AI Boom or Bubble? The Economics Behind the Circular AI Investment Cycle

Every major technological breakthrough brings a rush of optimism, and every so often that optimism gets ahead of itself. That's usually how a bubble starts. Once investors become convinced that a technology's future will be wildly profitable, they start bidding up valuations and pouring in money based on what they expect to happen rather than what's actually happening yet. The dot-com boom of the late 1990s is the textbook case. The internet genuinely did transform the economy, but investors got so far ahead of themselves that companies with barely any revenue, let alone profit, were pulling in enormous amounts of capital. When reality eventually caught up with those expectations, the crash wiped out billions in market value and pushed many of those companies out of business.

AI is now raising a similar question. It's already producing real products, real revenue and genuine productivity gains, but the sheer scale of investment around it has become extraordinary. Companies are pouring hundreds of billions of dollars into chips, data centres and computing capacity, much of it justified by expectations of enormous future demand rather than demand that exists today. At the same time, the money is increasingly circulating among the same small group of companies that build, finance and use all this infrastructure.

So is this simply the investment required for a transformative technology, or are expectations beginning to outrun reality?

1. The Boom Is Being Built on Expectations

What's most striking about the AI boom isn't how much money it's making today, but how much is being bet on its future. The world's biggest tech companies are pouring staggering sums into data centres, chips and computing power, with the five largest hyperscalers alone expected to spend over $1 trillion on AI infrastructure in 2025 and 2026. The logic behind this spending is fairly straightforward, if AI becomes as widely used in software, advertising, healthcare, manufacturing and finance as many expect, the companies providing the infrastructure will be in a strong position to benefit. Investing heavily now is therefore partly about making sure they are ready for that future, and not left behind if demand takes off.

Another force driving this spending has less to do with opportunity than with the fear of falling behind. For companies like Microsoft, Google and Amazon, spending too little on AI could be far more costly than spending too much. If AI does become as transformative as expected, no company wants to be the one that failed to invest. This creates a kind of arms race, where firms keep spending partly because their competitors are doing the same. When everyone is trying to secure a position in the same future market, investment can grow far beyond what current demand alone would justify. 

None of this means the spending is irrational, AI could genuinely reshape the economy and create enormous value. The problem is that the case for spending so much today increasingly depends on assumptions about how large, how fast and how profitable the AI economy will eventually become. That creates a simple cycle:

Expectations of enormous future demand → massive investment today → greater AI capacity → stronger expectations of future growth → even more investment.

The problem is that expectations can keep the cycle going for a surprisingly long time, even when the returns have not yet caught up. As long as investors and companies believe that demand will eventually grow into all this new capacity, the spending continues to make sense. That assumption becomes much harder to sustain if demand fails to grow as quickly as expected. What happens when it doesn't catch up with the investment being made today?

2. When the Industry Starts Buying From Itself

This cycle gets even more interesting once you look at where the money is actually flowing. AI companies need huge amounts of computing power, so they buy chips from the likes of NVIDIA and rent infrastructure from cloud providers. But some of these same suppliers are also investing directly in the companies that will later become their customers. Take NVIDIA and OpenAI. In 2025, NVIDIA announced plans to invest up to $100 billion in OpenAI, while OpenAI committed to deploying at least 10 gigawatts of NVIDIA systems for its next generation of infrastructure. So NVIDIA isn't just selling OpenAI equipment, it's also helping fund the company that will go on to buy it. This can also be seen with  CoreWeave, an AI cloud computing company  in which NVIDIA invested $2 billion even though CoreWeave's infrastructure runs largely on NVIDIA GPUs. 

There is nothing unusual about these companies investing in one another. Suppliers often invest in important customers, especially in industries where future demand is uncertain. And the transactions themselves are real: companies are buying chips, building data centres and recording revenue. The concern is what happens when these relationships become widespread enough to create a cycle in which suppliers help finance the very demand they are relying on.

Consider a chipmaker investing in an AI company that then uses some of that capital to buy the chipmaker's products. The sale still counts as revenue, but it does not tell us as much about demand from the wider economy. That is very different from a bank, manufacturer or retailer paying for AI because it expects the technology to make its own business more productive. In the first case, capital is circulating within the same ecosystem; in the second, new demand is coming into it.

There is evidence that this external demand remains strong. Microsoft has said that close to 90% of its cloud revenue comes from customers outside frontier AI labs, suggesting that the AI industry is not simply selling to itself. Still, as these financial and commercial ties multiply, they make it harder to judge how much of the industry's rapid growth reflects genuine adoption and how much is being reinforced by investment from within the sector. That matters because a boom supported by widespread end-user demand is much less fragile than one that depends heavily on the companies inside it continuing to fund each other's expansion.

3. When the Economics Stop Justifying the Expectations

A bubble doesn't burst just because something is expensive, it bursts when investors collectively start doubting that future returns will actually match today's prices. For AI, that doubt could show up in a few ways: productivity gains turning out smaller than promised, competition squeezing the prices companies can charge, new models making existing infrastructure obsolete, or efficiency breakthroughs that let the industry do more with far less computing power than expected.

The DeepSeek episode in January 2025 showed how quickly these expectations can change. DeepSeek released a powerful AI model that appeared to require far less computing power to develop than many investors had expected. This raised a simple but important question: if AI models can become more efficient, do companies really need to spend as much on chips and data centres as everyone had assumed? Investors reacted quickly, and NVIDIA's stock fell sharply, along with much of the broader tech sector. Nothing had changed about AI's usefulness overnight, what changed was the expected amount of money the industry would need to build it. If AI can deliver similar results with less infrastructure, the future revenues of the companies supplying that infrastructure suddenly look less certain. That is the vulnerability at the heart of the current boom: when the assumptions supporting today's investment change, the spending built around those assumptions can change just as quickly

Once companies expect lower returns, they have less reason to keep investing at the same pace. Data centre projects get delayed, chip orders slow, and cloud providers stop expanding as aggressively. Those decisions do not affect just AI companies, they also hit the businesses supplying them, from semiconductor manufacturers to construction firms and energy providers that had planned around rising AI demand. Investors then begin lowering the valuations of companies that were priced for years of rapid growth, making it harder for those firms to raise new capital. The result is that the same cycle that drove the boom starts working in reverse: weaker expectations lead to less investment, less investment slows growth, and slower growth reinforces even weaker expectations.

This is where the fallout would spread well beyond tech stocks. The AI boom now ties together chipmakers, cloud providers, data centre operators, construction firms, energy producers and financial markets, so a correction wouldn't stay confined to AI labs. It would ripple into every business that expanded to serve the boom, and every investor who priced those businesses assuming it would keep going.

Conclusion

The AI boom is not a bubble simply because valuations are high or companies are investing unprecedented amounts of money, the technology is real, adoption is growing and AI is already creating genuine economic value. The bigger question is whether the scale of investment can eventually be justified by the returns that AI generates. When companies invest based on expectations of enormous future demand, while suppliers simultaneously help finance that demand, the distinction between genuine growth and self-reinforcing optimism becomes increasingly difficult to see.

That does not mean a crash is inevitable. It means the industry has become increasingly dependent on assumptions about how quickly AI adoption will grow, how profitable it will become and how much computing power it will require. If those assumptions hold, today's spending may look perfectly rational in hindsight. If they do not, the same circular investment cycle that accelerated the boom could amplify the downturn.