The Symmetry Problem
In the first half of 2025, OpenAI spent $5.02 billion running its models on Microsoft's Azure cloud infrastructure . Over the same period, the company was on track to generate roughly $13 billion in annual revenue —a respectable sum by any ordinary measure, except that OpenAI had just committed to pay Oracle $300 billion over five years for additional computing power [8,13], agreed to hand CoreWeave $11.9 billion across the same timeframe for data centres and services , and signed deals with chipmakers to secure 900,000 semiconductor wafers per month . The arithmetic produces a certain vertigo. By some reckonings, OpenAI requires $1.4 trillion worth of compute to realise its vision , a figure more than a hundred times its current annual revenue. The company's leadership believes that demand for ever-better iterations of its products will eventually pay off this staggering bill. Whether they are right is the question on which trillions of dollars now rest.
This is not hyperbole. Across the past eighteen months, the AI sector has attracted investment that exceeds the capital committed to any recent technology boom . OpenAI alone has racked up $1 trillion in deals [8,9], a sum so large it requires a moment's pause to comprehend. The company closed a funding round in late 2024 at a post-money valuation of $852 billion , making it one of the most richly valued private enterprises in history. Meanwhile, Nvidia—whose graphics processors have become the essential infrastructure of the AI age—became the world's first company to reach a $5 trillion market capitalisation [5,6,7], joining Microsoft and Apple in the rarefied territory above $4 trillion [2,3,4]. The numbers are dizzying, and they raise an uncomfortable question: are we witnessing the birth of a new economic order, or the inflation of a bubble whose collapse will reshape the global economy?
The Deals That Ate the World
To understand the scale, it helps to catalogue the contracts. OpenAI's $300 billion commitment to Oracle is only the largest in a cascade of agreements that have redefined corporate spending [8,13]. Meta, Facebook's parent company, has pledged over $10 billion to Google for cloud services across six years , and separately contracted with CoreWeave—a GPU cloud provider that emerged from the cryptocurrency mining industry—to provide $14.2 billion of AI infrastructure . CoreWeave's stock surged 12 per cent on the announcement , a notable vote of confidence in a company that barely existed in its current form a few years ago. Broadcom, a chipmaker previously known for networking semiconductors, has entered a joint venture with OpenAI to build and deploy 10 gigawatts of custom AI accelerators , a deal that sent Broadcom's stock up 9 per cent .
These transactions share a peculiar structure. In many cases, the same companies appear on both sides of the ledger. Microsoft, for instance, is OpenAI's largest investor and also its primary cloud provider; OpenAI pays Microsoft billions for compute, which Microsoft books as revenue, a portion of which flows back to Microsoft as OpenAI's backer . Google sells cloud services to Meta whilst competing with Meta in AI product development . Oracle provides infrastructure to OpenAI even as it positions itself as a rival in enterprise AI . The result is what some observers have termed a "circular flow of capital" —a system in which money moves in loops, inflating valuations without necessarily producing corresponding real-world growth. Portfolio managers at major investment funds have begun to voice concern , noting that the metrics by which AI companies are valued bear little resemblance to traditional measures of profitability or cash flow.
The scale of investment has material consequences. Micron, one of the world's largest memory manufacturers, recently announced the discontinuation of its Crucial consumer product lines, citing a RAM shortage driven in part by the voracious appetite of AI data centres . OpenAI's agreement with Hynix and Samsung for 900,000 wafers per month represents roughly 40 per cent of global DRAM production, according to some estimates —a level of demand that has sent memory prices soaring and constrained supply for other industries. The scarcity is expected to worsen , a sign that the AI boom is no longer an abstract financial phenomenon but a force reshaping supply chains and hardware markets worldwide.
The Revenue Question
Against this backdrop of trillion-dollar commitments, OpenAI's actual revenue appears almost quaint. The company's $13 billion annual run rate is impressive by the standards of a startup, but it pales beside the obligations it has assumed. The gap invites scepticism. If OpenAI is spending $5 billion every six months on inference alone —the process of running trained models to answer user queries—and committing tens of billions more to secure future capacity, when does the business become self-sustaining? The company's answer is that today's spending is an investment in tomorrow's dominance. OpenAI's models are already embedded in products used by hundreds of millions of people, from ChatGPT to Microsoft's Copilot suite. Each new generation of models expands the range of tasks they can perform, opening fresh revenue streams. The theory is that network effects and switching costs will entrench OpenAI's position, allowing it to capture value commensurate with the infrastructure it is building.
