Every few weeks someone asks me, usually half-joking, whether the whole AI thing is about to pop. It is a fair question, and it deserves a straight answer rather than a cheerleading one or a doom one. Here is how I read the AI bubble debate in mid-2026, from the operator’s chair rather than the trading desk.
Both things can be true
Let me give you the punchline first, because the rest of this article is evidence for it: AI can be a financial bubble and a genuine technological revolution at the same time. These are not contradictory. They are the normal pattern.
The economist Carlota Perez described how every major technology goes through an “installation” phase of frenzied over-investment and speculation, a correction, and then a longer “deployment” phase where the technology actually reshapes the economy. Railways did it. Electricity did it. The internet did it: the dot-com crash vaporised trillions in market value, and also left behind the fibre and the know-how that powered everything since. (Forbes)
So “is it a bubble?” is slightly the wrong question. The better questions are: how stretched are the asset prices, how real is the underlying value, and what survives if the prices correct? Let me take each.
The case that it is a bubble
The bull case for a bubble is genuinely strong, and honest people should not wave it away.
The capex is staggering and accelerating. Combined hyperscaler capital expenditure is heading toward roughly US$725 billion in 2026, up around 77% on 2025, with Amazon, Google, Meta and Microsoft all spending at once. Goldman has projected something like US$5.3 trillion combined over 2025-2030. (ValueAdd VC)
That spend is straining even these balance sheets. PIMCO has estimated capex could consume around 94% of hyperscaler operating cash flow by 2026-27, and Moody’s has flagged hundreds of billions in off-balance-sheet data-centre leases. (CNBC)
The revenue to justify it does not yet exist. Bain estimated the industry needs roughly US$2 trillion in new annual revenue by 2030 to fund the compute being built, and that even after AI-driven savings there is an annual shortfall on the order of US$800 billion. (Bain)
The financing is looking circular. Nvidia’s much-touted up-to-US$100bn OpenAI investment (which Jensen Huang later softened to an “invitation” rather than a commitment), the OpenAI-Oracle cloud deal, and the OpenAI-AMD GPU deal have critics describing a “money-go-round” where the same dollars inflate everyone’s headline demand. (NBC)
The grown-ups are nervous. The IMF’s Kristalina Georgieva compared valuations to the peak of the 2000 dot-com bubble; the Bank of England warned of a possible “sharp correction.” (CNBC) And “Big Short” investor Michael Burry disclosed large put positions against Nvidia and Palantir, accusing hyperscalers of flattering earnings by extending the assumed useful life of their chips. (CNBC)
If you only read that section, you would short the whole thing. So let me give the other side its due.
The case against (or at least, “not like 2000”)
Unlike the dot-com era, the leaders have real profits. The companies doing most of the spending are among the most profitable enterprises in history, funding much of this capex from genuine cash flow rather than from pure promises and IPO hype. That does not make the valuations safe, but it makes the base very different from Pets.com.
Demand is starting to catch up to depreciation. By Q1 2026, global AI sales (around US$25bn in the quarter) had just begun to exceed the associated depreciation for a second consecutive quarter. It is thin, and depreciation still eats a large share of revenue, but the line is moving in the right direction. (Bloomberg)
Physical limits act as a brake. Unlike purely financial bubbles, this one is gated by power, land, and grid connections. You cannot conjure a gigawatt of capacity overnight, which paradoxically makes a violent oversupply-driven collapse harder to engineer.
The stat that should actually worry you (and it is not a valuation)
Here is the number I care about most as an operator, and it has nothing to do with Nvidia’s share price.
An MIT study in 2025 found that around 95% of enterprise generative-AI pilots delivered no measurable P&L impact, despite US$30-40 billion invested. (Fortune)
Read that alongside the capex numbers and you see the real risk. It is not that the technology does not work. It is that most organisations have not figured out how to turn it into value. The same study found that tools bought from specialist vendors and embedded into real workflows succeeded far more often than internal build-your-own efforts. The gap is organisational: a learning and integration gap, not a capability gap.
This is the whole thesis of a piece I wrote earlier, why 95% of AI pilots fail, and it is why I keep pushing clients toward the honest ROI numbers rather than the demo magic. If a correction comes, it will not be because AI is fake. It will be because too much money got spent by organisations that never crossed that integration gap.
The Australian angle: check your super
For Australians, the most direct exposure to an AI correction is not your job. It is your superannuation.
- Around 20% of the roughly A$4.3 trillion super system is invested in US companies, and recent strong returns (the median growth fund returned around 9.5% in FY2026, a fourth straight above-average year) have leaned heavily on US AI exposure. (Yahoo Finance AU)
- Some big funds are already de-risking. Colonial First State has been reviewing its US-tech exposure, and AustralianSuper’s CIO publicly lamented a “late tilt to AI.” (Bloomberg)
- The RBA has said the quiet part. In its March 2026 Financial Stability Review, it warned that “a correction in any major market would not stay offshore” and would transmit rapidly into Australian funding costs. Others have flagged “hidden AI concentration risk.” (RBA)
None of this is financial advice, and I am not qualified to give it. But the sensible, boring action is to ask your fund a direct question: how concentrated am I in the largest AI names, and what happens to my balance if they fall 30%? You may be comfortable with the answer. You should at least know it.
What I actually do about it
I am neither buying the hype nor shorting the future. As an operator, here is the posture that makes sense regardless of whether prices correct next quarter or in three years.
- Build capability, not exposure. The durable value is not owning AI stocks at these prices. It is having real, workflow-embedded AI capability in your organisation. That survives any financial correction, just as the fibre survived the dot-com crash.
- Land in the 5%, not the 95%. Integrate AI into back-office and high-volume workflows where the ROI is measurable, buy proven tools rather than over-building, and measure outcomes ruthlessly. The AI factory mindset is exactly this.
- Assume the tools get cheaper, not more expensive. If there is a correction, compute and models get cheaper and more abundant, not less. An organisation with real AI capability and a cost-disciplined approach is a winner in that scenario, not a loser.
So, is the AI bubble real? Probably, in the narrow sense that some asset prices are stretched and a correction is plausible. Is AI real? Unambiguously yes, and the correction, if it comes, will not change that. The dot-com crash did not un-invent the internet. It just cleared out the companies that had confused a valuation for a business. My job, and yours, is to make sure we are building the second kind.
Amjid Ali is a technology and AI leader based in Melbourne who helps organisations build durable AI capability rather than chase AI hype. If you want to land in the 5% that gets real value, get in touch.