The Trillion-Dollar Question
OpenAI earned approximately $13 billion in revenue in 2025.
Around the same time, Sam Altman was describing plans for 30 gigawatts of artificial intelligence computing capacity requiring roughly $1.4 trillion of investment.
Put those numbers beside each other and the strategy appears absurd. How can a company generating tens of billions of dollars support infrastructure plans measured in trillions?
The apparent contradiction has fuelled claims that OpenAI sits at the centre of an artificial intelligence bubble—one sustained by technology companies investing in one another, financing their customers and recording revenue from deals that ultimately circle back through the same small group of firms.
There is a genuine problem here, but it is not quite as simple as OpenAI owing someone $1.4 trillion.
The headline figure combines different types of arrangements spread across several years: cloud-service contracts, data-centre projects, chip purchases, proposed investments, conditional warrants, capacity targets and estimated deal values. Some are binding. Some depend on deployment milestones. Some represent investments by OpenAI’s partners rather than direct spending from its own balance sheet.
That makes the situation less immediately impossible than the headline suggests.
It does not necessarily make it safe.
OpenAI is becoming the central customer, investment target and strategic partner in an increasingly interconnected AI economy. The companies providing its capital also sell it computing capacity. The companies selling it chips may invest in OpenAI. Cloud providers build data centres to serve demand from a company whose continued growth depends on raising more capital from the same ecosystem.
The central question is therefore not whether OpenAI must produce $1.4 trillion in cash tomorrow.
It is whether real, independently financed demand for artificial intelligence will grow quickly enough to justify the enormous amount of infrastructure now being built around OpenAI’s ambitions.
What OpenAI Has Actually Committed To
The first mistake in assessing OpenAI’s position is to add every announced figure together as though each represents an immediate and unconditional payment obligation.
The underlying agreements are not all equivalent.
| Partner or project | Publicly announced arrangement | What the figure actually represents |
| Stargate | Up to $500 billion and 10 gigawatts | A multiyear infrastructure initiative involving OpenAI, SoftBank, Oracle and other partners—not a single bill owed by OpenAI |
| Nvidia | At least 10 gigawatts, supported by an intended investment of up to $100 billion | A September 2025 letter of intent under which Nvidia would invest progressively as capacity was deployed |
| AMD | Six gigawatts of AMD GPUs | A multiyear purchase and deployment agreement supported by a contingent warrant for up to 160 million AMD shares |
| Broadcom | Ten gigawatts of custom AI accelerators | A development and deployment partnership; no official $350 billion contract value was disclosed |
| Microsoft | An incremental $250 billion of Azure services | A large, long-term cloud-computing contract following the restructuring of the Microsoft–OpenAI relationship |
| Amazon Web Services | An initial $38 billion agreement, later expanded | Multiyear access to AWS infrastructure, Nvidia GPUs and Amazon’s Trainium chips |
| CoreWeave | Approximately $22.4 billion across several agreements | Contracted cloud-computing capacity from a specialised GPU-cloud provider |
| Oracle | A reported arrangement valued at roughly $300 billion | Long-term cloud and data-centre capacity associated with OpenAI’s broader infrastructure programme |
The original Stargate announcement said the initiative intended to invest $500 billion in American AI infrastructure over four years, beginning with $100 billion. Later expansions brought the project close to seven gigawatts of planned capacity and more than $400 billion of stated investment.
But Stargate is a financing and development platform involving multiple companies. It is not the same thing as OpenAI signing a cheque for $500 billion.
The Nvidia partnership was similarly conditional. Nvidia said it intended to invest up to $100 billion progressively as OpenAI deployed at least ten gigawatts of Nvidia systems. The first gigawatt was expected in the second half of 2026.
The AMD agreement was described as a definitive six-gigawatt partnership, beginning with one gigawatt of AMD Instinct MI450 GPUs. Yet even here, the full scale would unfold over multiple product generations and depend on technical and commercial milestones.
Broadcom’s agreement involved ten gigawatts of custom accelerators between 2026 and 2029. The companies announced the capacity target but not a total contract value. The widely repeated $350 billion estimate was an outside calculation, not an official price.
