In 2025, OpenAI reportedly generated approximately $13 billion in revenue. Around the same period, Sam Altman was describing ambitions to develop 30 gigawatts of artificial intelligence computing capacity, requiring an estimated $1.4 trillion in infrastructure investment. The difference between those numbers was extraordinary: a company earning billions was helping set in motion infrastructure plans worth trillions.
The announcements raised an uncomfortable question. How could OpenAI afford such enormous commitments, particularly when many of the companies supplying its computing power were also investing in OpenAI itself? Critics saw the beginnings of a financial bubble, sustained by interconnected investments, ambitious revenue projections and a technology industry increasingly dependent on its own expectations of future growth.
But the most alarming interpretation of the numbers is also misleading. OpenAI does not have a single $1.4 trillion invoice waiting to be paid. Its infrastructure ambitions involve a complicated collection of cloud contracts, chip purchases, proposed investments, conditional incentives, construction projects and financing arrangements spread across multiple companies and several years.
That distinction matters, but it does not eliminate the financial risk. By October 2026, OpenAI had raised enormous amounts of capital, expanded its infrastructure partnerships and reportedly reached an annualised revenue run rate of approximately $50 billion. Meanwhile, the companies building the infrastructure around it were taking on commitments whose financial success depended on AI demand continuing to expand at an exceptional pace.
The real question is not whether artificial intelligence works or whether OpenAI can raise enough money to announce another data centre. It is whether paying customers will eventually generate enough economic value to justify the infrastructure being constructed in anticipation of their arrival.
The Deals Behind the $1.4 Trillion Headline
Understanding OpenAI’s infrastructure strategy requires separating the different financial arrangements behind its headline numbers. A proposed equity investment is not equivalent to a contractual purchase obligation, while a data-centre development project is not necessarily an asset that OpenAI must buy or finance directly.
The scale of the agreements is nevertheless remarkable. Several of the world’s largest technology companies have entered arrangements involving OpenAI, with announced capacity targets and financial values that would have seemed extraordinary for a relatively young company only a few years earlier.
| Partner or project | Announced arrangement | Financial meaning |
| Stargate | Up to $500 billion and 10 gigawatts | Multiyear infrastructure development involving several investors and technology partners |
| Nvidia | At least 10 gigawatts, with intended investment of up to $100 billion | Proposed investment tied to progressive deployment of Nvidia computing systems |
| AMD | Six gigawatts of AMD GPUs | Multiyear hardware agreement with conditional equity warrants |
| Broadcom | Ten gigawatts of custom AI accelerators | Chip development and deployment partnership without an officially disclosed $350 billion contract price |
| Microsoft | Additional $250 billion in Azure services | Long-term cloud-computing purchase commitment |
| Amazon Web Services | Initial $38 billion cloud agreement, subsequently expanded | Multiyear access to AWS computing infrastructure and AI accelerators |
| Oracle | Reported arrangement valued at approximately $300 billion | Long-term cloud and infrastructure capacity agreement |
| CoreWeave | Approximately $22.4 billion across agreements announced in 2025 | Contracted capacity from a specialised AI cloud provider |
The Stargate project, announced in January 2025, initially proposed investing $500 billion in American AI infrastructure over four years, beginning with $100 billion. OpenAI, SoftBank, Oracle and other partners were involved in the initiative. Subsequent project announcements expanded its planned capacity and brought the stated investment pipeline above $400 billion, but the programme was never equivalent to OpenAI personally financing every facility.
Nvidia’s September 2025 partnership announcement contemplated deploying at least ten gigawatts of Nvidia systems, supported by an intended investment of up to $100 billion in OpenAI. That investment was designed to occur progressively as infrastructure came online, rather than through an unconditional transfer of $100 billion immediately after the announcement.
AMD’s agreement covered six gigawatts of accelerators, beginning with the planned deployment of its MI450 generation. The arrangement included a highly unusual incentive allowing OpenAI to acquire AMD shares if specified commercial and financial milestones were achieved. That equity incentive was not simply another cash obligation that could be added to OpenAI’s infrastructure bill.
The Broadcom collaboration targeted the development and deployment of ten gigawatts of custom accelerators between 2026 and 2029. Although estimates placed its potential value at approximately $350 billion, the companies did not announce that figure as an official contract price. Treating it as a confirmed payment commitment substantially overstates what was publicly established.
Other arrangements more directly concern computing services. Under the restructured Microsoft–OpenAI partnership, OpenAI committed to purchasing an additional $250 billion in Azure services. Its Amazon partnership initially involved $38 billion in AWS computing capacity before the relationship expanded, while CoreWeave’s disclosed agreements with OpenAI reached approximately $22.4 billion during 2025.