But the theory depends on a set of assumptions that remain untested at scale. Chief among them is the belief that demand for AI services will grow exponentially, and that customers will pay prices sufficient to cover the staggering cost of the underlying compute. So far, much of the revenue in the AI sector has come from enterprises experimenting with the technology rather than integrating it into core operations. The leap from pilot projects to production workloads is not guaranteed. There is also the question of competition. OpenAI is not the only company pouring capital into AI infrastructure. Google has described its Gemini model as "a fundamentally new kind of AI" and "the company's most powerful to date" , a signal that it intends to compete aggressively. Meta, Amazon, and a host of startups are building rival systems. If the market fragments, the economies of scale that justify today's investments may fail to materialise.
"OpenAI needs $1.4tn worth of compute, a number dwarfing its current revenues, because it believes demand and ever better iterations of its products will pay it off."
The stakes extend beyond any single company. If OpenAI cannot make the economics work, the ripple effects will be profound. Investors have poured capital into AI startups on the assumption that a few winners will emerge with sufficient market power to recoup their expenditures. If that assumption proves false, the resulting write-downs could rival those of the dot-com collapse. Portfolio managers have begun to draw the comparison explicitly , noting that the current frenzy shares features with past bubbles: soaring valuations, speculative capital flows, and a disconnect between present performance and future expectations.
The Nvidia Singularity
No company embodies the AI economy's contradictions more fully than Nvidia. In early 2025, its market capitalisation surpassed $5 trillion [5,6,7], a milestone that reflects both the centrality of its products and the expectations now priced into its stock. Nvidia's GPUs—originally designed to render video game graphics—have become the default hardware for training and running large AI models. The company has leveraged this position into a near-monopoly, capturing an estimated 80 per cent of the AI accelerator market. Its revenue has grown in tandem with the AI boom, but its valuation has grown faster still, implying that investors believe the company's dominance will persist and deepen.
The risk is that Nvidia's position is less secure than it appears. Broadcom's deal with OpenAI to build custom accelerators is a sign that large AI developers are seeking alternatives to off-the-shelf GPUs, which remain expensive and supply-constrained. Google has long used its own Tensor Processing Units for internal workloads, and Amazon has developed custom chips for AWS. If OpenAI, Meta, and others succeed in building bespoke hardware, Nvidia's margins could compress. The company's $5 trillion valuation assumes not only that AI spending will continue to grow but that Nvidia will capture a stable share of that spending. Both assumptions are contestable.
Yet Nvidia's rise is also a testament to the reality of the AI transformation. The company's revenue growth is not speculative; it is shipping silicon in quantities that strain global manufacturing capacity. The deals OpenAI and others have signed are not hypothetical; they commit real capital to real infrastructure. The AI economy may be overheated, but it is not a mirage. Data centres are being built, chips are being fabricated, and models are being trained at a scale that would have seemed fantastical a decade ago. The question is not whether AI is transformative—it plainly is—but whether the financial structure being erected on top of that transformation is sustainable.
The Circular Capital Problem
The phrase "circular flow of capital" captures a unease that has taken root among investors and analysts. In a conventional market, capital flows from investors to companies, which use it to build products and services that generate revenue from customers. The revenue covers costs, produces profit, and returns to investors as dividends or capital gains. The system is linear and accountable. But the AI economy increasingly resembles a closed loop. Companies invest in each other, buy services from each other, and book the transactions as revenue, which justifies higher valuations, which attract more investment, which funds more deals. The loop can sustain itself for a time, but it is vulnerable to any shock that breaks the circuit.
