Has Oracle spent too much on AI?
In our AI infrastructure market deck, you will find everything you need to understand the market
SUMMARY
Yes. Oracle has spent too much on AI relative to its balance sheet, even though the demand behind the expansion is real.
The problem is not that Oracle built empty data centers. OCI revenue is growing quickly, GPU utilization is close to full, and the company has signed an enormous backlog of contracted business.
The problem is that Oracle is funding a hyperscaler-sized construction program with a much smaller cash engine than Microsoft, Amazon or Google. Capital expenditure reached 83% of revenue and 174% of operating cash flow in fiscal 2026.
Oracle’s backlog reduces demand risk but does not remove financing risk. Much of the revenue will arrive years after Oracle has paid for buildings, power, networking and equipment.
OpenAI is both the strongest part of the investment case and its biggest weakness. A roughly $300 billion commitment gives Oracle a huge anchor customer, but S&P estimates that OpenAI represents about half of Oracle’s backlog.
The reported 97.5% GPU utilization rate is impressive, although the renewal data shows that a small number of large customers control most of the hardware. The customer count looks broader than the economic exposure really is.
The hardest commitments are not necessarily the GPUs. Oracle disclosed roughly $260 billion of future operating lease commitments, mostly for data centers that may run for 15 to 19 years once they begin.
Management says mature infrastructure projects can produce returns on invested capital in the high 20% range and OCI margins of 30% to 40%. Those economics would justify the buildout, but investors have not yet seen them across the new fleet of giant campuses.
Oracle can still finance the expansion because it remains investment grade, owns a large recurring software business and has access to debt and equity markets. The downgrade to BBB- shows that the cushion is already thin, not theoretical.
The decisive test will come after the main sites open. Cash flow should recover, debt issuance should slow, OCI margins should rise and OpenAI should become a smaller share of the backlog.
Oracle may still produce an excellent outcome. But the company has accepted so much debt, dilution, customer concentration and long-term lease exposure that several uncertain things now need to go right at the same time.
This market map, featured in our AI infrastructure market deck, highlights top companies and startups in the AI infrastructure market
Why is Oracle’s AI spending so controversial now?
Oracle’s AI spending is controversial now because the company is building at hyperscaler speed with far less financial room than Microsoft, Amazon or Google.
Oracle’s latest annual results put the problem in plain numbers. Revenue rose 17% to $67.4 billion and Oracle Cloud Infrastructure revenue jumped 77%, even as capital expenditure absorbed 83% of total revenue and 174% of operating cash flow. Free cash flow ended the year at negative $23.7 billion.
S&P Global then lowered Oracle’s credit rating to BBB-, one level above speculative grade. The agency pointed to rising capital needs, uncertain AI infrastructure profitability, intense competition and heavy customer concentration. It had already called Oracle’s infrastructure strategy the most aggressive among major technology companies in terms of leverage.
Oracle is trying to become a major AI infrastructure provider faster than its own cash generation allows. Signed contracts support the expansion. Debt, equity, customer prepayments and long leases cover the years between construction and payment. The uncomfortable question is how much risk Oracle accepted to win that future revenue.
What would “too much AI spending” mean for Oracle?
Oracle has spent too much on AI if the company needs unusually favorable assumptions about customers, margins and financing for the investment to pay off.
A huge capital program can still be sensible when demand is contracted, assets stay busy and returns comfortably exceed the cost of funding. Oracle already has evidence on the first two points. Its backlog is enormous, cloud infrastructure revenue is growing quickly and delivered GPUs are almost fully used.
The weaker parts of the case are timing, concentration and proven returns. Oracle pays for construction and equipment before recognizing most of the related revenue. A large share of future business appears linked to OpenAI. Management expects attractive infrastructure returns once sites reach full contractual revenue, but investors have yet to see those economics across the new fleet of giant campuses.
Our test is simple: can Oracle turn its AI contracts into strong cash flow without constant fundraising, near-perfect construction and one dominant customer staying financially healthy? Right now, it falls short.
If you want more recent data on this point, please see our latest AI infrastructure market report.
As this chart shows, and as featured in our AI infrastructure market deck, search interest in AI infrastructure has risen sharply
How extreme is Oracle’s AI spending compared with Microsoft and Meta?