Even these figures cannot always be added without qualification. The same underlying infrastructure may involve a developer constructing a facility, a cloud provider operating it, a chipmaker supplying its accelerators and OpenAI purchasing the resulting computing service. Counting every transaction as separate infrastructure investment can exaggerate the total economic activity represented by a project.
The $1.4 trillion figure is therefore most useful as an indication of the enormous computing ecosystem OpenAI wants to mobilise. It is not a conventional debt balance, and not every announced dollar represents a binding obligation. Yet the underlying agreements still require somebody to provide capital, construct facilities and eventually earn a return.
Why OpenAI Is Racing to Secure Compute
OpenAI’s demand for computing power comes from two related activities: training AI models and operating them after they have been developed. Both require substantial infrastructure, but they create different financial pressures.
Training a frontier model involves running enormous clusters of advanced processors for extended periods, using large datasets to adjust the model’s internal parameters. These projects are expensive, technically demanding and increasingly dependent on specialised networking, memory systems and electricity supplies.
Inference is the process of using a trained model to produce answers and perform tasks. Every ChatGPT conversation, generated image, coding session or AI-powered research task requires computing resources. More sophisticated reasoning models may perform substantial internal computation before responding, while autonomous agents can execute multiple steps, call tools and continue working over longer periods.
Unlike a traditional software product that can distribute another copy at relatively low marginal cost, an AI service must continually perform computation for its customers. As adoption increases, inference can become an enormous recurring expense even if each individual task becomes cheaper.
This explains why OpenAI is attempting to secure infrastructure years before all the expected demand has materialised. Data centres take time to build, electricity connections can be difficult to obtain, and the most advanced chips depend on complicated global supply chains. Waiting until capacity is urgently needed could leave the company unable to serve customers or compete with better-resourced rivals.
Supplier diversification also matters. OpenAI’s early dependence on Microsoft Azure and Nvidia hardware gave those partners considerable importance. Agreements with AMD, Amazon, CoreWeave and Broadcom provide additional capacity and potentially strengthen OpenAI’s negotiating position.
Custom processors are especially attractive because general-purpose GPUs, despite their flexibility, are expensive. Chips designed around particular inference workloads might deliver more useful computation for each dollar spent and each unit of electricity consumed.
The Broadcom relationship illustrates this strategy. In June 2026, the companies unveiled Jalapeño, an inference processor designed for OpenAI workloads, moving the collaboration beyond its initial capacity announcement. OpenAI is increasingly interested in the architecture beneath its models, including processors, networking, memory, scheduling systems and data-centre design.
Success could reduce dependence on external suppliers and improve profit margins. Failure could leave OpenAI and its partners with enormous amounts of expensive, specialised hardware whose economic value depends on future demand.
Who Actually Pays for the Data Centres?
A major source of confusion is the assumption that OpenAI must personally fund every facility associated with its infrastructure plans. In practice, large computing projects can involve several businesses with different financial responsibilities.
Consider a hypothetical AI data centre. An infrastructure developer acquires land and constructs the building, financing part of the project with loans. A cloud provider signs a long-term lease, purchases processors and networking equipment, and agrees to obtain electricity from a utility. OpenAI then commits to purchasing computing capacity from the cloud provider over several years.
OpenAI may never own the building, the electrical equipment or the GPUs. Its obligation might instead consist of paying for access to a specified amount of computing power, subject to the terms of its agreement.
The cloud provider can use that anticipated revenue to justify the investment and attract lenders. Those lenders, in turn, may evaluate the creditworthiness of OpenAI and the cloud company when deciding whether the project’s expected income is sufficiently reliable.
Much of the capital can therefore come from businesses other than OpenAI. Microsoft, Amazon and Nvidia have substantial operating cash flows and access to financial markets. Oracle, CoreWeave and specialised developers can use debt, leases and infrastructure financing, while private-credit funds and institutional investors can fund individual projects.
Utilities and energy companies may separately invest in generation, transmission and grid connections. Some of those facilities could serve other customers, but their construction may still be encouraged by the expectations surrounding AI.
This arrangement makes an extraordinarily large infrastructure programme possible without requiring OpenAI to have trillions of dollars in cash. It also distributes the risk across a much wider financial network.
The crucial distinction is between who finances the construction and who ultimately pays for the capacity. If OpenAI enters a long-term contract with minimum purchase requirements, its supplier may be entitled to payment even when demand disappoints. If the supplier borrows against that contract, the lender expects repayment regardless of whether the projected AI boom materialises.
Contractual details make an enormous difference. A non-binding capacity target creates less exposure than a take-or-pay agreement requiring payment for reserved capacity. A cloud contract with flexible cancellation provisions is different from a long-duration lease secured by costly equipment. An investor buying equity accepts the possibility of losing its investment, whereas a creditor generally expects scheduled payments.
