Should Enterprises Worry About an AI Bubble?
A review of the arguments around over-valued AI firms and what this might mean for enterprise AI based on lessons from previous stock market bubbles
In mid-August, Portugal is in beach mode and Europe is in peak holiday season, so rather than focus on the minutiae of enterprise AI this week, I will slip in a more meta piece about the relationship between AI investment and global economic and geo-political risks to reflect on what this might mean for enterprise AI investment strategy.
TL;DR: If there is a bubble (not convinced), its blast radius will be limited, and it will leave behind tech and infrastructure at a capability level and price that could transform businesses who apply it in the right way.
The 1873 railroad bubble
The (UK) Times newspaper recently wrote about the Microsoft CEO, among others, reading up on the 1873 global financial panic associated with an investment bubble in railway development, which is perhaps the closest historical analogue to today’s so-called ‘AI bubble’. Roughly $3bn had been invested in US railroads between the end of the civil war and 1873, with a third of that coming from European investors. In total, this was equivalent to roughly 30% of the USA’s annual national output at the time, which means it was more substantial than the AI investment we have seen to date. Railroads were a powerful general purpose technology, and like AI and data centres today, they catalysed downstream innovation and created lots of economic opportunities beyond just the railroad companies themselves.
By 1873, panic about this investment bubble in Europe triggered a crash that led to the Austrian stock exchange losing 45% of its value in a day; global prices fell by about 30%, and a recession followed that lasted until the end of the decade.
But the railroads were still there after the crash, and this new infrastructure was vital in the next phase of social and economic development. The technology was real, even if the bubble around it tended to overvalue the early pioneers who built it.
Something similar happened more recently with the dotcom crash and telecoms firms who thought owning the pipes would mean also owning the value chains they spawned. Without that crash we would not have had the social technologies that followed.
Is there a bubble and, if so, who might be exposed?
Today, there are lots of AI doomers who believe the level of spending on AI in the pre-profit stage combined with the vertigo-inducing valuations of AI-related companies will probably lead to a stock market crash. But the evidence does not yet fully support these fears, and AI is not the only reason for a frothy stock market.
However, if there is even a partial retrenchment, we might yet see some major players over-reach and suffer the consequences.
Oracle’s Larry Ellison is doing a good impression of a regime-linked Russian oligarch, buying up media companies that the emperor wants to de-fang, regardless of the price. But he has also built up astronomical levels of debt to turn Oracle into an AI hyperscaler to avoid being left behind by the AI bubble. As this long New York Times profile suggests, funding these two big bets at the same time using various forms of debt is a bold and very risky strategy.
As Mike Brock put it recently:
Every wing of the House [of Ellison] rests on one column: Oracle equity. The equity rests on the OpenAI contract. The contract rests on the AGI story, and the story is being repriced out of Hangzhou at 87 cents per million tokens. A federal judge holds the Warner deal; twelve attorneys general hold the lawsuit; S&P Global holds the rating one notch off the floor. Underneath all of it, an eighty-one-year-old man holds a $40.4 billion promise, irrevocable by its own terms, written against a stock that has lost two-thirds of its value since the morning the promise became imaginable.
Microsoft suffered a long pull-back in its stock price this year based on fears that its future revenue was too dependent upon OpenAI, but its fundamentals look solid and it is hard to imagine the company not being a huge beneficiary of enterprise AI one way or another.
Google has recently been criticised for falling behind in the frontier AI race to AGI, but its strategy has arguably been misunderstood, and it seems to be doing very well indeed from AI in search and its cloud platform, even if its models are not the very best. A recent wave of senior departures from DeepMind and the engineering team suggest that Google is betting on diffusion over invention in the next phase of AI development, and is counting on its Google Cloud Platform to lead this charge.
Tim O’Reilly chimed in on what this might mean for Google, referencing another historical example that suggests the firm is taking a leaf out of the Westinghouse playbook in building on the inventions of Nikola Tesla and others:
Jeff Ding’s book Technology and the Rise of Great Powers traces the relative impact of invention and diffusion during technology revolutions. Ding argues that nations that dominate the “leading sector” of a general purpose technology don’t reliably grow more powerful as a result. Diffusion is the defining factor. He posits that Britain’s edge in the first industrial revolution came less from inventing the steam engine and advances in steelmaking than from diffusing machinery through the whole economy so that many businesses, not just the steam engine manufacturers and the steelmakers, became more profitable.