Oracle is spending far more aggressively than Microsoft or Meta when we compare capital expenditure with revenue and operating cash flow.
Oracle invested $55.7 billion during fiscal 2026, up 162% in one year. Revenue grew by $10 billion over the same period. Capital expenditure increased by more than $34 billion. Oracle has guided to roughly $70 billion of net capital cash outlay in fiscal 2027. Reported capital expenditure could reach about $90 billion to $95 billion once customer prepayments and financing-related timing are included.
Microsoft and Meta also spend heavily on AI infrastructure. Both companies funded their latest completed annual programs from operating cash flow and remained strongly free-cash-flow positive. Oracle spent $1.74 on capital expenditure for every $1 generated from operations. Microsoft spent about $0.47 and Meta about $0.62.
The absolute totals hide the real difference. Meta invested more dollars than Oracle. Microsoft invested a similar amount. Both companies had much larger revenue bases and cash engines. Oracle attached a hyperscaler-sized construction program to a business that generated $32 billion of operating cash flow.
| Latest completed year | Oracle | Microsoft | Meta |
|---|---|---|---|
| Revenue | $67.4B | $281.7B | $201.0B |
| Capital expenditure | $55.7B | $64.6B | $72.2B |
| Operating cash flow | $32.0B | $136.2B | $115.8B |
| Capital expenditure as a share of revenue | 83% | 23% | 36% |
| Capital expenditure as a share of operating cash flow | 174% | 47% | 62% |
| Free cash flow | Negative $23.7B | About $71.6B | $43.6B |
Did Oracle build AI data centers before finding customers?
Oracle already has real customers for its AI data centers, so empty speculative capacity is currently a secondary concern.
Oracle finished its latest fiscal year with $638 billion in remaining performance obligations, up 363% from the previous year. Oracle said most of the recent increase came from large AI infrastructure contracts. Four customers each committed more than $8 billion during the latest quarter, and Oracle delivered more than 1.2 gigawatts of customer capacity during the year.
Actual usage also looks strong. On Oracle’s latest earnings call, management reported 97.5% global GPU utilization. Among 35,000 GPUs reaching renewal across 59 customers, 49% of the customers renewed, representing 92% of the GPUs. Oracle said most of the remaining capacity was sold to other customers during the same quarter.
The renewal figures also expose the concentration behind the high utilization. Oracle can quickly reassign capacity when smaller customers leave. A limited group of large users still controls most of the hardware. The demand is real, just narrower than the customer count first suggests.
This chart, included in our AI infrastructure market deck, shows annual VC investment in AI infrastructure startups
How much of Oracle’s AI backlog will turn into revenue soon?
Oracle’s AI backlog gives the company years of contracted revenue, but the money arrives too slowly to fund the current buildout by itself.
Oracle’s chief financial officer expects the company to recognize 12% of its $638 billion backlog during the next 12 months, equal to about $76.6 billion. Another 34%, or roughly $216.9 billion, should arrive during months 13 through 36. More than half sits beyond that window. These figures cover Oracle’s wider business, including software and applications, rather than AI infrastructure alone.
The payment schedule lets Oracle report record contracted revenue alongside deeply negative free cash flow. Data centers, power systems and networking must be prepared before customers consume the contracted capacity. Several large projects will only begin contributing meaningful revenue after years of spending.
Customers are carrying part of the upfront cost. Oracle says $75 billion of large AI contracts involves prepaid GPUs or customer-supplied hardware. Its fiscal 2026 cash-flow statement recorded a much smaller $4.6 billion increase in deferred revenue from customer prepayments with a significant financing component. The $75 billion figure describes contract structures rather than cash already sitting in Oracle’s bank account.
The backlog is far stronger than a loose order pipeline. Its value still depends on construction, margins, customer finances and years of successful delivery.
| Oracle backlog timing | Share of backlog | Approximate revenue |
|---|---|---|
| Expected within 12 months | 12% | $76.6B |
| Expected during months 13 to 36 | 34% | $216.9B |
| Expected after 36 months | 54% | $344.5B |
| AI contracts with prepaid or customer-supplied GPUs | About 12% of total backlog | $75.0B |
Has Oracle become too dependent on OpenAI?