Many of the most consequential terms are private. Outside investors may not know the full cancellation provisions, minimum utilisation requirements, guarantees, penalties, equipment financing terms or refinancing conditions.
This lack of visibility creates uncertainty about where financial losses would ultimately accumulate. A project can appear attractive when viewed from the standpoint of a cloud provider’s contracted revenue while carrying substantial risks for its lenders, landlords or equipment financiers.
OpenAI’s $1.4 trillion ambition is therefore better understood as a network of capital commitments than as one corporate expenditure account. That structure makes the expansion financially feasible, but it also means the consequences of disappointment may spread far beyond OpenAI itself.
The Circular Money Machine: Investors, Customers and Suppliers
The most controversial feature of the AI infrastructure boom is that many participants are simultaneously investors, customers and suppliers to one another.
Ordinarily, a company raises money from investors and uses it to buy the products and services necessary to run its business. Its suppliers receive revenue because the company has demand for what they provide. If the company grows successfully, its investors benefit from the value it creates.
AI infrastructure financing complicates this arrangement because suppliers may provide capital to the businesses purchasing their products. A company can help finance its customer, receive revenue when that customer spends the money, and then use its growing sales to justify further investment.
Microsoft and OpenAI provide the clearest starting point.
Microsoft invested more than $13 billion in OpenAI through arrangements that included a financial stake and commercial rights. These were not all equivalent to unrestricted cash transfers; the broader partnership also involved access to Azure computing resources and a complex commercial relationship.
OpenAI became a major purchaser of Azure capacity. Microsoft, needing to provide that capacity, invested heavily in data centres and bought substantial quantities of Nvidia hardware.
Nvidia benefited from growing demand for its accelerators. It then announced its intention to invest up to $100 billion in OpenAI as OpenAI deployed additional Nvidia computing systems.
The relationship creates a reinforcing financial cycle. Microsoft provides capital and computing services to OpenAI. OpenAI pays Microsoft for capacity. Microsoft purchases Nvidia equipment to deliver that capacity. Nvidia benefits from the sales and proposes investing in OpenAI, which could then acquire more infrastructure containing Nvidia processors.
Importantly, this does not mean that every dollar moves through the entire circle unchanged, or that each transaction is immediately completed. The companies have separate finances, contractual obligations and commercial activities. Nevertheless, the arrangements connect the financing of AI demand with the revenues earned by the companies supplying it.
Amazon introduces another version of this relationship. Its strategic investment in OpenAI is connected with an expanding AWS relationship that includes access to Nvidia GPUs and Amazon’s Trainium processors. A business can therefore benefit both from the possible appreciation of its investment in OpenAI and from selling the infrastructure that OpenAI uses.
A similar pattern can exist in specialised cloud computing. A provider obtains contracts from an AI company, raises debt or equity to purchase hardware, and uses that hardware to fulfil the contracts. The supplier’s rapidly growing revenue can then support more financing and expansion.
The relationship becomes especially complicated when the projected revenue from a small number of customers supports enormous infrastructure valuations.
Imagine a chipmaker investing $10 billion in an AI developer. The developer uses that capital to purchase cloud services, while the cloud provider uses the revenue to finance equipment purchases from the same chipmaker. The chipmaker receives legitimate payment for equipment, and the developer receives legitimate computing services.
The transactions can be genuine without proving that independent customers ultimately value the AI service at a level sufficient to sustain the entire cycle.
That is the distinction between commercially real revenue and economically self-sustaining demand.
Revenue normally has to be recognised under applicable accounting rules, and strategic relationships require their own disclosures and treatment. The existence of an interconnected investment is not evidence that recorded revenue is fabricated. Chips may have been delivered, computing services may have been supplied, and customers may have made real payments.
The concern is what happens when the outside financing stops. If an AI developer cannot generate enough money from subscriptions, enterprises or API customers to support its operating costs, it may need another funding round. Its suppliers may also depend on the developer’s future purchases to justify their own expansion.
A financial ecosystem can therefore grow rapidly while remaining dependent on continuous capital inflows. Supplier revenue is not necessarily false, but its persistence may rely on the willingness of investors and creditors to keep funding the customer.
The smaller startup ecosystem deserves similar scrutiny, although its arrangements should not be confused with direct supplier financing. OpenAI’s associated startup-investment activities include the separately capitalised OpenAI Startup Fund, which invests in companies developing applications and other businesses around artificial intelligence. Some portfolio companies may purchase OpenAI services, but that does not establish that their financing is automatically recycled into OpenAI revenue or that they are required to spend their investments with the company.
The same caution applies to other AI startups receiving capital from technology companies. A supplier investing in an early customer can be a rational strategy for supporting a promising market. The financial concern arises when much of the industry’s apparent demand depends on investments from the companies that expect to benefit from that demand.
