle has not yet developed the financing characteristics that made the dot-com build-out especially fragile.

In other words:

A lot of this spending has been backed by businesses capable of generating enormous amounts of their own cash.

That’s important.

Because cash gives you time.

Subscribe to our newsletter if you enjoy insights like this.

Debt Changes the Equation

Imagine two companies each decide to build a $10 billion data centre.

Company A has $100 billion sitting on its balance sheet.

It pays cash.

Company B contributes $2 billion and borrows the remaining $8 billion.

The physical asset might be identical.

The demand assumptions might be identical.

The expected revenue might even be identical.

But financially, these are very different investments.

If demand takes two years longer than expected to materialise, Company A may be disappointed.

Company B still has interest payments.

If utilisation reaches only 60% instead of 90%, Company A earns a lower return.

Company B earns a lower return while still servicing its debt.

Debt doesn’t necessarily make an investment bad.

But it reduces the room for error.

That’s why S&P Global describes one of the key risks in the current AI cycle as a transition from cash-funded investment toward debt-funded investment before the returns have been validated.

I think that sentence deserves more attention from investors.

The Bond Market Is Starting to Notice

This isn’t theoretical anymore.

Reuters reported recently that gross debt issuance from hyperscalers is expected to reach around US$420 billion in 2027, based on Goldman Sachs data.

That would be roughly 60% higher than estimated 2026 issuance.

Interestingly, bond investors aren’t necessarily worried that companies like Meta or Alphabet are suddenly going to default.

Their balance sheets remain strong.

The concern is more subtle.

There is simply a huge amount of borrowing coming to market, while investors still have limited visibility over the ultimate return on all this AI infrastructure.

That is already affecting pricing.

Reuters reported that spreads on AI-related issuers were running at roughly 115 basis points, compared with around 78 basis points for the broader investment-grade market.

Essentially, lenders are beginning to say:

If you want us to finance this much AI infrastructure, we want to be compensated for it.

That is a healthy market mechanism.

But it also tells us something.

The cost of funding the AI boom is becoming part of the investment thesis.

Sometimes the Debt Isn’t Even on the Big Tech Company’s Balance Sheet

This is where things get slightly more complicated.

Not every data centre is directly owned and financed by Microsoft, Google or Meta.

Instead, infrastructure providers may borrow money to build the facility.

A technology company signs a long-term lease.

That lease helps support the financing.

Banks provide loans.

Institutional investors buy the debt.

Suddenly, the economic risk created by AI demand is spread across multiple balance sheets.

A recent example is Oracle-linked Project Jupiter in New Mexico.

Reuters reported that around US$18 billion of loans tied to the Oracle-leased data-centre project had been quoted at roughly 89 to 91 cents on the dollar as efforts to distribute the debt encountered difficulties.

The project forms part of Oracle’s broader agreement with OpenAI to supply AI computing capacity.

I wouldn’t use one project to conclude that the entire AI financing system is in trouble.

That would be a stretch.

But it is a useful example of something investors increasingly need to understand:

The infrastructure risk doesn’t disappear just because somebody else owns the building.

Sometimes it simply moves elsewhere in the financial system.

This Is Where Return on Invested Capital Becomes Critical

When capital is cheap and internally generated, companies can afford to experiment.

When capital becomes increasingly external and carries an explicit cost, the hurdle rises.

Let’s say a company borrows at 6%.

It doesn’t make much sense to build infrastructure that ultimately generates a 5% return.

At least not indefinitely.

The next dollar invested needs to produce a return meaningfully higher than the cost of financing it.

Otherwise, you’re not creating value.

You’re simply growing the asset base.

This is why I think investors should pay increasingly close attention to incremental return on invested capital.

Not just:

How much AI revenue did the company generate?

But:

How much additional capital did it require to generate that revenue?

Because revenue growth and value creation are not the same thing.

Investment Booms Tend to Follow a Familiar Pattern

S&P Global describes capital expenditure cycles using four stages:

Build.

Stretch.

Overshoot.

Digest.

The build phase makes sense because returns justify additional investment.

The stretch phase continues because expected demand remains strong.

The danger appears when the expected return from the next investment no longer exceeds its cost of financing.

That’s when the boom can move into overshoot.

I don’t know whether AI is approaching that stage.

Nobody does.

But that’s exactly what we should be watching.

Not whether data-centre construction is increasing.

Not whether AI demand is growing.

But whether each additional dollar being invested continues to generate an attractive economic return.

Want a concise version of Charlie Munger’s 25 human misjudgements? Download here!

This Doesn’t Mean AI Spending Is About to Collapse

This is important.

I am not saying:

AI debt is rising, therefore the AI boom is ending.

In fact, bond investors quoted by Reuters explicitly said the widening spreads were largely about supply, portfolio concentration and uncertainty around future returns — not fears that hyperscalers are suddenly poor credit risks.

The major technology companies still generate enormous amounts of cash.

AI demand remains significant.

And some of these investments could produce returns far above their cost of capital.

The point is simply that the nature of the investment cycle is evolving.

And as investors, our analysis should evolve with it.

What I Would Watch From Here

I’d watch four things.

First, free cash flow.

How quickly is capex growing relative to operating cash generation?

Second, debt issuance.

Are companies increasingly relying on financing rather than internally generated cash?

Third, utilisation.

It is one thing to build data centres.

It is another to keep them full at attractive prices.

And fourth, incremental returns on capital.

If another $100 billion goes into AI infrastructure, how much additional operating profit eventually comes out?

That’s ultimately what matters.

AI Can Be Revolutionary and Still Have a Financing Problem

One of the easiest mistakes investors can make is turning every discussion into two camps.

AI is the future.

Or:

AI is a bubble.

Reality is usually more complicated.

AI can transform the economy.

Demand can remain enormous.

The technology can genuinely improve productivity.

And companies can still overinvest.

All of those things can be true at the same time.

So I don’t think the most useful question is whether debt proves something is wrong with AI.

Instead, I would ask:

At what point does an investment boom become dependent on increasingly expensive capital to keep going?

And:

Are the returns arriving quickly enough to justify that capital?

During the first phase of this AI boom, the strongest companies had an enormous advantage.

They could fund the future using cash generated from the present.

If the next phase increasingly relies on borrowing against the future instead...

That doesn’t mean the story is over.

But it does mean investors should start reading the balance sheet as closely as they read the AI headlines.