Meanwhile, Microsoft confirmed that OpenAI had contracted to purchase an additional $250 billion in Azure services. OpenAI also signed an initial $38 billion agreement with AWS, while CoreWeave disclosed approximately $22.4 billion in total OpenAI contract value.
These are extraordinary numbers.
They are simply not interchangeable.
Some measure potential capital investment. Others measure services OpenAI has agreed to consume. Others depend on future deployment. Several projects overlap because one company may own the data centre, another may operate the cloud service and a third may provide the chips inside it.
The $1.4 trillion figure is best understood as the estimated cost of the computing empire OpenAI wants to assemble—not as a conventional debt balance sitting on its books today.
Why OpenAI Needs So Much Compute
OpenAI’s demand for infrastructure comes from two different activities: training models and running them.
Training is the process of building a model. It requires large clusters of advanced chips operating for weeks or months, processing enormous quantities of data and adjusting billions or trillions of internal parameters.
Inference begins after the model has been trained. Every time someone asks ChatGPT a question, generates an image, uses a coding assistant or deploys an AI agent, computing resources are needed to produce the answer.
Training attracts most of the public attention because a new frontier model may require an immense one-time computing run.
Inference may ultimately become the larger expense.
A successful consumer or enterprise product must serve millions of requests continuously. More capable models often perform additional internal reasoning before answering. AI agents may run for minutes or hours, use tools, call other models and revise their work repeatedly.
The cost is no longer confined to creating intelligence.
It includes delivering intelligence every time someone asks for it.
OpenAI is therefore trying to secure computing capacity before it needs it. From the company’s perspective, waiting for demand to arrive before constructing data centres would risk falling behind competitors that already possess the hardware, power and network infrastructure required to serve users.
Diversifying suppliers is also strategically important.
OpenAI began with a deep dependence on Microsoft Azure and Nvidia GPUs. Its newer arrangements with AMD, AWS, CoreWeave and Broadcom reduce the risk of relying on one cloud provider or one chipmaker. Competition between suppliers may also help OpenAI negotiate lower prices.
Custom chips could be especially important. General-purpose GPUs are powerful, but they are expensive and designed to handle many different workloads. A processor designed specifically around OpenAI’s inference patterns could potentially perform the same work with less energy and lower cost.
That is the logic behind the Broadcom partnership.
OpenAI is not merely buying more chips. It is trying to control more of the system beneath its models: processors, networking, memory, scheduling, data-centre design and software.
If successful, that integration could improve its margins.
If unsuccessful, it could leave OpenAI committed to an enormous amount of specialised infrastructure whose economic value depends on demand continuing to rise.
The Difference Between a Commitment and a Bill
A conventional company buys a factory, records the asset and gradually pays for it using cash, debt or operating profits.
AI infrastructure is rarely that straightforward.
A data centre may be funded by an infrastructure investor, built by a property developer, leased to a cloud provider, filled with chips financed through debt and ultimately used to serve an OpenAI workload under a multiyear computing contract.
OpenAI may never own the building.
It may instead agree to purchase a minimum amount of computing capacity over several years. The cloud provider then uses that contract to justify constructing the facility or raising money from lenders.
This distinction matters because OpenAI does not necessarily need $1.4 trillion in cash to initiate $1.4 trillion of infrastructure development.
Its partners can supply much of the capital.
Microsoft, Amazon and Nvidia can invest using their own cash flows. Oracle, CoreWeave and data-centre developers can borrow money. Infrastructure funds and private-credit firms can finance individual facilities. Power companies can build generation and transmission capacity under long-term agreements.
The structure spreads the initial cost across an ecosystem.
It also spreads the risk.
A long-term contract from OpenAI may allow a provider to secure financing, but someone must still pay the lender if OpenAI later reduces its usage, renegotiates the agreement or cannot meet its obligations.
The exact danger depends on the contract.
A non-binding capacity target creates less financial exposure than a take-or-pay agreement requiring payment whether the computing power is used or not. A cancellable cloud contract is different from a 15-year data-centre lease. An equity investment can fall in value, while a loan must normally be repaid.
Most OpenAI contracts are private, leaving outsiders unable to see minimum purchase requirements, cancellation rights, pricing adjustments, guarantees or penalties.