Oracle is now too dependent on OpenAI for a company taking on this much debt and long-term infrastructure risk.
The Wall Street Journal reported that OpenAI agreed to buy roughly $300 billion of Oracle computing capacity over about five years. Oracle’s backlog is unusually concentrated: S&P estimates that OpenAI represents roughly half of it and describes the customer as a key credit risk.
An even five-year split would average about $60 billion a year, close to Oracle’s entire fiscal 2026 revenue. OpenAI’s chief financial officer said its annualized revenue had passed $20 billion, which leaves a vast gap between the customer’s current scale and the average value of the Oracle commitment. Actual invoices will follow capacity delivery and vary from year to year. Even so, Oracle’s growth plan assumes that OpenAI becomes much larger and keeps raising extraordinary amounts of capital.
OpenAI has shown extraordinary growth and access to capital, including a reported $122 billion funding round. The Wall Street Journal has since reported that OpenAI’s planned cloud spending had climbed to about $750 billion through 2030. It also reported that Nvidia was discussing a guarantee of up to $250 billion for a separate OpenAI data-center project in Ohio. Those figures keep the expansion credible, but they also show how heavily it depends on outside financing.
Oracle gains a huge anchor tenant. OpenAI keeps several infrastructure suppliers. OpenAI can redirect future growth among providers more easily than Oracle can redirect a campus designed around an anchor contract. That imbalance makes the relationship especially risky.
If you want more recent data on this point, please see our latest AI infrastructure market report.
This chart, included in our AI infrastructure market deck, shows why CoreWeave is winning in AI infrastructure
Are Oracle’s AI data centers already generating good returns?
Oracle’s AI data centers are growing fast. The returns on the new infrastructure remain unproven.
Oracle Cloud Infrastructure revenue reached $18.1 billion in fiscal 2026 and grew 77%. In the final quarter, infrastructure revenue rose 93%. Operating cash flow increased by $11.2 billion during the year, which confirms that the cloud expansion is already adding real cash to Oracle’s business.
Capital spending grew three times faster in dollar terms. Oracle added roughly $34.5 billion of annual capital expenditure against an $11.2 billion increase in operating cash flow. The gap may narrow when more contracted capacity starts producing revenue, although the company is moving deeper into the investment phase with another large increase planned.
Management says mature infrastructure projects can earn returns on invested capital in the high 20% range and that OCI margins could eventually reach 30% to 40%. Those targets would make the strategy attractive. They remain forecasts built around full contractual utilization, successful delivery and expected pricing.
Oracle’s overall gross margin has already fallen by roughly five percentage points as infrastructure becomes a larger part of revenue. Cloud capacity needs chips, power, cooling, buildings and regular equipment replacement. Oracle’s older software model earned far higher margins from each additional sale. The promised growth now has to compensate for a structurally more expensive business.
Can Oracle keep financing its AI expansion safely?
Oracle can keep financing its AI expansion for now, but the company has already lost much of the financial cushion that made the bet possible.
Oracle ended fiscal 2026 with about $31.9 billion of cash and marketable securities and $129.5 billion of borrowings. That leaves net borrowings close to $97.6 billion, slightly above three years of the latest operating cash flow if that cash flow stayed flat. Annual interest expense rose 29% to $4.6 billion.
During the year, Oracle raised $43 billion in debt and $5 billion in equity. Management expects another $40 billion of debt and equity financing during fiscal 2027, including a $20 billion program that can sell common shares into the market. Equity protects the credit profile, but existing shareholders carry the dilution.
S&P’s recent downgrade to BBB- shows that the balance-sheet pressure is already real. Oracle remains investment grade, has a large recurring software base and can still access capital markets. Another major disappointment could raise borrowing costs, trigger collateral requirements and make future projects harder to finance.
A recent Wisconsin dispute shows how quickly credit quality can affect operations. Utility rules connected to a planned data-center campus could require Oracle to provide more than $7 billion of collateral after the downgrade, according to recent reporting. Oracle is challenging the requirement. Even if it succeeds, the dispute shows that weaker credit can create cash demands far away from the bond market.
If you want more recent data on this point, please see our latest AI infrastructure market report.
This chart, included in our AI infrastructure market deck, shows annual funding in AI infrastructure startups
Are Oracle’s long-term data-center leases more dangerous than the GPUs?