Nvidia continues to innovate in financial engineering as well as hardware and software, forming an investment partnership to unlock $500bn of lending for firms to build more data centres and buy more GPUs, offering chips as collateral in addition to contracted cashflows, and underwriting up to 25% of the debt itself. And, at the same time, it is launching its own open model, despite its dependence on OpenAI and its expensive frontier models for future revenues.
But whilst the hyperscalers and chip makers are at risk if there is indeed an AI bubble, it is large-spending firms with no other revenue streams that are most vulnerable, such as OpenAI, Anthropic and their competitors, GPU neoclouds and other AI infrastructure companies.
Noted AI doomer Ed Zitron is convinced the first big domino to fall will be OpenAI, and that will lead to a cascade effect.
But would it?
Leverage multiplies the risk
Beyond the fate of individual tech firms, is there a risk that the US economy, and particularly its stock market, is already so pumped and leveraged by debt that even a partial deflation of the AI boom could have much wider knock-on effects for the global economy?
This question is made more interesting by the high degree of systemic concentration and counter-party risk that is accumulating due to the circular nature of much AI pump-priming investment, or what people are calling round-tripping.
Hyperscalers (e.g., Microsoft, Amazon, Google) and chipmakers (Nvidia) invest billions in vendor financing for frontier labs and GPU neoclouds, but these agreements typically include covenants requiring the recipient to spend the capital on the investor’s cloud compute or hardware. Hyperscalers recognise these expenditures as top-line cloud growth, while Nvidia books hardware sales when hyperscalers and neoclouds buy GPUs to fulfil those compute commitments. These rising top-line revenues lift the stock price of Nvidia and the hyperscalers, providing additional paper wealth and operating cash flow to reinvest into the next round of financing. It’s a circle!
So this means if investors lose confidence in one of the frontier model firms, the collateral damage could propagate back towards the hyperscalers and chip makers as well.
The Financial Times today leads with news of a global bond sell-off, which it partly ascribes to newly created debt instruments designed to fund the AI capex build-out over the next few years. The risk here is not AI bonds per se, but also the wider context of the Trump regime playing fast and loose with US debt, combined with the failing Iran war and other policies that increase the risk of other countries deciding to — or even just threatening to — sell US Treasuries.
Long-term borrowing costs across major economies hit multi-decade highs on Tuesday as inflation concerns, deficit fears and surging AI bond issuance put pressure on government debt around the world.
The recent debacle at the AI-focused hedge fund Situational Awareness, which ran a concentrated book of AI investments with an estimated 4x leverage provided by lenders, is a good example of the way leverage can increase the risk of AI investment. A sharp 25-40% fall in some of these stocks last month led to the fund being margin-called and ultimately selling off a large part of their portfolio cheaply to avoid being wiped out, with an estimated total loss of $30-35bn.
But given the long-term nature of the infrastructure investments AI needs today, such as data centres, compute and model development, it is not surprising that hyperscalers and other big spenders are looking to borrow to fund much of this work. These things take a long time to build, and even longer before they produce returns; plus estimating forward demand and capacity needs is very hard to get right.
The question is, will all this spending and borrowing produce long-run returns for the firms investing most heavily — and more broadly, will the total aggregate spend produce enough value for the economy as a whole to have been worthwhile?
Returning to historical analogies for a second, the Economist recently likened the AI investment boom to canal mania, railroad mania, the roaring 20’s and the dotcom boom, and concluded that projected revenues lag projected AI costs by a significant margin, and would require AI to start delivering a clear improvement in business productivity, and also a substantial increase in the intangible capital that firms expend on re-tooling their structures and processes.
On the other hand, Azeem Azhar’s team has produced some very thorough analysis that suggests expected AI revenue growth should be just enough to cover capex and depreciation in this early phase of infrastructure spending, even if the majority of current spending is now coming from borrowing rather than free cashflow.