In a healthy version of this cycle, outside customers increasingly pay for valuable AI products. Their spending supports the developer, which pays cloud providers, which purchase hardware and expand infrastructure. The ecosystem eventually becomes self-sustaining because its ultimate revenues come from independent economic activity.
In a weaker version, investments and debt keep flowing into AI companies so those companies can buy more computing power, while their suppliers interpret the resulting sales as evidence that the market can support continuous expansion.
The difference may be difficult to recognise during a boom. It becomes impossible to ignore when investors begin asking which businesses can survive without another round of financing.
AMD’s Penny Warrants and the Price of Buying Demand
The OpenAI–AMD partnership illustrates how aggressively suppliers are prepared to compete for customers in artificial intelligence.
AMD agreed to support a planned deployment of six gigawatts of its accelerators and issued OpenAI a warrant allowing it to acquire up to 160 million AMD shares for an exercise price of one cent per share. If fully exercised, the shares could represent approximately 10% of AMD.
At a sufficiently high share price, the potential value of those shares would reach tens of billions of dollars. This created the impression that AMD was offering OpenAI an enormous financial reward for becoming a customer.
However, the terms described in AMD’s regulatory filing are substantially more complicated than an immediate stock giveaway.
The warrant is tied to deployment and purchase milestones, beginning with the initial gigawatt and extending toward the full six-gigawatt agreement. Additional technical, commercial and share-price requirements affect whether and when shares can vest and become exercisable.
OpenAI does not simply receive 160 million unrestricted shares because it signed the agreement. It must satisfy contractual conditions, while AMD must meet specified performance thresholds. The eventual financial value also depends on AMD’s share price and the extent to which the warrant becomes exercisable.
Those distinctions matter because the theoretical market value of the maximum share allocation is not equivalent to AMD spending that amount in cash. The primary financial effect for AMD’s existing shareholders would be potential dilution if additional shares were issued.
Why would AMD offer such an incentive?
Nvidia has dominated the market for advanced AI accelerators, supported by its hardware capabilities, software ecosystem and relationships with leading cloud providers. For AMD, securing substantial OpenAI deployments could demonstrate that its products can support demanding frontier-model workloads.
That validation could encourage cloud providers, developers and other customers to adopt AMD hardware. If the agreement helps AMD capture a significant share of an expanding market, its shareholders might ultimately benefit despite the dilution associated with the warrant.
OpenAI has incentives of its own. Using AMD hardware could diversify its infrastructure, reduce dependence on Nvidia and create competition among suppliers. The warrant also allows OpenAI to participate financially if AMD’s market position strengthens as a consequence of the partnership.
The arrangement therefore aligns the companies’ commercial interests while introducing an unusual financial incentive into a purchasing decision.
OpenAI is not necessarily choosing between AMD and Nvidia solely on hardware performance, price and availability. Its potential ownership stake gives it another reason to help AMD succeed.
There is nothing inherently illegitimate about such an arrangement. But an investor analysing AMD’s future sales must recognise that some of the expected demand comes from a strategically incentivised customer rather than an entirely conventional purchasing relationship.
The relevant question is whether the hardware will generate enough economic value to justify its deployment, regardless of the financial rewards associated with signing the agreement.
Can Real AI Revenue Catch Up With the Spending?
OpenAI’s extraordinary growth is the strongest argument against dismissing the entire infrastructure programme as speculative.
The company reportedly generated approximately $13 billion in revenue during 2025. By early 2026, its annualised revenue run rate had moved above $25 billion. According to Reuters reporting from 19 October 2026, that annualised figure had reached approximately $50 billion.
These numbers indicate remarkable growth, but they measure different things. Revenue earned during a completed financial year is not the same as annualised revenue calculated from a recent period of activity.
For example, a company earning $4 billion in one month can describe that pace as a $48 billion annualised run rate. It does not follow that the company actually collected $48 billion during the preceding twelve months.
Run-rate figures can be useful when a business is expanding rapidly because they provide a snapshot of its current scale. They become misleading when compared directly with cumulative spending commitments without acknowledging the difference in measurement periods.
Revenue is also not profit, and neither is equivalent to free cash flow.
A company might collect substantial customer payments while incurring enormous expenses to serve those customers. Its gross profit reflects the revenue remaining after the direct costs of providing its services, while operating profit also accounts for expenses such as research, administration, staff and sales.
Free cash flow measures the cash remaining after operating needs and capital expenditure. It is particularly important when assessing whether a company can support long-term infrastructure commitments without continually obtaining additional financing.
OpenAI’s economics are complicated by the recurring cost of inference. A traditional software company can distribute another licence at comparatively low incremental cost after developing the product. An AI service requires continuing computation for every conversation, coding session, generated video or autonomous task.