Oracle’s long-term data-center leases create a bigger downside risk than the GPUs because buildings and power commitments are much harder to move.
Oracle disclosed about $260 billion of additional operating lease commitments in its annual filing, substantially all related to future data centers. The leases had yet to begin, leaving them outside the reported lease liabilities on the balance sheet. Many are expected to run for 15 to 19 years once they start.
Oracle’s cloud infrastructure is designed to be multi-tenant, and the company has shown that GPUs can be reassigned when customers leave. The latest renewal cohort supports that claim: capacity representing most of the non-renewed GPUs found other buyers during the same quarter.
A purpose-built campus creates a different problem. Power contracts, substations, cooling systems and long leases remain tied to a location. Oracle may find another customer for thousands of GPUs and still struggle to fill several gigawatts of capacity on the same economic terms. The risk becomes larger if multiple cloud providers bring new supply online at once.
The leases secure scarce power and construction capacity, but they reduce Oracle’s freedom to slow down later. They could stay expensive long after a particular generation of GPUs becomes outdated.
Does Oracle have a real advantage in AI infrastructure?
Oracle has a real AI infrastructure advantage, although it is narrower than the spending program suggests.
Oracle has become unusually good at delivering large GPU clusters quickly. The company brought more than 1.2 gigawatts of customer capacity online during fiscal 2026 and expected the following quarter’s delivery to approach another gigawatt. High utilization shows that Oracle has matched scarce capacity with customers effectively.
Oracle also brings assets that specialist GPU clouds lack. Its databases contain core financial, healthcare, supply-chain and customer information for large organizations. Oracle’s multicloud database business grew 404% in the latest quarter, allowing customers to use Oracle databases inside other cloud platforms. That creates a stronger link between enterprise data and AI workloads.
Bulk compute offers Oracle less protection. Microsoft, Amazon, Google, CoreWeave and several newer providers are expanding aggressively. Power access and fast construction currently command a premium. Those advantages may fade as industry supply grows. Databases, networking, reliability and enterprise relationships will decide whether Oracle keeps its lead.
Oracle has earned a place among the serious AI infrastructure providers. Whether that position requires almost an entire year of revenue in annual capital spending remains doubtful.
This chart, included in our AI infrastructure market deck, compares the main business model options for AI cloud infrastructure providers
What happens to Oracle if AI compute becomes much cheaper?
Cheaper AI compute would first squeeze Oracle’s renewal prices, then threaten the economics of its long-lived data centers.
AI computing costs can fall through newer chips, better networking, smaller models, quantization and more efficient inference. Lower prices can also create more demand because companies run more models and automate more tasks. Oracle’s current utilization suggests that rising usage is still absorbing efficiency gains.
The risk grows when new supply and efficiency arrive together. Customers may find comparable capacity at lower prices as Oracle continues paying for sites agreed during a shortage. Long contracts can protect revenue for a period. Customers facing much cheaper alternatives may still renegotiate, reduce later commitments or move new workloads elsewhere.
Oracle says some contracts include mechanisms that pass unexpected component-cost increases to customers. Those clauses help with inflation in chips or memory. Falling market prices create the opposite problem and are harder to pass through.
The current capacity shortage favors Oracle. Many sites will earn most of their return in a market that may have far more power, chips and competitors than the one in which the contracts were signed.
If you want more recent data on this point, please see our latest AI infrastructure market report.
What would prove that Oracle’s AI spending was too high?
Oracle’s AI spending will be proven excessive if the new data centers mature without restoring cash flow, margins and customer diversification.
Construction-phase negative free cash flow is only an early warning. Oracle deliberately spends before revenue starts. The decisive test comes when the largest contracted sites are fully operating and the company has had time to collect from them.
We would expect three clear improvements from a successful buildout. OCI margins should move toward management’s 30% to 40% range. Operating cash flow should begin covering most of the annual capital program. OpenAI should become a smaller share of Oracle’s backlog as other customers grow.
