On October 14, 2026, research firm Gartner published an estimate that owners of AI data centers will need cumulative revenue of $8.5 trillion by 2030 to recover their investments and earn their target returns. Over the same period, enterprise and consumer AI revenue expected to flow to AI data centers is no more than $3.6 trillion. Subtracting one from the other leaves a gap of $4.9 trillion.
But Gartner's October 14 announcement contains another important point. Gartner forecasts that companies owning large-scale AI infrastructure will capture less than half of all AI revenue.
In other words, even if the companies that provide AI services grow their revenue, that growth does not necessarily translate into profit for the companies that build and operate the data centers. Equipment owners put up huge sums first and bear depreciation and interest costs, and even after they win contracts, it takes time before those contracts can be booked as revenue. In assessing the profitability of AI investment, what matters is not only how far demand grows, but who captures the revenue and who bears the investment risk.
Gartner's $8.5 trillion revenue requirement and the $4.9 trillion gap
According to Gartner's estimate, AI providers will spend about $1 trillion on infrastructure in 2026, with another $6 trillion expected between 2027 and 2030. In total, that works out to roughly $7 trillion in capital expenditure through 2030.
Recovering these investments requires covering not only data center construction and equipment costs but also the financing costs that come with raising capital. Gartner assumed a 12% return on invested capital (ROIC) as the benchmark for the return investors demand, and estimated that the cumulative revenue needed to meet that condition is $8.5 trillion.
However, the 12% ROIC is an assumption Gartner set for its analysis. It is neither the profit margin actually achieved across the AI infrastructure industry nor a common standard that every investor demands.
Meanwhile, revenue expected to flow to AI data centers from AI services used by enterprises and consumers through 2030 is at most $3.6 trillion. Subtracting that from the required $8.5 trillion produces the $4.9 trillion gap.
That said, the $4.9 trillion is the difference between two published figures, not a confirmed loss amount announced by Gartner. It shows the gap between projected future revenue and the amount needed to secure a certain rate of return.
Will Sommer, senior director analyst at Gartner, said in reports including The Register that although AI infrastructure investment is expanding rapidly, revenue from enterprises and consumers alone is unlikely to generate sufficient returns on all the planned capital spending.
Not every company will fare the same, of course. Some will outperform the market on profitability, while others will fail to earn the returns they expected or be forced to exit the market.
Gartner further warned that if the current pace of investment continues, more than $3.5 trillion of AI infrastructure investment could be exposed to asset impairment by the end of 2030. Over the period through 2032, the value of assets exposed to impairment risk could exceed $4 trillion.
In this scenario, Gartner predicts that the 12% ROIC investors demand would not be achieved until the end of 2032 at the earliest.
The outlook was especially harsh for "neoclouds," providers of cloud services specialized in AI and high-performance computing.
Gartner predicted that, owing to rising cost pressure and the difficulty of differentiating their business models, more than 80% of neocloud providers will cease to operate independently by 2030.
This does not mean that more than 80% of these companies will go bankrupt. It also includes cases where they stop being independent companies through acquisitions or mergers. What Gartner is forecasting is a large-scale reshuffling of the neocloud market.
On October 14, neocloud-related stocks also fell sharply. According to Reuters' report of the same day, shares of CoreWeave, Nebius and others plunged, heightening concern about the profitability of AI infrastructure investment. However, wariness over U.S. inflation data and rising interest rates also pushed markets lower that day, so the decline cannot be attributed to Gartner's forecast alone.
Even setting aside the 12% return, investment and revenue are far apart
Gartner's $8.5 trillion revenue requirement assumes a 12% ROIC. Whether that rate is reasonable is open to debate.
But even if the 12% condition is set aside for a moment and only capital expenditure and expected revenue are compared, there is a large gap between the two.
Gartner estimates total AI infrastructure investment from 2026 to 2030 at about $7 trillion. Meanwhile, AI revenue expected to flow to AI data centers over the same period is at most $3.6 trillion.
A simple calculation shows that $3.6 trillion is about 51% of $7 trillion.
In other words, the AI-related revenue projected over the period amounts to only a little more than half the planned capital spending.