More capable models may also require greater computation per request. Consequently, customer adoption can increase revenue while simultaneously increasing infrastructure requirements.
Technology may improve this relationship. Better processors, more efficient models, improved scheduling and lower-precision computation can reduce the cost of delivering a task.
But lower costs can stimulate additional usage. If a service becomes cheaper, customers may use it more frequently, request more complicated work or replace simple interactions with longer-running agents.
The economic result depends on whether revenue and customer value grow faster than the total cost of delivering those services.
OpenAI’s spending plans have also evolved. Earlier reporting discussed approximately $600 billion in computing expenditure through 2030, while October 2026 reporting described projections closer to $750 billion. These estimates concern a more specific spending horizon than the broader $1.4 trillion infrastructure ambition and should not be added to it as separate liabilities.
The industry-wide economics are equally challenging. In its 2025 Global Technology Report, Bain & Company estimated that supporting projected AI demand could require approximately $500 billion in annual spending on new data centres by 2030. Its analysis described a potential $800 billion annual revenue gap after accounting for expected efficiency and productivity improvements.
The often-repeated figure of $2 trillion in required new revenue concerns the broader AI ecosystem. It is not a forecast that OpenAI itself must earn $2 trillion, nor does it prove that the industry’s projected infrastructure cannot generate acceptable returns.
The more useful interpretation is that the industry may need an enormous expansion of commercially valuable AI applications to support the infrastructure it wants to build.
Some of that demand could come from consumer subscriptions, but enterprise spending may become more important. Businesses must move beyond experimenting with AI and begin paying for tools that reduce costs, improve productivity, create new products or replace expensive workflows.
The danger is that infrastructure construction proceeds on the assumption that these applications will appear quickly enough to justify the investment.
OpenAI can continue growing rapidly and still face serious financial challenges if the cost of serving that growth, acquiring customers and expanding infrastructure rises at a similarly extraordinary pace.
Where the Financial Risk Is Concentrated
Not every company participating in the AI infrastructure boom faces the same risk.
Microsoft, Amazon and Nvidia have enormous established businesses, substantial operating cash flows and considerable access to financing. Although a reversal in AI demand could damage their profitability and market valuations, they possess other businesses that can help absorb losses.
Specialised infrastructure providers may be much more exposed.
A company building data centres for a small number of AI customers can report rapidly growing contracted revenue while concentrating its future income in those customers. If one major customer delays expansion or renegotiates a contract, the financial consequences may be severe.
CoreWeave illustrates this vulnerability. Its agreements with OpenAI reached approximately $22.4 billion in 2025, creating substantial expected future demand for its cloud infrastructure. Yet providing that capacity requires expensive equipment, facilities and financing.
An October 2026 Reuters examination of CoreWeave’s financing described a network carrying approximately $50 billion in debt. The scale underscores how quickly infrastructure exposure can accumulate when borrowing is used to finance expansion against anticipated customer payments.
Oracle faces a related challenge, although its business is more diversified. Its increasing infrastructure expenditure and debt exposure have attracted scrutiny because its existing cloud cash flows have historically been much smaller than those of Microsoft and Amazon.
Customer concentration is therefore only one concern. Financing costs, debt maturity schedules, lease obligations and the resale value of computing equipment can all determine whether an infrastructure provider can survive a slowdown.
Data centres are also unusually difficult to build quickly. Projects require land, construction, power connections, cooling infrastructure and technical equipment. Developers often make substantial financial commitments years before the facilities begin earning revenue.
By the time unexpectedly weak demand becomes visible, a large portion of the investment may already be irreversible.
A further risk is the useful life of the equipment. AI accelerators can become less competitive as new hardware generations arrive. If projected demand disappoints, collateral securing a loan may be worth less than lenders expected.
This creates refinancing risk. A borrower might successfully service its debt during the early years of a project but struggle to obtain replacement financing if customer contracts weaken, equipment depreciates rapidly or investors demand higher interest rates.
Private credit, equipment financing, project loans and long-term leases can distribute these exposures across institutions that are less visible than publicly traded technology companies.
The system does not require fraudulent accounting or worthless technology to become unstable. It only requires too many participants to make long-term financial commitments based on similarly optimistic assumptions about future demand.
What Happens if OpenAI’s Growth Slows?
A substantial slowdown in OpenAI’s growth would not automatically cause the entire AI industry to collapse. The consequences would depend on how quickly demand weakened, which contracts remained enforceable and how much flexibility individual companies retained.
The first response could be to postpone infrastructure deployments or renegotiate purchasing commitments. OpenAI might delay capacity additions, shift workloads between providers or seek more favourable terms from suppliers.
Some suppliers might accept renegotiation rather than lose a strategically important customer. Others could insist on contractual payments, depending on their agreements and financial circumstances.
