Is the AI industry in a financial bubble created by NVIDIA?

NVIDIA’s profits give it an extraordinary ability to bring the AI economy forward. As it helps finance its customers, more of the investment around AI depends on the same judgement: that profitable demand will arrive quickly enough.

I’ve been doing a lot of reading about NVIDIA recently, including all the investments it’s making in Australia. I thought I’d take a global view of the company and where it is at. What interests me most is how far the company’s role now extends beyond selling chips. It is putting its financial strength behind the businesses that will buy and use them, and helping determine how quickly the rest of the economy gets access to AI.

NVIDIA has a highly profitable business, customers willing to pay for its technology and the resources to accelerate a market in which it already leads. There is a sensible commercial argument for using those resources to help the market grow.

The complication is that the supplier is also becoming an investor in its customers and, under some agreements, a buyer of their unused capacity. If demand takes longer to develop than expected, NVIDIA can feel the effects in several places at once. The same applies to investors and lenders whose portfolios depend on that expansion.

I’m old enough to have lived through the dot com bubble and crash, and saw how Cisco played its role and navigated through it. It’s their experience that’s also worth revisiting to see how closely it aligns with where we are at with NVIDIA. It shows how a successful infrastructure company can correctly identify a technological change and still misjudge the spending cycle around it. But there are good reasons to take NVIDIA’s present strength seriously before reaching for the dotcom comparison.

The profits give NVIDIA room to move

For the quarter ended 26 July 2026, NVIDIA reported US$96.2 billion in revenue, up 106% on the previous year, with a 75% gross margin. Operating cash flow for the first half was US$74.4 billion. The scale is remarkable. This business is already generating the money with which to pursue its ambitions. ^[1]

That allows it to accept risks and payback periods that a smaller company could not. It can support customers while their businesses develop, invest ahead of demand and influence which projects get financed. Those choices can be valuable even when a project’s return takes years to emerge.

Its balance sheet is substantial too, although the composition matters. At the end of that quarter NVIDIA held US$22.4 billion in cash and cash equivalents, US$34.1 billion in marketable debt securities and US$42.8 billion in marketable equities. Together they came to roughly US$99.4 billion. Let’s just call that an even US$100 billion. Calling all of that cash would overstate the money immediately insulated from a fall in share prices. ^[1]

An equity portfolio can lose value at the same time customers need more support. Even allowing for that possibility, the company has considerable resources and earning power. I would be wary of an argument that treats every new financial commitment as a sign of weakness without considering the business funding it.

NVIDIA is helping its customers obtain the money

In August, NVIDIA announced agreements with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR intended to mobilise more than US$500 billion of third-party capital for AI infrastructure over time. The proposed platforms remain subject to definitive agreements. It is an ambition for outside investment, not US$500 billion that NVIDIA has already spent or promised to supply itself. ^[2]

There is a practical problem for those platforms to solve. Being an AI provider is capital intensive. It needs expensive GPUs, buildings and power before it can earn revenue from customers. The demand may be promising, but a lender still wants assurance that someone will pay. Support from the company supplying the equipment can make a project easier to finance and bring useful capacity into service sooner.

CoreWeave’s September 2025 securities filing gives a specific example. It describes an order with an initial value of US$6.3 billion under which NVIDIA purchases specified remaining unsold computing capacity through April 2032, subject to the agreement’s conditions. NVIDIA is also identified as a supplier and shareholder. ^[3] Sounds a little circular to me.

The commitment provides support for part of CoreWeave’s future revenue. If customers buy the capacity, the computing service earns its money from them. If the specified capacity remains unsold, NVIDIA has an obligation to purchase it under the contract. The arrangement can help get infrastructure built, while returning some of the risk of weak demand to the supplier.

NVIDIA’s involvement in Ohio extends further into the physical infrastructure. It disclosed credit support for SB Energy’s campus, a US$1.5 billion investment in SB Energy, and facilities intended for an OpenAI lease, with the first phase expected in 2028. Its interests now reach into the land, power and buildings needed to operate the equipment. ^[4]

I can see why NVIDIA would do this. Waiting until every future customer can finance everything independently could leave the industry growing more slowly than the technology allows. Infrastructure often has to come first. A company with NVIDIA’s resources can help close that gap and earn a commercial return from doing so.

Eventually, though, the services running on the equipment must bring in enough money to justify it. Financial support buys time for demand to develop. The investment succeeds when customers find enough value in AI to keep paying for it without requiring ever more support from the businesses selling it to them.