That opacity is part of the problem.
Investors can observe the enormous headline values without always knowing where the firm obligation sits—or who ultimately bears the loss if projected demand does not materialise.
Can OpenAI Grow Into the Spending?
OpenAI’s growth has been remarkable.
The company reportedly generated approximately $13 billion of revenue in 2025, exceeding its earlier expectations. By early 2026, its annualised revenue had risen above $25 billion.
That does not mean it collected $25 billion during the previous 12 months. An annualised run rate takes the revenue generated during a recent month and projects it across a full year.
If a company makes $2 billion in March, it can describe that as a $24 billion annualised run rate even if its revenue during the previous year was considerably lower.
Run rate is useful when a company is growing quickly.
It can also make the company appear larger than the amount of revenue it has actually earned.
There is an even greater difference between revenue and cash available for infrastructure.
Revenue is the amount customers pay. Gross profit is what remains after the direct cost of serving those customers. Operating profit also subtracts research, salaries, sales, administration and other expenses. Free cash flow accounts for the cash required to operate and invest in the business.
OpenAI can therefore grow revenue rapidly while continuing to consume cash.
According to figures reported in early 2026, OpenAI generated $13 billion in 2025 while spending approximately $8 billion. It was targeting about $600 billion in total computing expenditure through 2030, a figure that appears to represent a more defined near-term spending plan within the broader $1.4 trillion infrastructure ambition.
The economics become harder because AI does not naturally resemble traditional software.
Once Microsoft builds a copy of Word, distributing another licence costs relatively little. Every ChatGPT conversation, coding session or generated video requires additional computing resources. Greater usage creates greater costs.
Efficiency improvements can lower the cost per task, but lower prices may encourage people to use AI far more frequently. This is beneficial only when revenue expands faster than the cost of supplying the service.
Bain estimates that meeting projected global AI demand could require approximately $500 billion in annual spending on new data centres by 2030. Even after accounting for possible productivity savings, its analysis found an estimated $800 billion gap between the annual revenue needed to support the infrastructure and the revenue the industry might realistically generate. This was an industry-wide estimate, not a claim that OpenAI alone must earn $2 trillion.
OpenAI does not need to become a trillion-dollar-revenue company immediately.
It does need several things to happen simultaneously:
Revenue must keep growing at an exceptional rate. The cost of inference must fall. Enterprise adoption must expand beyond experimentation. Paying users must grow faster than free usage. New products must produce meaningful revenue. Capital markets must remain willing to finance losses.
That is possible.
It is also an extremely narrow path.
How the Circular AI Money Loop Works
Imagine that a chipmaker invests in OpenAI.
OpenAI uses part of the capital to purchase computing capacity containing that chipmaker’s hardware.
The cloud provider receiving OpenAI’s payment orders more chips to build the required capacity.
The chipmaker records increased demand, earns more revenue and becomes more valuable.
Its rising value gives it greater capacity to invest in OpenAI or in another AI company that will purchase more chips.
Nothing in this cycle is automatically illegitimate.
OpenAI receives real computing power. The cloud provider receives a paying customer. The chipmaker sells physical products. Investors gain exposure to a potentially transformative company.
But the loop complicates the interpretation of demand.
Microsoft invested more than $13 billion in OpenAI while serving as its primary cloud provider. OpenAI purchased Azure capacity. Microsoft purchased large quantities of Nvidia hardware to expand Azure. Nvidia then announced that it intended to invest up to $100 billion in OpenAI as OpenAI deployed Nvidia systems.
AMD offered OpenAI a warrant tied to future purchases of AMD hardware.
Amazon combined a major investment in OpenAI with agreements under which OpenAI would use AWS and Amazon’s Trainium chips.
The same money can therefore appear at different points in the network as investment capital, capital expenditure, supplier revenue and customer spending.
That does not mean the revenue is fake.
It means the quality of the demand deserves scrutiny.
The strongest form of demand comes from independent customers paying for AI because it produces measurable economic value. A weaker form comes from an AI company spending money supplied by the businesses whose products it is purchasing.
The distinction may not matter during rapid expansion. It becomes crucial when financing slows.
If customers eventually pay enough for AI products, the circle turns into a productive industrial ecosystem.