That said, this 51% cannot be treated as a recovery rate on invested capital. Assets acquired through capital spending may remain usable beyond 2030, so subsequent revenue and the residual value of the assets must also be considered. In addition, Gartner's revenue forecast covers AI revenue attributable to AI data centers, not all sales of AI-related companies.
It would therefore be inappropriate to conclude from this comparison that half of the capital spending cannot be recovered. Even so, the fact remains that revenue expected over the period is far smaller than the scale of investment through 2030.
Note also that the $7 trillion total includes Gartner's rough estimate of about $1 trillion for 2026. The 51% ratio is therefore an approximate comparison, not a precisely measured figure.
Even more important is how revenue is distributed.
Gartner pointed out that most AI-native companies use an "asset-light" business model, in which they do not own large-scale facilities themselves. As a result, it predicts that companies owning large-scale AI infrastructure will capture less than half of all AI revenue.
This "less than half" forecast is a separate metric from the 51% comparison of capital spending and revenue above.
Still, taken together, the two reveal a structural problem in the AI market. Even if companies providing AI services grow and AI-related spending rises, not all of that revenue goes to the owners of data centers.
Meanwhile, companies that own facilities must keep bearing the cost of GPUs, servers, power equipment and buildings.
Growth in the AI market as a whole may not line up with the investment returns of the companies that own the equipment.
Oracle's $664 billion backlog: about 13% will become revenue in the next 12 months
Oracle's contract backlog is a useful reference for thinking about this structure.
According to the quarterly report (Form 10-Q) for the quarter ended August 31, 2026 that the company filed with the U.S. Securities and Exchange Commission (SEC), its remaining performance obligations (RPO) have reached $664 billion.
RPO is the unfulfilled portion of contract value corresponding to products or services to be provided under customer contracts. Even when a contract has been signed, services not yet delivered are in principle not recorded as revenue.
Oracle's 10-Q outlines revenue recognition as follows:
| Period until revenue recognition | Share of RPO | Approximate amount from $664 billion |
|---|---|---|
| 12 months from August 31, 2026 | About 13% | About $86.3 billion |
| Months 13–36 | About 37% | About $245.7 billion |
| Months 37–60 | About 34% | About $225.8 billion |
| After 60 months | Remaining about 16% | About $106.2 billion |
Note: The amount for each period is an approximation obtained by multiplying the percentage Oracle disclosed by $664 billion.
According to this disclosure, even with a backlog as huge as $664 billion, only about 13%, or roughly $86 billion, is expected to be recognized as revenue in the 12 months through the end of August 2027.
Meanwhile, most of the remaining revenue recognition is expected to stretch over the following several years.
This shows that winning a long-term contract does not mean its value immediately becomes revenue.
Companies that operate data centers must invest in equipment up front in order to provide services. They procure servers and GPUs, build out facilities and bring them online, and then recognize revenue as they deliver services to customers.
During that time, depreciation on the equipment and interest on borrowings keep accruing. Particularly when large capital spending is financed with debt, the pace of monetization and cash management matter even if contract values are enormous.
However, Oracle's RPO also includes contracts for businesses other than AI. And RPO indicates the outlook for future revenue recognition, not profit or cash receipts directly.
The roughly $86 billion therefore cannot be regarded entirely as revenue from AI data centers, nor interpreted to mean that the same amount of cash will be recovered as is.
CoreWeave holds a $104 billion backlog while posting a large loss
The burden that comes with monetizing AI infrastructure also shows up in the results of neocloud providers.
CoreWeave provides GPU-centered computing resources for AI as a cloud service. The company has secured a backlog of about $104 billion, while its full-year 2026 revenue outlook is $12.4 billion to $13.2 billion.
Furthermore, in CoreWeave's second-quarter 2026 results, it posted a net loss of $626 million on revenue of $2.575 billion.
Net interest expense for the quarter reached $640 million. Although revenue has grown substantially, the burden from financing and capital spending is squeezing profit.
Of course, the existence of a net loss alone does not allow the conclusion that CoreWeave's business is unsustainable. In a growth phase where capital spending comes first, profit improvement can lag revenue expansion.
On the other hand, a large backlog does not guarantee future profitability.
What matters is whether contracted services can be delivered as scheduled and whether the resulting revenue comfortably exceeds depreciation and financing costs.