Cisco was a successful business too

Cisco came through the dotcom crash, but getting to the other side meant absorbing a sharp fall in sales and billions of dollars in inventory losses. Cisco was profitable. In fiscal 2000 it reported US$18.9 billion of revenue and US$2.7 billion of net income. Yet by the fourth quarter of fiscal 2001, its quarterly revenue was down 25% from the previous year. The internet’s long-term importance did not prevent a sharp reversal in equipment spending. ^[5]

Cisco’s inventory accounts show what that reversal meant. Its 2001 annual report records a US$2.25 billion additional excess-inventory charge after stock and purchase commitments had accumulated during shortages. US$105 million of the affected inventory was scrapped in fiscal 2001. A subsequent filing records another US$975 million scrapped in fiscal 2002. Other stock was used or sold, and purchase commitments were settled. ^{[6] [7]}

I vaguely remember back in 2001 (which I have just subsequently investigated) that Cisco literally crushed equipment to restore the balance between supply and demand. The filings confirm more than US$1 billion of scrapping across those two years. They do not establish how the equipment was destroyed, where it went or a motive of supporting prices. Those parts of the story remain unproven.

Deliberately destroying usable equipment to defend its price would be hard to justify today. But I would not attach that motive to Cisco on the strength of a rumour. The documented inventory loss is enough to make the economic point: resources had been committed to demand that did not arrive on the expected timetable.

That is the danger I take from the comparison. A technology can become essential over a decade while the equipment market becomes oversupplied next year. Being right about the destination does not settle how much capacity customers can profitably use along the way.

The differences matter

NVIDIA’s current profits give it considerable capacity to absorb setbacks. That deserves weight in any comparison. Cisco was also profitable before its correction, so profitability alone cannot settle the argument, but neither should a historical resemblance obscure the resources available to NVIDIA today.

The commitments themselves also differ. An equity investment exposes the money invested. A conditional purchase agreement creates obligations under particular circumstances. A proposed platform for outside capital is another arrangement again. Adding all their headline values together as though they were NVIDIA debts falling due tomorrow would produce a misleading picture.

Let’s look at the timeline, as it does give us some indicators. CoreWeave’s disclosed term runs through 2032 and Ohio’s first phase is expected in 2028. The useful questions concern the conditions under which payments arise, how those payments are distributed over time and what cash the business can generate during that period.

There may also be value left in equipment after its first owner’s plans disappoint. A processor can remain useful to a different customer or workload. But continued usefulness does not guarantee the rental income, resale value or operating life assumed when someone financed it. A useful asset can still be a poor investment at the price paid.

Competition creates a further reason for NVIDIA to support a broader customer base. Google’s Ironwood processor is an example of a large AI customer developing its own hardware for inference. More AI use does not necessarily translate into the same share of chip purchases for NVIDIA. ^[8]

My reading is that helping independent computing providers expand makes strategic sense here. It gives more businesses the means to compete and creates buyers beyond the largest customers developing alternatives of their own. NVIDIA’s financing activity has a commercial purpose beyond keeping this quarter’s sales high.

Why this reaches the wider economy

The effects of AI investment are already broader than the sale of processors. Constructing a data centre creates work and orders before the services it will deliver have earned their long-term return. Investors finance those buildings and equipment. Share prices affect household wealth, and expectations of further growth influence spending elsewhere.

On 10 September, the Bank for International Settlements drew attention to concentrated equity valuations, growing debt and private-credit exposure, and financial connections between chip suppliers, cloud operators and AI businesses. It examined how disappointing returns could reverse investment and carry the effects through the financial system. ^[9]

A portfolio can contain several apparently different investments that rely on the same outcome. Shares in a supplier, a loan to a computing provider and an interest in a data-centre building are different assets. If all three need rapid growth in paying AI customers, their risks may be much more closely connected than their labels suggest.

This is why NVIDIA’s financing role matters even to people who never choose to buy its shares directly. As the company helps more projects proceed, its confidence in future demand supports more spending elsewhere. Successful projects can strengthen growth. This structure, if it goes pear shaped, can affect creditors, construction and employment well beyond NVIDIA’s own accounts.

NVIDIA could remain a successful company through a correction that leaves other investors with substantial losses. The economic value of AI, the returns on particular assets and the fortunes of the strongest supplier need not move together.

The return has to come from using AI

More available computing can make useful services cheaper, give smaller businesses access to expertise they could not previously afford and allow work that was uneconomic to become worthwhile. NVIDIA has the means to help that happen sooner.

For the economy, that is the return that matters. Building another facility adds activity today. Using its capacity to deliver something customers value creates the income that can sustain the investment. Over time, the evidence of success should become easier to find in businesses outside the infrastructure industry.

The risk is delay. Demand does not have to disappear for a project to run into trouble. If profitable uses develop more slowly than expected, too much capacity can compete for the same customers while the financing costs continue. NVIDIA’s cash generation gives it more room than most to endure that period, but its willingness to support customers also increases its exposure to it.

I see a company strong enough to accelerate a real economic opportunity, choosing to commit more of its own resources to the outcome. That is a reason for confidence in its position, with close attention to how quickly the investment becomes productive elsewhere.

Cisco’s experience should make us careful about the timetable. NVIDIA’s strength gives us a reason to believe it can help improve it. The bet now reaches far enough into markets and economic activity that the result will matter to a great many people who never bought a chip.