When Money Circles Back on Itself, Is That Real Demand?

Imagine rolling a snowball down a slope.

It starts small. As it rolls, it gathers snow, grows heavier, and accelerates. Before long, from a distance, it looks like an immense force growing under its own power.

But it is easy to forget one question as we watch:

How much of that growth is the core, and how much is simply the volume gathered by motion?

This question increasingly comes to mind when I look at the AI economy today.

The investment figures are staggering. Data centers are being built, chip orders are growing, energy agreements are being signed, and multibillion-dollar financing packages are being announced. There is no doubt that a real technological transformation lies behind it all.

But when you start following the money, the picture becomes more interesting.

The same companies invest in one another. Become one another’s customers. Buy capacity from one another. Sometimes participate in one another’s financing.

And that is when the finance professional’s instinct kicks in:

How much of this growth comes from real customer demand outside the cycle, and how much from the ecosystem financing itself?

Where does the money go?

Let us make this concrete.

In 2025, Nvidia announced plans to invest up to $100 billion in OpenAI. The other side of the agreement involved OpenAI deploying at least 10 gigawatts of Nvidia systems.

So Nvidia was not merely OpenAI’s supplier.

It was also an investor.

Microsoft has been one of OpenAI’s largest investors for years. It also owns Azure, a major provider of the computing infrastructure OpenAI uses. OpenAI now works with other infrastructure providers too, but the economic logic of the relationship makes the issue clear enough: capital enters an AI company, the AI company buys computing capacity, the cloud company builds data centers, and the data centers buy Nvidia chips.

The money does not disappear.

It becomes chips. Servers, data centers, electricity, cooling, and land.

So this is not an imaginary economy.

But there is another question, and I think it is the central one:

Who is bringing money into the system from outside this cycle?

If a bank buys an AI service and pays for it, that is external demand.

If a manufacturer reduces costs using AI and pays for the system, that is too.

If a law firm, retailer, hospital, SME, or consumer pays for an AI product out of their own pocket, the system is generating external revenue.

That is the core of the snowball.

Nvidia investing in an AI company, that company buying Nvidia chips, and renting cloud capacity from another technology company creates economic activity. But it does not, by itself, prove that final demand is growing at the same pace.

$100 billion circulating within the system does not mean there is $100 billion in customer demand outside it.

To me, that distinction captures the entire debate in one sentence.

A Ponzi scheme?

At this point, the comparison immediately appears.

‘Is this a Ponzi scheme?’

No.

The term dramatizes the issue more than it explains it.

In a classic Ponzi scheme, there is no underlying real economic activity. New investors’ money is paid to earlier investors as returns, and the system collapses when new money stops coming in.

In the AI economy, very real things exist.

The chips are real. The data centers are real. The energy consumption is real. The products are real. The customers are real. And in some use cases, the productivity gains are real too.

Moreover, companies such as Microsoft, Alphabet, Amazon, and Meta, which account for a substantial share of investment, have enormous existing businesses and strong cash flows behind them.

So the problem is not that the technology is fake.

The question is how much economic expectation is being purchased around a real technology.

The language used in the IMF’s 2026 assessment matters here. It specifically highlights circular financing relationships in which a small group of AI-ecosystem companies become one another’s customers, investors, and financiers at the same time.

These relationships are not inherently wrong.

But they make it harder to read the source of demand.

The right word, then, is not Ponzi.

It is fragility.

Because revenue and economic value are different things

OpenAI is a good example of this debate.

According to company projections reported by the Financial Times in September 2026, OpenAI’s revenue is expected to rise from approximately $36 billion in 2026 to $350 billion in 2030.

Enormous growth.

But those same projections have another side: the company is expected to generate approximately $280 billion in negative free cash flow between 2026 and 2030.

Putting those two figures together makes the issue clearer.

We are no longer debating whether AI generates revenue. It does.

The real question is:

How much capital is consumed to generate that revenue?

This is precisely where finance distinguishes growth from value creation.

A company can multiply its sales while needing far more capital than those sales to finance its growth. That does not mean it is a bad company. But it fundamentally changes its valuation, financing needs, and investment return calculations.

This distinction is especially important in AI, because behind what we call software lies a surprisingly heavy physical economy.

The cloud is actually quite heavy

AI brings a digital world to mind.

Yet behind that world are GPUs, enormous buildings, transformers, power lines, water, cooling systems, and miles of infrastructure.

According to Berkeley Lab’s study, US data centers consumed approximately 176 TWh of electricity in 2023, around 4.4% of the country’s total electricity consumption. The estimate for 2028 is 325–580 TWh, or between 6.7% and 12% of total US electricity consumption.

An update published in June 2026 shows the trend continuing: data centers could consume around 11.8% of US electricity in 2030, rising to as much as 15.3% under some scenarios.

Beyond a certain point, AI’s limits are no longer only about developing better models.

Where to find the electricity becomes an issue.

Grid capacity becomes an issue.

Cooling becomes an issue.

Land becomes an issue.

And who will finance it all becomes an issue.