If the companies mainly depend on continued investment from one another, the circle begins to resemble vendor-financed demand.
The AMD Warrant: The Most Unusual Deal in the Network
The OpenAI–AMD agreement offers the clearest example of how far suppliers are willing to go to secure AI demand.
AMD issued OpenAI a warrant allowing it to purchase up to 160 million AMD shares for one cent each. At full exercise, the shares could represent close to 10% of the company.
Taken alone, that sounds like AMD gave OpenAI tens of billions of dollars in stock in exchange for becoming a customer.
The actual agreement is more complicated.
According to AMD’s regulatory filing, the shares vest in tranches. The first tranche is connected to delivery of the initial one gigawatt of AMD GPUs. Full vesting depends on OpenAI or authorised purchasers acquiring six gigawatts.
Additional technical, commercial and share-price conditions also apply. Later filings confirmed that the warrant depends on specified AMD stock-performance thresholds and other conditions before the shares become exercisable.
OpenAI does not receive 160 million unrestricted shares on the day the agreement is signed.
It earns the opportunity to acquire them as it generates business for AMD and as AMD achieves certain market-value targets.
For AMD, the arrangement is an aggressive customer-acquisition strategy.
Nvidia dominates the AI accelerator market. Convincing OpenAI to deploy AMD hardware could validate AMD’s technology, encourage cloud providers to support its chips and attract other customers. If the agreement helps AMD capture a large new market, dilution may be an acceptable price.
For OpenAI, the warrant creates a potential financial reward for diversifying away from Nvidia. If OpenAI helps increase AMD’s sales and share price, it participates in the value it helped create.
The arrangement aligns incentives.
It also blurs the normal relationship between customer and supplier.
OpenAI is not simply choosing the product offering the best performance and price. It could become one of AMD’s largest shareholders if it buys enough AMD products and helps the company succeed.
That does not invalidate the deal.
It does mean that the resulting demand cannot be interpreted in exactly the same way as an ordinary customer placing an ordinary purchase order.
When Reciprocal Deals Become Dangerous
Reciprocal relationships become risky when they begin to conceal how much demand exists outside the financial network supporting them.
Several warning signs matter.
The first is customer concentration. If a cloud provider builds an enormous amount of infrastructure for one customer, its future depends heavily on that customer’s ability to pay.
CoreWeave’s OpenAI agreements reached approximately $22.4 billion in 2025. Oracle’s infrastructure expansion has also become closely connected to OpenAI and other large AI customers. By June 2026, Oracle’s rising capital expenditure and debt load were already worrying investors because it lacked the enormous existing cloud cash flows of Microsoft and Amazon.
The second warning sign is supplier-supported demand.
When a supplier invests in a customer that purchases its products, reported sales may reflect both genuine usage and the supplier’s willingness to finance market expansion.
The third is infrastructure built before utilisation has been demonstrated.
Data centres require years of planning, grid connections, construction and equipment procurement. Companies must make decisions based on forecasts. By the time weak demand becomes obvious, much of the money may already have been spent.
The fourth is dependence on future fundraising.
OpenAI can continue signing contracts as long as investors believe its future value exceeds its present cash consumption. If that belief weakens, the company may need to delay capacity, renegotiate contracts or accept financing on less favourable terms.
The fifth is hidden leverage.
Microsoft, Amazon and Nvidia have strong balance sheets. Not every company in the network does. Data-centre developers, specialised cloud providers and infrastructure vehicles may rely on debt, leases or private credit that is less visible than public-company borrowing.
The system does not need fraudulent accounting to become unstable.
It only needs too many companies to make long-term commitments based on the same optimistic assumptions.
What Happens If OpenAI Misses Its Growth Targets?
An OpenAI slowdown would not necessarily produce an overnight collapse.
The first stage would probably involve renegotiation.
OpenAI could delay deployments, reduce minimum commitments or shift workloads between providers. Suppliers might accept changes rather than risk losing a strategic customer.
Cloud providers would then cut or postpone expansion.
Microsoft and Amazon could redirect some capacity to other customers. Nvidia and AMD could sell general-purpose accelerators elsewhere, although a simultaneous slowdown across the AI industry would make that more difficult.