Debt also needs to peak. Repeated large financing rounds after the new campuses reach full revenue would suggest that the projects cannot support their own expansion. Another credit downgrade would make that conclusion harder to avoid.
| What to watch | Evidence Oracle’s spending is working | Evidence Oracle spent too much |
|---|---|---|
| OCI margins | Move toward 30% to 40% | Stay well below management’s target |
| Free cash flow | Recovers as contracted sites open | Remains deeply negative after the main sites mature |
| Customer mix | OpenAI becomes a smaller share | One customer continues to drive about half of the backlog |
| Financing | Debt and dilution slow sharply | Large debt or equity raises continue |
| Utilization | New capacity stays heavily used | Utilization falls as industry supply expands |
| Credit rating | Oracle remains securely investment grade | A further downgrade raises funding and collateral costs |
This chart, featured in our AI infrastructure market deck, shows the share of revenue generated by each customer segment in the AI infrastructure market
Has Oracle spent too much on AI?
Yes. Oracle has spent too much on AI relative to its balance sheet, even though the company has correctly identified a huge and genuine demand opportunity.
The positive evidence is substantial. OCI revenue is rising at exceptional speed. Oracle’s delivered GPUs are almost fully utilized. Customers have signed enormous long-term contracts, supplied hardware and made prepayments. Oracle also has a valuable enterprise position through its databases and applications.
The size and structure of the commitment have gone too far. Oracle has accepted a capital program that exceeds its operating cash flow and pushed borrowings above $129 billion. Oracle has also accumulated roughly $260 billion of future lease commitments and tied a large part of its backlog to OpenAI. S&P’s recent downgrade confirms that this combination has already weakened Oracle’s credit profile.
Oracle may still earn excellent returns. Strong demand could continue, OpenAI could grow into its commitments and the new campuses could reach the high margins management expects. But the company has left itself little room for a delayed site, a weaker customer, lower compute prices or another financing shock.
Our judgment goes beyond calling the bet risky. Oracle has overextended itself to capture the AI infrastructure boom. The spending can still produce a great outcome, but several uncertain things now need to go right together. A company that must execute close to perfectly has usually spent beyond the prudent limit.
If you want more recent data on this point, please see our latest AI infrastructure market report.
OUR METHODOLOGY
This analysis tests whether Oracle has spent too much on AI by comparing the scale of its infrastructure investment with its financial capacity, contracted demand, utilization, customer concentration, expected returns, financing requirements, long-term lease commitments and competitive position.
We prioritized observed results over forecasts. Reported revenue, OCI growth, capital expenditure, operating cash flow, free cash flow, debt, backlog, GPU utilization and signed commitments carry more weight than management targets for future margins or returns on invested capital.
Oracle’s $638 billion of remaining performance obligations is treated as contracted business rather than a loose sales pipeline. It is not treated as cash already received, and it includes Oracle’s wider software and applications business rather than AI infrastructure alone.
We separated demand risk from timing and financing risk. High utilization and signed contracts show that Oracle has found customers, but they do not guarantee that revenue will arrive quickly enough, at high enough margins, to fund the construction program without repeated debt or equity raises.
The comparison with Microsoft and Meta is designed to measure financial intensity rather than absolute spending. We compare capital expenditure with revenue and operating cash flow because a similar dollar investment can create very different balance-sheet pressure at companies with different cash engines.
OpenAI concentration is assessed using reported contract values and S&P’s estimate of its share of Oracle’s backlog. The roughly $60 billion annual figure is only an even five-year average used to show scale; actual invoices will depend on capacity delivery and will vary from year to year.
We treat Oracle’s infrastructure return targets, including OCI margins of 30% to 40% and returns on invested capital in the high 20% range, as management forecasts rather than proven fleet-wide economics. The conclusion can improve if those targets appear in reported cash flow and margins as the largest campuses mature.
Key sources include Oracle’s fiscal 2026 results, Oracle’s fiscal 2026 Form 10-K, Oracle’s Q4 earnings commentary, Oracle’s Financial Analyst Meeting materials, S&P Global Ratings on Oracle’s downgrade, S&P Global Ratings on AI infrastructure credit risk, Microsoft’s 2025 annual report, Meta’s 2025 full-year results, The Wall Street Journal on the Oracle-OpenAI computing agreement, and OpenAI’s financing disclosure.
This chart, included in our AI infrastructure market deck, shows how GPU cloud infrastructure technology has evolved over time
Related blog posts