Oracle and CoreWeave both hold huge backlogs, but their scale, business content and revenue recognition periods differ. The ratio of annual revenue to backlog alone cannot be used to directly compare the two companies' monetization speed or management efficiency.
Still, they share one thing: between winning contracts and generating sufficient profit from them lies a burden of time and capital for AI infrastructure companies.
Revenue gathers at Anthropic and OpenAI, while funding burdens remain with equipment owners
What further complicates the problem is the relationship between companies that provide AI services and those that own the computing infrastructure behind them.
For example, Anthropic provides its AI model "Claude" to businesses and developers and earns revenue from usage fees and other charges. Meanwhile, much of the computing resources needed to train and run its models are secured through contracts with major cloud providers.
According to what CNBC reported in August 2026 based on information for investors, Anthropic's annualized revenue reached $65 billion as of the end of July.
However, this is the revenue pace at that point converted to a one-year basis, not a statement that it actually booked $65 billion in annual revenue. It is also a figure based on investor information, so it must be distinguished from audited annual results.
While AI model developers like Anthropic grow their revenue, the cloud providers supplying computing resources must press ahead with capital spending to meet that demand.
That said, not owning equipment does not mean AI model developers bear no funding burden or business risk. Contracts to use large-scale computing resources can carry long-term payment obligations.
It is therefore difficult to split AI companies cleanly into equipment owners and non-owners. Even so, the differences in who holds ownership of the equipment, who bears depreciation and debt, and who earns service revenue are important in judging business profitability.
SoftBank Group raises $11.1 billion through high-yield bonds
The financing that supports massive investment in AI also reflects part of this burden.
On September 24, 2026, SoftBank Group announced that it would issue about $11.1 billion of bonds, combining dollar- and euro-denominated tranches.
The dollar-denominated bonds carry coupons of 8.625%, 9.25% and 9.75% per year, and the euro-denominated bonds 7.125% and 8% per year.
One of the main uses of the proceeds is a $10 billion payment, the final tranche of its additional investment in OpenAI. SoftBank Group said in its October 1 announcement that it had carried out this investment.
This financing is not for SoftBank Group to directly buy AI data center equipment itself. It raised the funds needed for equity investment in OpenAI and similar purposes.
Still, it is clear that investors and funding companies are accepting heavy interest burdens to support the growth of the AI business.
For SoftBank Group, what matters is whether it can earn sufficient returns from its OpenAI investment while paying interest on the bonds. Meanwhile, as investment in OpenAI expands, it also affects the capital spending decisions of companies that supply computing resources in anticipation of its growth.
In the AI market, service providers, data center operators, funding companies and financial institutions are intricately linked, and risk is not necessarily contained within a single company.
Estimate: about half of the four major cloud providers' backlog is concentrated in two companies
Customer concentration is also an important issue.
According to Jefferies' estimate, about half of the roughly $2.3 trillion of backlog held by Microsoft, Alphabet, Amazon and Oracle is tied to two companies, OpenAI and Anthropic.
This estimate suggests that demand for AI infrastructure may depend heavily on specific large customers.
The rapid growth of AI model developers and their expanding large contracts with cloud providers are evidence of strong demand. But when much of that demand is concentrated in a few companies, a change in those companies' business plans or funding conditions can easily affect the suppliers of the equipment.
That said, the "about half" figure is Jefferies' estimate, not the result of each company disclosing amounts by counterparty. Oracle's 10-Q, for example, does not disclose the exact share of RPO accounted for by OpenAI.
The backlog with a specific customer therefore cannot be pinned down from secondary-report figures alone.
OpenAI itself is also deepening its involvement in building AI infrastructure and raising funds through initiatives such as "Stargate." If AI model developers come to own more equipment directly or take on long-term funding burdens, the distinction between equipment owners and service providers will also shift.
Amazon says equipment can't keep up with demand, a bullish view on AI investment
In contrast to Gartner's cautious forecast, some argue that demand for AI computing resources remains extremely strong.
Amazon CEO Andy Jassy has explained that after raising the company's 2026 capital expenditure outlook to about $220 billion, it still cannot secure enough computing capacity to meet all of that year's demand, and that the same situation will continue in 2027.