Specialised infrastructure would be harder to repurpose.
A data centre designed around a particular cooling system, power density or accelerator architecture may not easily attract an alternative tenant. Custom OpenAI chips would have less obvious external value than standard Nvidia GPUs.
The next effect would reach suppliers.
Chip orders could be postponed. Networking, memory and cooling-equipment demand would weaken. Manufacturers that expanded production based on aggressive forecasts could face excess inventory or lower prices.
Infrastructure owners could record impairments if buildings or equipment proved less valuable than expected.
Highly leveraged providers would face the greatest pressure. Debt and lease payments continue even when customers use less capacity. Refinancing becomes difficult when lenders no longer trust future demand projections.
Public technology giants would probably survive.
They might suffer lower margins, falling share prices and billions of dollars in write-downs, but Microsoft, Amazon and Nvidia possess large profitable businesses beyond OpenAI.
Smaller cloud providers, developers and financial vehicles could fare much worse.
The most realistic failure scenario is therefore not every AI company disappearing at once.
It is a chain of cancelled expansion, underused data centres, supplier write-downs, credit losses and collapsing valuations—concentrated among the companies that borrowed most aggressively against OpenAI’s projected growth.
Is This Another 2008?
The comparison with the global financial crisis is emotionally powerful but structurally weak.
The 2008 crisis began inside the financial system and household economy. Risky mortgages were packaged into securities, distributed throughout global markets and amplified through leverage, derivatives and short-term funding.
Financial institutions depended on one another for liquidity. When confidence disappeared, banks could not determine which counterparties were solvent. Credit froze, firms failed and the shock rapidly reached households and businesses.
The Financial Crisis Inquiry Commission identified failures in mortgage lending, securitisation, risk management, credit ratings, derivatives and financial regulation as interconnected causes of the crisis.
OpenAI’s computing contracts have not been securitised throughout the global banking system in a comparable way.
There is no known equivalent of millions of vulnerable household mortgages sitting beneath layers of securities and credit derivatives. Consumers do not lose their homes merely because an AI data centre operates below capacity.
Many of the most exposed companies are also financially stronger than the institutions at the centre of the mortgage crisis. Microsoft, Amazon, Nvidia and major private-equity firms can absorb substantial losses without becoming insolvent.
That does not eliminate systemic risk.
AI infrastructure is increasingly funded through bonds, private credit, leases and project-finance structures. OpenAI’s importance to cloud providers and chipmakers also means that disappointment could produce a broad market correction.
But the transmission mechanism is different.
A severe AI downturn would initially look more like an industrial investment bust than a banking panic.
The Better Comparison: The Telecom and Dot-Com Infrastructure Boom
The telecommunications boom of the late 1990s offers a more useful warning.
The internet was real.
Data traffic was growing. Communication was moving online. Fibre-optic networks were clearly going to become essential infrastructure.
Investors were correct about the technological transformation.
They were wrong about how quickly demand would justify the amount of capacity being built.
Telecommunications companies laid vast fibre networks, purchased equipment and borrowed heavily. Suppliers financed customers so they could continue buying equipment. Rising sales reinforced optimistic forecasts, which encouraged more investment.
When growth failed to match expectations, capital expenditure collapsed.
Fibre sat unused. Companies failed. Equipment suppliers lost customers. Investors suffered enormous losses even though internet usage continued to expand.
A Federal Reserve Bank of Richmond study described how telecommunications equity values, investment and employment surged from 1997 before collapsing after 2000 as forecasts proved highly inaccurate.
AI shares several characteristics with that period.
The technology is real. Demand is increasing. Infrastructure is essential. Suppliers are supporting customers. Companies are building ahead of proven utilisation. Market forecasts cover outcomes ranging from useful workplace tools to a transformation of the entire global economy.
The AI boom could therefore succeed technologically while failing financially.
People may use artificial intelligence every day in 2035, just as they use the internet today, while many investors in the first infrastructure wave still lose money.
A technology can change the world without validating every price paid or every data centre built in its name.
The Bull Case for OpenAI
The sceptical case is powerful, but it is incomplete without acknowledging what could go right.