Cloud providers would then reconsider their construction plans. Microsoft and Amazon could potentially redirect some computing capacity toward other customers because their cloud businesses serve many different workloads.
Specialised providers would have fewer alternatives, particularly if their facilities were designed primarily around demanding AI computing systems and a limited number of tenants.
Standard Nvidia and AMD accelerators might find alternative customers, although a simultaneous industry-wide slowdown would reduce demand for such equipment. Custom chips and highly specialised infrastructure could prove more difficult to redeploy.
The next effects would move through suppliers. Orders for processors, networking equipment, memory systems and cooling technology could be reduced or postponed. Manufacturers that expanded capacity based on optimistic forecasts might face excess inventory or lower selling prices.
Infrastructure owners could record impairments if their buildings, machinery or equipment were worth less than expected. Those accounting losses would not necessarily mean immediate insolvency, but they could weaken balance sheets and make further financing more expensive.
Lenders would face greater pressure when cash flows from contracted capacity became uncertain. Highly leveraged providers might be forced to refinance on unfavourable terms, sell assets or restructure obligations.
The distinction between equity and debt becomes particularly important during such a downturn. Equity investors can accept declining valuations or lose their investments, while borrowers must still satisfy interest and principal payments unless creditors agree to changes.
Public technology giants could experience substantial share-price declines, lower margins and expensive write-downs without necessarily approaching insolvency. Smaller cloud providers, developers and financial vehicles with concentrated customer exposure and large debt obligations could face much greater danger.
A severe AI downturn would therefore be more likely to unfold through cancelled construction, underused infrastructure, equipment write-downs and credit stress than through the simultaneous disappearance of every major AI company.
The extent of the damage would depend less on the amount of money announced during the boom than on where legally enforceable obligations and borrowed money ultimately reside.
Is This Another 2008—or a Repeat of the Dot-Com Boom?
The comparison between the AI infrastructure boom and the 2008 financial crisis is tempting because both involve complex financial relationships, rising valuations and the possibility of risks spreading between institutions.
But the underlying mechanisms are substantially different.
The 2008 financial crisis developed through a financial system heavily exposed to risky mortgage lending, securitisation, derivatives, excessive leverage and unstable short-term funding.
Banks and other institutions packaged mortgages into securities, distributed them throughout global markets and used additional financial instruments to transfer or amplify exposure. When housing prices declined and mortgage defaults increased, investors struggled to determine where the losses were concentrated.
The resulting collapse of confidence damaged financial institutions that depended on one another for funding and liquidity. Credit markets froze, major firms failed or required government intervention, and the crisis spread into the wider economy.
AI infrastructure contracts have not been distributed through the financial system in a comparable way. There is no known equivalent of millions of vulnerable household mortgages supporting layers of widely held credit derivatives and short-term banking liabilities.
Moreover, several of the central technology companies have strong balance sheets and profitable businesses outside AI infrastructure.
That makes an exact replay of 2008 unlikely, although it does not eliminate broader financial contagion. AI infrastructure increasingly relies on debt, leases and private-credit arrangements, and losses could spread across creditors, investors and suppliers if assumptions about demand prove incorrect.
A more illuminating comparison comes from the telecommunications and dot-com investment boom of the late 1990s.
During that period, investors correctly recognised that the internet would transform communication, commerce and business. Demand for connectivity was expanding, and telecommunications networks were becoming essential infrastructure.
Companies responded by building enormous amounts of fibre-optic capacity. Equipment manufacturers supplied the infrastructure, telecommunications operators borrowed heavily, and some suppliers helped finance customers so those customers could continue purchasing equipment.
The expansion was based on expectations that internet traffic and commercial demand would grow fast enough to justify the investment. Rising equipment sales and market valuations reinforced those expectations, encouraging even greater construction.
But the investment cycle moved faster than the economic demand needed to support it.
When growth disappointed, telecommunications companies cut capital expenditure. Equipment orders collapsed, unused fibre capacity accumulated, businesses failed and investors suffered enormous losses.
A Federal Reserve Bank of Richmond analysis highlighted the parallels between the telecommunications investment cycle and modern AI infrastructure, including the possibility that technology-related capital spending can expand faster than the market ultimately supports.
The broader dot-com bubble followed a related pattern. Technology valuations surged during the late 1990s, the Nasdaq peaked in March 2000, and the subsequent decline destroyed trillions of dollars in market value. Yet the internet did not disappear, and surviving companies such as Amazon continued to develop enormous businesses.
The important lesson is that technological progress and financial returns are separate questions.
The investors financing infrastructure before demand materialises may lose money even when the technology eventually becomes indispensable. A network built too early, too expensively or with too much debt can produce disappointing returns despite being useful decades later.