Amazon's second-quarter 2026 results also showed that AWS growth is accelerating. AWS's backlog reached $496 billion as of June 30, 2026.
Major cloud providers like Amazon have a different business structure from neocloud companies.
Beyond AWS's computing platform for AI, they have a wide range of cloud services as well as retail, advertising and other businesses, so their revenue sources are diversified. Even if the profitability of AI-related equipment deteriorates temporarily, it may be supported by profits and cash flow from other businesses.
Gartner itself points out that this business diversification is a strength for the large hyperscalers.
Goldman Sachs, too, discusses the differences between current AI investment and the past dot-com bubble in a report examining whether AI investment amounts to a bubble.
In particular, many large technology companies generate substantial profits and cash flow from existing businesses. The view is that their financial foundations differ from the dot-com bubble era, when emerging companies expanded equipment relying only on expectations of future revenue.
However, this counterargument does not completely rebut Gartner's forecast.
That large technology companies generate ample profit and that individual neocloud providers can recover enormous capital spending are separate questions.
Furthermore, even if current demand exceeds supply capacity, there is no guarantee that demand will continue at the same prices. If the supply of computing resources grows or the efficiency of running AI models improves, prices and profitability of services that use GPUs could also change.
Gartner forecasts that spending on AI-optimized IaaS will grow from $66 billion in 2027 to $174 billion in 2030.
The market is expected to grow substantially, but Gartner still believes that growth will not be enough for many AI computing infrastructure providers to achieve sufficient investment recovery before the current generation of AI chips is replaced by the next.
In short, strong demand and profitable capital spending do not necessarily go together.
What to watch from 2027 to 2030
To judge whether Gartner's forecast is close to reality, it will be necessary to follow upcoming corporate earnings and moves in market consolidation.
First to watch is the speed at which Oracle's backlog converts into actual revenue.
For its $664 billion in RPO as of August 31, 2026, the company has indicated that about 13% will be recognized as revenue over the next 12 months. That is about $86 billion, covering the period through the end of August 2027.
Note that Oracle's fiscal year ends in late May, so this 12-month period does not coincide only with its fiscal 2027. To check results through the end of August 2027, one needs to look at disclosures through the first quarter of fiscal 2028.
If revenue recognition proceeds as scheduled, it will be evidence of the company's ability to convert a huge backlog into actual revenue.
However, even if about $86 billion of revenue is recognized, that alone does not prove the profitability of AI infrastructure investment. RPO also includes non-AI contracts, and it will be necessary to check how much profit and cash flow remain after deducting equipment and financing costs from revenue.
For neocloud providers, the funding and profitability trends of CoreWeave, Nebius, IREN and others will be important.
In particular, tracking capital expenditure, debt burden, operating computing capacity and the conversion of backlog into revenue together makes it easier to assess business sustainability.
Moves such as acquisitions, mergers and capital alliances will also serve as indicators of market consolidation. However, even if any of the three major players survives as an independent company at the end of 2027, that alone cannot refute Gartner's forecast that "more than 80% will cease independent operation by 2030." To verify the forecast, one needs to look at how many providers across the whole market maintain their independence.
It is also important to see how far AI model developers such as OpenAI and Anthropic come to take on ownership of computing infrastructure and long-term capital spending burdens.
If these companies raise the share of equipment they own, the current structure, in which those earning revenue from AI services and those bearing capital spending risk are separate, may also change.
Gartner's $8.5 trillion revenue requirement and maximum $3.6 trillion revenue outlook show a large gap between the scale of AI infrastructure investment and expected revenue.
But for the AI market going forward, what matters is not the simple question of whether this gap becomes a loss as is. It is, as AI demand expands, who captures the revenue, who takes on the burden of capital spending and financing, and ultimately which companies secure sufficient profit.
While major cloud providers may be able to support investment with revenue from diversified businesses, neocloud providers that depend heavily on AI computing infrastructure are more exposed if monetization is delayed.
The focus going forward will likely shift from the growth rate of the AI market as a whole to whether individual equipment owners can convert their backlogs into profit and cash flow. Oracle's revenue recognition through the end of August 2027 and neocloud providers' financial condition heading into the end of 2027 will be key indicators of how far this structural problem comes to the surface.