This is not to suggest fraud or malfeasance. The deals are real, the services are delivered, and the companies involved are building genuine technological capabilities. But the structure creates a kind of self-reference that obscures underlying value. When OpenAI pays Microsoft for compute, and Microsoft counts that payment as revenue, and Microsoft's valuation rises on the strength of that revenue, and Microsoft's higher valuation justifies further investment in OpenAI, the system begins to resemble a perpetual motion machine. It works until it doesn't.
The concern is amplified by the sheer scale of the commitments. OpenAI's $1 trillion in deals [8,9] is not spread across a diversified portfolio of experiments; it is concentrated in a handful of enormous bets on infrastructure that will take years to deploy and longer still to pay off. If demand falters, or if competition erodes margins, or if the models plateau in capability before reaching the thresholds needed to justify the spending, the losses will be commensurate with the scale of the investment. CoreWeave's $14.2 billion deal with Meta and $11.9 billion deal with OpenAI have turned a once-obscure GPU cloud provider into a linchpin of the AI economy. If either customer scales back, CoreWeave's business model collapses, and the investors who drove its stock up 12 per cent will face steep losses.
The Historical Echo
Comparisons to the dot-com bubble are inevitable, and not entirely fair. Jerome Powell, chair of the US Federal Reserve, has argued that AI is fundamentally different from the late-1990s internet frenzy, describing it as "a major source of economic growth" rather than speculative excess . Powell's point is well-taken: AI is already producing measurable gains in productivity, from drug discovery to logistics optimisation. The technology is not vaporware. But the dot-com bubble was not built on nothing, either. The internet did transform the global economy, just not quickly enough or profitably enough to justify the valuations of 1999. Many of the companies that embodied the first internet wave collapsed; the infrastructure they built was eventually put to use by a second generation of firms that emerged from the wreckage.
The parallel suggests that even if AI fulfils its transformative potential, the current financial structure may not survive the transition. Bubbles do not require the underlying technology to fail; they require only that expectations outrun reality by a sufficient margin. The $1.4 trillion in compute that OpenAI believes it needs may eventually prove justified, but if the revenue needed to support that infrastructure takes a decade to materialise rather than five years, the gap will bankrupt companies and vaporise fortunes. The question is not whether AI is real but whether the timeline implicit in today's valuations is plausible.
Some analysts have begun to argue that the bubble will burst imminently, with predictions that 2025 will be the year of reckoning . The logic is straightforward: the gap between investment and revenue is widening, not narrowing, and at some point investors will demand evidence that the spending is producing returns. When that moment arrives, the circular flow of capital will reverse. Companies that relied on ever-rising valuations to fund operations will find themselves unable to raise fresh capital. Deals will be renegotiated or cancelled. Stock prices will correct, and the correction will be severe.
The Path Forward
Yet it is also possible that the sceptics are wrong. The AI economy's scale and speed are unprecedented, but so is the technology's capability. OpenAI's models can now perform tasks—writing code, conducting legal research, generating synthetic media—that were the exclusive domain of highly trained professionals just a few years ago. Each new generation of models expands that frontier. If the trajectory continues, the addressable market for AI services is not millions or billions but trillions of dollars, encompassing entire sectors of the global economy. In that scenario, today's spending is not irrational exuberance but rational preparation.
The difficulty is that both narratives are plausible, and the outcome will not be clear for years. In the meantime, the AI economy will continue to evolve in ways that defy easy categorisation. Memory shortages will worsen , forcing industries to compete for scarce resources. New deals will be announced, each larger than the last, pushing the boundaries of what seems financially possible. Nvidia's valuation will fluctuate, a real-time barometer of investor sentiment [5,6,7]. And OpenAI, the company at the centre of the maelstrom, will continue to spend sums that dwarf its revenue [8,16], betting that the future will validate the present.
What is certain is that the AI economy has reached a scale where its internal dynamics matter to everyone. The capital being deployed is not confined to a speculative corner of the market; it is reshaping semiconductor supply chains, electricity grids, and labour markets. If the spending pays off, the gains will be transformative. If it does not, the losses will be catastrophic, not only for investors but for the industries and workers caught in the downdraft. The trillion-dollar wager is no longer a side bet. It is the main event, and the outcome will define the next decade of global economic life.