AI could follow a similar path. People may routinely use sophisticated artificial intelligence throughout their working lives, while some of the data centres, processors and financing arrangements created during the initial expansion turn out to have been poor investments.
The most credible danger is therefore not that artificial intelligence will suddenly stop working. It is that the technology succeeds on a slower timetable than the financing assumptions used to construct its infrastructure.
Could OpenAI Become Too Big to Fail?
The phrase too big to fail normally describes an institution whose disorderly collapse could cause such widespread economic damage that governments feel compelled to intervene.
OpenAI is not formally in that category simply because it has an enormous valuation or influential technology. Yet its growing importance to American infrastructure and industrial strategy creates the possibility that a severe financial crisis involving the company would become a political concern.
OpenAI’s commercial relationships influence demand for advanced chips, electricity, networking equipment, cloud services, construction and specialised manufacturing. Large projects increasingly depend on assumptions about the company’s future computing requirements.
Artificial intelligence also occupies an important position in American national-security and industrial-policy discussions. Policymakers view advanced computing infrastructure, domestic chip production and energy availability as matters of strategic importance.
That creates incentives for governments to support the infrastructure surrounding AI development through permits, public investment, tax incentives, manufacturing programmes or financing assistance.
Supporting an industry, however, is not the same as guaranteeing a particular company’s financial obligations.
The distinction became controversial in November 2025 when OpenAI CFO Sarah Friar discussed the possibility of governmental support that could help finance infrastructure development. Her remarks were widely interpreted as suggesting that a federal backstop might be sought.
OpenAI subsequently clarified that it was not asking the government to guarantee its data-centre commitments. Sam Altman also rejected the suggestion that taxpayers should protect OpenAI from private financial losses, while supporting government action that could strengthen the broader energy and computing infrastructure ecosystem.
Those positions do not amount to an existing federal guarantee.
Nevertheless, a severe downturn could create difficult political choices. If the failure of a major AI company threatened strategically important projects, employment, lenders or critical computing capacity, policymakers might consider indirect intervention.
Possible responses could include supporting certain infrastructure projects, facilitating restructuring, helping viable assets find new owners or protecting capabilities judged important for national security.
Such interventions would not necessarily rescue OpenAI’s shareholders or guarantee all its contracts. Governments could prioritise the preservation of infrastructure while allowing investors and lenders to absorb losses.
OpenAI may never receive the kind of direct rescue associated with the financial crisis of 2008. Its strategic importance does, however, mean that a disorderly failure might be difficult to contain within ordinary corporate bankruptcy proceedings.
The Case That the Infrastructure Bet Could Work
The sceptical interpretation of OpenAI’s infrastructure strategy is incomplete without considering what might make the enormous investment economically rational.
OpenAI has already demonstrated genuine commercial demand. ChatGPT became a widely used consumer product, while APIs, coding tools, enterprise services, research applications and AI agents have expanded the ways businesses and individuals can pay for advanced models.
The growth in reported revenue suggests that AI is not simply an imaginary industry supported by speculative expectations. Customers are purchasing services, businesses are building applications around models, and useful products are emerging.
OpenAI has also displayed a remarkable ability to attract capital. In March 2026, the company announced $122 billion in committed funding at a reported post-money valuation of approximately $852 billion.
Access to capital on that scale gives the company time to expand, experiment with business models and develop infrastructure before it reaches sustainable profitability. It does not guarantee financial success, but it weakens the argument that OpenAI must immediately finance its entire programme using operating cash flow.
The financing environment continued to evolve during 2026. Reports described preparations for a possible public listing, while Bank of America reportedly extended a $520 million credit line in July, an early step toward more conventional bank borrowing.
These developments broaden the company’s financing options beyond strategic equity investors. They also create new expectations about financial transparency, repayment and the ability to demonstrate sustainable business economics.
The infrastructure arrangements themselves are staggered across several years. Conditional investments can depend on deployment milestones, and some planned capacity can be adjusted if circumstances change.
A failure to achieve every announced ambition would not necessarily mean the company had defaulted on a corresponding amount of debt. The flexibility available would depend on the terms of its actual contracts.
Technological progress also offers a plausible route toward better economics. More efficient models, improved inference techniques, cheaper computation, specialised accelerators and better management of hardware could reduce the cost of serving customers.
OpenAI’s relationships with Nvidia, AMD, Amazon and Broadcom could encourage supplier competition and reduce dependence on a single architecture.
The development of its Jalapeño inference processor is an example of infrastructure strategy becoming tangible rather than remaining an announcement. Hardware specifically designed for OpenAI’s workloads could improve performance and economics if it proves competitive at scale.
Some facilities may also remain useful even if OpenAI grows more slowly than expected. General-purpose cloud capacity and widely supported accelerators can serve other customers, particularly where operators have diversified customer bases.