An increasingly asset-heavy economy is forming behind technology that appears asset-light.

Which brings us back to the same question:

Who will pay for the return on all that capital?

Does it resemble the dot-com era?

Somewhat.

But the comparison has limits.

In the late 1990s, people also believed a new technology would change the world. They were right.

The internet really did change the world.

But that did not mean every internet-related investment in 1999 was correctly priced.

I think that is the most important historical lesson for today too.

You can be right about a technology and wrong about an investment in that technology.

AI may genuinely transform the economy.

At the same time, some AI companies may be overvalued, some data centers may have been built too early, and some capacity contracts may rest on overly optimistic demand forecasts.

Those two statements do not contradict each other.

There is also an important difference from the dot-com era. Most of the large companies at the center of AI infrastructure are highly profitable. So far, they have been able to finance a significant share of investment from their own cash flows.

The IMF highlights this too.

But the same institution is now examining the next stage: as investment needs exceed existing corporate cash flows, debt financing plays a larger role.

That is the part I want to watch more closely.

Because an expectation financed with equity and an expectation financed with debt leave different marks on the economy.

If the expectation fails to materialize, the equity holder loses money.

Debt still demands repayment after the story has fallen apart.

What happens if the avalanche comes?

That is why considering a possible AI correction solely through technology stocks gives an incomplete picture.

An IMF analysis published in January 2026 contains an interesting estimate. A moderate correction in AI stocks, accompanied by tighter financial conditions, could reduce global growth in 2026 by approximately 0.4 percentage points relative to the baseline.

That does not mean 2008.

Indeed, current data is not sufficient to suggest a 2008-style banking crisis is the most likely scenario.

But there is another channel.

Equity markets fall.

Wealth declines.

Consumption is affected.

Companies’ cost of capital rises.

New investments are postponed.

Companies dependent on external financing find it harder to raise money.

Buyers of AI infrastructure scale back their plans.

And what initially looked like a valuation problem in the technology sector gradually seeps into the real economy.

Today’s market structure is also different. According to the IMF’s calculations, the total value of the US stock market was approximately 132% of GDP in 2001, compared with around 226% today.

So even a smaller market correction could have a larger economic effect through wealth and consumption.

The danger of the snowball is not only how large it is.

It is what it is connected to.

Perhaps we are asking the wrong question

I find the question ‘Is AI a bubble?’ increasingly uninteresting.

Because the answer will probably not be a single word.

AI itself may not be a bubble.

But the valuation of certain AI companies may be.

The technology may be extraordinary.

But the returns on some investments may fall short.

Demand may be real.

But the capacity being built may have run ahead of it.

Revenue may grow rapidly.

But the capital required to produce that revenue may grow faster.

So instead of asking whether AI is real, finance professionals need to ask more ordinary, but more useful, questions:

Where does the money come from?

Where does it go?

Who finances it?

Who borrows?

And most importantly:

Who pays the bill at the end of the chain?

Ultimately, the sustainability of this entire system will not be determined by how much money Nvidia transfers to OpenAI, Microsoft to OpenAI, or an investor to another AI company.

At some point, a real customer outside the cycle needs to pay out of their own pocket.

A bank.

A manufacturer.

A retailer.

A hospital.

An SME.

A consumer.

And there must be a real reason behind that payment: higher revenue, lower costs, faster processes, better decisions, or greater productivity.

That is the core.

Time will tell how much of the rest is lasting.

Looking inside the snowball

This is also why, at Ainsideme, we insist so strongly on connecting AI to measurable business outcomes.

When a company says, ‘We use AI,’ my main question is not which model it uses.

What is the result?

What happened to cash flow?

Did costs fall?

Did revenue rise?

Did risk decline?

Did decision quality improve?

Was measurable value created in return for the capital spent?

Because believing in a technology is one thing.

Believing in its economics is another.

When we see the next billion-dollar AI deal, perhaps our first question should not be ‘How big is it?’

A simpler question may tell us much more:

Where does this money come from, and whose pocket does it ultimately come out of?

If the answer keeps circling the same few companies, we should not mistake all of the apparent scale for genuine final demand.

The size of a snowball does not prove it is solid inside.

What matters is the size of its core.

Sources

• Reuters — Nvidia’s plan to invest up to $100 billion in OpenAI and the agreement to deploy at least 10 GW of Nvidia systems, September 2025.

• IMF Annual Report 2026 — Circular financing in the AI ecosystem and interconnectedness risks arising as companies become one another’s customers, investors, and financiers.

• Financial Times — OpenAI’s revenue, spending, and free-cash-flow projections for 2026–2030, September 2026.

• Lawrence Berkeley National Laboratory / U.S. Department of Energy — United States Data Center Energy Usage Report and its 2026 update.

• IMF / Pierre-Olivier Gourinchas — The AI investment boom, increasing debt financing, and comparisons with the dot-com era.

• IMF — World Economic Outlook Update, January 2026; the effects of a potential correction in AI valuations on growth, financial conditions, and consumption.