OpenAI has grown from almost no revenue to tens of billions of dollars in only a few years. ChatGPT has become a global consumer product, while coding, research, enterprise automation and AI agents are creating new ways to monetise the underlying models.
Infrastructure commitments are also staggered.
OpenAI does not need to fund every project simultaneously. Capacity can be added as usage expands, and conditional arrangements can be altered if technical or commercial milestones are not achieved.
The company has demonstrated an extraordinary ability to raise money.
In March 2026, OpenAI announced that it had closed a funding round with $122 billion in committed capital at an $852 billion post-money valuation. The scale suggests that many of the world’s largest investors and technology companies are willing to finance years of expansion.
Costs may also fall faster than sceptics expect.
New chip architectures, better model design, improved scheduling and lower-precision calculations can make each unit of intelligence cheaper to produce. OpenAI’s ability to shift workloads among Nvidia, AMD, Amazon and custom Broadcom hardware could create competition and reduce supplier dependence.
The Broadcom partnership has already moved beyond a distant announcement. In June 2026, the companies unveiled Jalapeño, OpenAI’s first custom inference processor. The chip was designed around the workloads OpenAI runs through ChatGPT, Codex and its API.
Some capacity may also be reusable.
General-purpose cloud infrastructure and Nvidia GPUs can serve other customers if OpenAI grows more slowly. The major cloud providers are not constructing every facility exclusively for one model company.
Most importantly, AI may create markets that are difficult to measure today.
Software development, scientific research, advertising, education, entertainment, customer service and professional work could generate enormous demand for models and agents. If AI meaningfully increases productivity, businesses may be willing to spend far more than current software budgets suggest.
The bull case is therefore not that the numbers are modest.
It is that the economic opportunity may eventually become large enough to justify numbers that currently appear unreasonable.
Is OpenAI Too Big to Fail?
“Too big to fail” has a specific meaning.
A company is not too big to fail merely because it is valuable, famous or strategically important. The phrase implies that its sudden collapse would create consequences severe enough for the government to intervene.
OpenAI is certainly becoming more important.
It sits at the centre of large American data-centre projects. Its contracts influence demand for chips, cloud services, electricity, networking equipment and construction. Its technology is used by businesses, consumers and government agencies. American policymakers increasingly view frontier AI as a national-security competition.
That creates political incentives to support the infrastructure around OpenAI.
Governments may accelerate permits, fund power generation, offer tax incentives, support domestic chip manufacturing or provide loan programmes for strategically important projects.
That is not the same as guaranteeing OpenAI’s obligations.
The controversy intensified when OpenAI CFO Sarah Friar discussed the possibility of governmental support that could make infrastructure financing easier. Her comments were widely interpreted as a request for a federal backstop.
OpenAI later clarified that it was not seeking government guarantees for its data-centre commitments. Sam Altman said the company did not want taxpayers protecting private losses, although he supported public investment in energy, manufacturing and infrastructure that could benefit the broader industry.
The distinction may become harder to maintain during a crisis.
If OpenAI’s failure threatened major data-centre projects, strategic chip supply chains or national AI capabilities, the government might intervene indirectly—even without writing a cheque to OpenAI.
It could support lenders, encourage a restructuring, help another company acquire assets or protect projects considered essential to national security.
OpenAI is not formally too big to fail.
It may be becoming too strategically entangled to disappear cleanly.
What Changed by Mid-2026
The OpenAI described in late 2025 is not the same company that existed by the middle of 2026.
Its financial position became stronger in the short term.
The $122 billion funding round gave OpenAI access to capital on a scale few private companies have ever achieved. Its annualised revenue exceeded $25 billion, and the company disclosed that it was generating approximately $2 billion per month.
It also moved towards the public markets.
In June 2026, OpenAI confidentially filed for an initial public offering. The company had not committed to a timetable, but a listing could provide access to an even larger pool of capital and make its financial performance more transparent to investors.
In July, Bank of America reportedly extended OpenAI a $520 million credit line—the company’s first conventional bank loan. The amount was small relative to OpenAI’s broader plans, but symbolically important. It suggested that OpenAI was beginning to build relationships with lenders rather than relying almost entirely on equity investors and strategic partners.
The operational story also advanced.