Perhaps the strongest argument is that today’s AI market may represent only a small portion of the eventual economic opportunity.
Coding, scientific research, customer service, advertising, education, professional services, creative production and enterprise automation could support markets much larger than current chatbot subscriptions.
If AI systems increasingly complete valuable work rather than simply generate answers, businesses may become willing to pay based on the economic benefits they receive. A company that saves substantial labour costs or accelerates product development might justify much greater AI expenditure than its existing software budget.
The bull case is therefore not that OpenAI’s infrastructure commitments are small or financially conservative. It is that the market for useful intelligence may eventually become large enough to justify infrastructure that currently appears excessive.
Whether that happens depends on the timing of adoption, improvements in cost efficiency and the ability to turn technical capability into profitable customer demand.
The Numbers That Will Decide Whether This Is a Bubble
The number of infrastructure agreements OpenAI signs is not the best measure of whether its strategy is succeeding. A more useful assessment focuses on the relationship between the economic value customers receive and the financial resources needed to deliver it.
Recognised revenue is the first measure. OpenAI’s rapid growth in annualised run-rate figures is important, but investors ultimately need sustained revenue recorded over complete reporting periods. The distinction matters especially as the company considers larger fundraising rounds and potential public-market scrutiny.
Free cash flow is equally important because extraordinary revenue growth can coexist with immense cash consumption. The company eventually needs to generate enough cash from its operations to support its business without depending indefinitely on further funding.
Gross margin reveals whether the cost of delivering AI services is improving relative to the revenue those services produce. Falling inference costs should create a greater contribution from each subscription, API call or enterprise contract, provided pricing and usage patterns do not offset the savings.
Infrastructure utilisation measures whether the enormous amount of installed computing capacity is actually being used productively. A data centre can be technically impressive and financially disappointing if demand does not justify its construction and operating costs.
Enterprise adoption is another important indicator. Consumer subscriptions created a substantial initial market, but businesses offer opportunities for larger and potentially more durable revenue streams. The relevant question is whether customers move from experimental deployments to permanent applications that produce measurable economic returns.
Customer concentration deserves particular attention among infrastructure suppliers. Investors should examine how much future growth at companies such as Oracle, CoreWeave and AMD depends on purchases from OpenAI or a similarly small group of AI developers.
Contract revisions can provide an early warning of changing expectations. Delayed deployments, reduced orders, altered minimum commitments or major renegotiations may reveal that actual demand is developing more slowly than projected.
Debt and financing costs reveal whether the infrastructure ecosystem is becoming financially more fragile. Investors should examine borrowing throughout cloud providers, developers, equipment-financing vehicles and private-credit structures, not merely the balance sheets of the largest technology companies.
Capital raised compared with capital consumed shows whether OpenAI is progressing toward self-financing operations. A company can remain solvent for a long time when investors are willing to provide extraordinary amounts of funding, but continued fundraising is not itself evidence that the underlying business model will become profitable.
Finally, the cost of delivering useful intelligence may be the most consequential long-term measure. If AI systems become more capable while the cost of completing commercially valuable work falls, the industry could develop a powerful economic foundation for continued investment.
If the cost of supplying better AI rises faster than customers are willing to pay, expansion could make the financial problem larger instead of solving it.
These measures should be assessed together. High revenue growth will not compensate indefinitely for persistent negative cash flow, while strong customer adoption will not justify every data centre if capacity has been constructed at excessive cost.
The outcome will depend on whether the infrastructure eventually supports a sustainable business ecosystem rather than one principally sustained by investors financing one another’s growth.
OpenAI’s infrastructure ambitions represent a genuine technological and industrial transformation. The company has valuable products, growing revenue, powerful partners and access to extraordinary financial resources. Its investments and commercial agreements are not automatically fraudulent, nor do their announced headline values represent a single immediate financial obligation.
The concern is that much of the surrounding infrastructure is being constructed in anticipation of future demand that has not yet been demonstrated on the necessary scale. Suppliers invest in customers, customers purchase services from investors, and specialised infrastructure companies borrow against expected revenues from a relatively concentrated group of AI developers.
Those relationships can support an enormously productive industry if paying customers eventually generate sufficient demand. They can also amplify losses if growth slows and highly leveraged providers discover that their obligations are less flexible than their projections assumed.
Artificial intelligence does not have to fail for significant parts of its infrastructure boom to become financially irrational. OpenAI could continue growing into one of the world’s most important technology companies while some of the facilities and financing arrangements built around its ambitions still produce disappointing returns.
The central risk is that the technology may ultimately succeed, but not quickly enough to justify what investors are spending in anticipation of that success.
Last Updated on October 21, 2026 by Aseem Gupta