What about open models?
If we zoom out further, we might conclude that most of the arguments and analyses above relate most closely to the US-led AI market and its impact on the stock market (albeit with high levels of contagion).
But it is also worth considering China’s AI strategy and how this acts as a counter-balance, whilst creating a cost floor that limits the blast radius of any frontier model collapse.
When we consider whether the AI bubble might lead to a stock market crash, one piece of evidence that suggests this risk might be lower than it seems comes from an analysis of the rise of Chinese open models and their economic impact in a recent paper by David Krause titled The Open-Source Chinese AI Shock: Fear, Volatility, and the Repricing of Tech Giants, which found the resulting market volatility was more pronounced in the tech / software sector than in the wider economy, suggesting (but not proving) that the impact of such events might be containable.
China’s promotion of open models acts as a brake on what frontier models can charge, given their decreasing lead against cheaper models. But beyond model and token costs, China is also pushing forward with the application layer that will turn raw intelligence into practical applications that companies can use to make or save money.
Writing in the Financial Times, Lizzi Lee, a fellow at the Asia Society Policy Institute’s Center for China Analysis, makes the case that Chinese AI will represent a fourth ‘China Shock’ that could be even more impactful than the previous waves of innovation in manufacturing, cleantech and ecommerce.
The momentum behind China shock 4.0 is strong. Geopolitical constraints are helping to crystallise a new and distinct model of technology governance. The next shock will not arrive in a container ship. It will spread via the principles that surround Chinese AI technologies. The world is not prepared for what will come next.
Just in this past week, we have seen another great example of how the ecosystem approach of China could help advance enterprise AI more effectively than the US focus on ever-stronger frontier models.
DeepSeek Harness seeks to improve the composability of AI agents and the tools built around them, with an architecture that treats everything as a plugin, hinting at a shareable ecosystem of add-ons. Launched as an open-source, MIT-licensed agent runtime and orchestration framework, it is designed to decouple the underlying language model from the execution, tooling, and governance layer of autonomous agents, which gives enterprises the option of a standardised, vendor-neutral agent runtime system.
What does this all mean for enterprise AI?
Things like DeepSeek Harness are good news for enterprise AI, as it promotes an open ecosystem and helps advance the application layer that is so under-developed right now, with frontier models, chips and infrastructure taking the lion’s share of investment funding.
Even for organisations that are more comfortable with US than Chinese government interference in AI, this ecosystem approach to open models, open standards and greater composability means that they will have much greater choice, and probably much cheaper tech in the future, even if they choose not to use Chinese models.
Should large firms be worried about an AI bubble bursting? I don’t think so, unless they are already vulnerable by being over-leveraged or otherwise exposed to stock market fluctuations. The tech is not going away, and it is likely to get a lot cheaper, especially if the hyperscalers over-invest in capacity in this pre-profit stage. But token pricing in particular still needs to be managed carefully and reviewed regularly as models and their capabilities change.
AI models are necessary but not sufficient for the intelligent enterprise, and the kind of intelligence each individual agent or component needs today is far below the level of frontier models. The best agentic operating systems will be granular, standardised and composable, with lots of small, specialised intelligent agents working together to create emergent outcomes. This does not need frontier models (and pricing) to be effective, and I would expect a lot more locally-hosted small and open models to do much of the repeatable work.
But in the face of tech sector volatility, it makes sense to reduce vendor lock-in and over-dependence on individual SaaS platforms, especially when agentic AI is just getting to the stage where it can be used to design and build your own workflow and process systems.
Even setting aside cost arguments, there is a strategic need to own more of your own organisational operating system rather than relying on rented platforms that enforce their own ways of working.
If there is a bubble and it pops, whether it is caused by the market over-valuing AI firms or by more craziness from Washington DC, well-run firms that are not over-reliant on renting intelligence from one supplier will adapt and thrive. The railroads survived the 1873 crash, and the dotcom crash was followed by a blossoming of cheaper, smarter tech that used the expensively-laid pipes for all kinds of purposes that were not considered viable just a few years previously.