The Broadcom chip moved into working hardware. AWS expanded its relationship with OpenAI. AMD continued preparing its first major deployment. Stargate projects progressed through planning and construction.
These developments weaken the argument that OpenAI’s infrastructure strategy consists entirely of announcements that will never materialise.
They do not resolve the economics.
OpenAI continued to consume large amounts of cash. It reportedly spent $3.7 billion in the first quarter of 2026 against $5.7 billion of revenue, and investor materials indicated that the company did not expect profitability until around 2030. Reuters could not independently verify all of the figures drawn from private financial documents.
OpenAI also reportedly missed some internal revenue and user targets as competition intensified, particularly in coding and enterprise markets.
The company is therefore better financed than the original bubble argument implied.
It remains far from proving that its infrastructure can eventually be funded through the cash generated by its products.
The Numbers That Will Reveal Whether the Bet Is Working
The success of OpenAI’s strategy will not be determined by the number of partnerships it announces.
It will be determined by whether the underlying economics improve.
The first number to watch is recognised revenue—not annualised run rate. OpenAI must convert rapid monthly growth into sustained annual sales.
The second is free cash flow. Revenue growth means less when spending grows just as quickly. The company eventually needs to demonstrate that each additional dollar of usage contributes towards recovering its immense fixed and development costs.
The third is gross margin. Falling inference costs should gradually improve the amount OpenAI retains from each subscription, API call or enterprise contract.
The fourth is infrastructure utilisation. Gigawatts of installed capacity matter only when customers use them productively.
The fifth is enterprise adoption. Consumer subscriptions created OpenAI’s initial business, but corporations offer larger and potentially more stable contracts. The key question is whether companies move from experiments to permanent, high-value deployments.
The sixth is customer concentration among OpenAI’s suppliers. Investors should track how much of Oracle, CoreWeave, AMD or other providers’ future growth depends on OpenAI.
The seventh is contract revision. Delayed projects, reduced purchases or renegotiated commitments would reveal that demand is not growing as expected.
The eighth is the amount and cost of debt throughout the ecosystem. Equity investors can tolerate long periods of uncertainty. Lenders expect regular payment.
The ninth is capital raised compared with capital consumed. OpenAI’s fundraising ability is extraordinary, but it cannot remain the primary proof of the business model forever.
The final number is the price of useful intelligence.
If OpenAI can deliver increasingly valuable work while reducing the cost of each task, the infrastructure may generate enormous returns.
If producing better models requires spending faster than customers are willing to pay, scale will magnify the problem rather than solve it.
Verdict: A Real Technology Boom with Bubble-Like Financing Risks
OpenAI does not face a single $1.4 trillion invoice.
Its infrastructure plans are a web of multiyear contracts, staged deployments, supplier investments, cloud purchases, data-centre projects and conditional incentives. Much of the capital will be supplied by partners, lenders and infrastructure investors rather than directly by OpenAI.
That is the reassuring part.
The concerning part is that many of the companies funding the expansion also benefit from OpenAI spending the money with them.
The resulting transactions can be commercially real while still making it difficult to determine how much demand originates from independent customers and how much is being encouraged by strategic investment.
OpenAI has several advantages the most notorious dot-com companies did not possess. It has enormous usage, rapidly growing revenue, valuable products, powerful partners and access to extraordinary amounts of capital.
The companies surrounding it are not all fragile startups. Microsoft, Amazon, Nvidia and other major technology firms can absorb losses that would destroy smaller businesses.
Yet the system is still making long-lived infrastructure commitments based on assumptions about future AI demand that have not been proven at anything close to the required scale.
The most likely danger is not a replay of 2008.
It is an investment bust resembling the telecommunications boom: a genuine technological revolution accompanied by excessive construction, vendor-supported demand, overoptimistic forecasts and painful losses when usage fails to arrive quickly enough.
OpenAI may continue growing and become one of the most important companies in the world.
Artificial intelligence may transform the economy.
Both could be true while large parts of the infrastructure boom still turn out to be financially irrational.
The technology does not have to fail for the bubble to burst.
It only has to succeed more slowly than the money expects.
Last Updated on July 14, 2026 by Aseem Gupta