Oracle GPU Utilization Hits 97.9%, but the AI Bubble Debate Is Not Over

Oracle reported 97.9% GPU utilization after delivering more than 300,000 GPUs, offering a sharp counterpoint to fears of unused AI infrastructure.

The Oracle GPU utilization figure describes activity across its multi-tenant fleet during the first quarter of fiscal 2027. It arrived alongside accelerating cloud revenue, a larger contract backlog, and continued demand for older chips.

That combination challenges one version of the AI bubble argument. Oracle is not struggling to find workloads for installed GPUs. Its immediate constraint appears to be delivering enough data center capacity for contracted customers.

Yet high utilization does not settle the broader debate. Oracle spent heavily to build that capacity, generated negative free cash flow, and accepted long-term obligations tied to data center expansion.

The conflict is therefore narrower than either side suggests. Current demand looks real, but Oracle must turn extraordinary infrastructure usage into durable margins and cash generation.

Amazon Web Services, Microsoft Azure, Google Cloud, and specialized GPU providers face versions of the same test. Oracle’s disclosure gives investors a rare view of how intensely one major cloud fleet is being used.

Oracle GPU Utilization Reached 97.9% as Capacity Expanded

Oracle added capacity at an exceptional pace without producing a visible drop in fleet usage.

Oracle disclosed the utilization figure during its fiscal first-quarter earnings call on September 10, 2026. The quarter ended August 31.

Management said the company delivered 850 megawatts of AI capacity since the end of the previous quarter. That equipment contained more than 300,000 GPUs.

The delivery was almost three times Oracle’s output during the preceding quarter. It also represented 73% of the capacity delivered during the entire previous fiscal year.

The company’s quarterly results reported total revenue of $19.3 billion, up 30% year over year. Cloud revenue rose 62% to $11.6 billion.

Oracle Cloud Infrastructure, or OCI, generated $7.4 billion in infrastructure-as-a-service revenue. That was a 121% increase from the same quarter one year earlier.

Remaining performance obligations reached $664 billion. This accounting measure represents contracted revenue that Oracle has not yet recognized.

The backlog increased by $209 billion from the prior-year period. It gives Oracle considerable visibility, but revenue recognition still depends on successfully delivering infrastructure.

Oracle also said it signed more than $30 billion in additional AI contracts during the quarter. Management said those agreements did not require more Oracle capital.

Some customers prepay for infrastructure, while others supply hardware that Oracle installs and operates. These structures can reduce Oracle’s initial cash burden, although they do not eliminate execution risk.

The 97.9% figure matters because Oracle expanded the fleet while keeping almost all available capacity occupied. Rapid additions often pressure utilization when customer deployments lag behind construction.

Oracle reported the opposite pattern. New capacity entered service, cloud infrastructure revenue accelerated, and fleet usage remained close to full.

Management also offered evidence about older equipment. GPUs reaching renewal were renewed or resold at prices 20% above their previous contracts.

Most of those renewing GPUs were at least four years old, according to the earnings transcript. That result pushes against concerns that each new chip generation immediately destroys the commercial value of older hardware.

Older GPUs cannot match the performance of Nvidia’s newest systems for every training workload. However, they can still serve inference, fine-tuning, development, and less demanding training jobs.

The disclosure does not reveal utilization for every GPU model or customer category. It also does not show how usage converts into margins at the individual cluster level.

Still, Oracle provided more operational detail than cloud providers usually disclose. The numbers show that deployed capacity was busy during the reported quarter.

That changes the immediate question. The issue is no longer whether Oracle can attract demand for its current fleet.

The harder question is whether it can keep delivering capacity, maintain renewal pricing, and earn satisfactory returns after the associated costs.

Near-Full Usage Pressures Every AI Infrastructure Provider

Oracle’s disclosure raises the standard for competitors because it connects AI demand with installed hardware, revenue growth, and contract renewals.

Cloud companies frequently describe AI demand as strong. They rarely publish a fleet-wide GPU utilization rate that outsiders can track from one quarter to another.

Oracle has now provided such a benchmark. At 97.9%, almost all available capacity was assigned to active customer work during the quarter.

That does not make Oracle the largest cloud provider. AWS, Microsoft, and Google operate broader platforms with larger established customer bases.

However, Oracle is positioning OCI around large, dense AI clusters and direct connections to its database business. It is also embedding Oracle database services inside competing clouds.

This strategy lets Oracle participate in enterprise workloads even when customers use Azure, AWS, or Google Cloud as their primary platform. Oracle said multicloud database revenue grew 353% year over year.

The rapid infrastructure expansion places pressure on two competitor groups.

The first group consists of hyperscalers. They must show that their own capital spending supports comparable workload growth, customer retention, and useful infrastructure life.

The second group includes specialized GPU clouds. These providers sell access to scarce accelerators without relying on a large software portfolio for financial support.

Companies in both groups face the same economic tension. High utilization helps revenue, but the hardware, networking, electricity, and buildings require substantial investment.

Oracle’s old-GPU renewals add another competitive complication. A functioning secondary market inside a cloud fleet can extend asset life and reduce the risk of rapid obsolescence.

Workloads do not all require the newest accelerator. A frontier-model training run may prioritize Nvidia Blackwell systems, while routine inference can remain economical on earlier generations.

Cloud operators can therefore segment customers by performance needs. They can reserve new clusters for demanding jobs and redirect older GPUs toward workloads that value availability or lower computing requirements.

That approach works only when software and networking make the fleet manageable. A large collection of chips has limited value if customers cannot schedule jobs efficiently across it.

Oracle’s reported utilization suggests that its operational systems kept pace with hardware deployment during the quarter. It does not independently establish the efficiency of every customer workload.

GPU utilization can also refer to different concepts. Oracle’s documentation defines utilization as GPU activity measured across time, while memory utilization tracks occupied memory separately.

A busy GPU is not necessarily completing economically valuable work at maximum efficiency. Jobs can encounter network delays, data bottlenecks, or poorly optimized code while still registering activity.

Even with that distinction, a nearly full commercial fleet indicates scarce available capacity. It suggests Oracle did not build hundreds of thousands of accelerators without customers ready to use them.

Nvidia benefits from this pressure across providers. Oracle’s expansion creates direct demand for accelerators and supports the case for continued purchases.

Customers face a less comfortable outcome. Tight capacity and premium renewals indicate that buyers should not expect an immediate collapse in AI computing prices.

Enterprises must also decide whether to reserve capacity through long contracts or buy computing resources as needed. Reservations improve availability but create commitment risk.

Developers feel the consequences through quotas, deployment schedules, and model-serving costs. High infrastructure use can delay access even when the underlying chips already exist.

For enterprise buyers, Oracle GPU utilization is therefore more than an investor statistic. It is evidence that the supply chain remains tight at the point of actual deployment.

The Real Reversal Is Demand Versus Delivery

The reported numbers shift Oracle’s near-term problem from finding AI customers to building infrastructure quickly enough to serve them.

AI bubble concerns often focus on stranded assets. Under that scenario, cloud companies buy expensive GPUs before discovering that customer demand cannot support the installed base.

Oracle’s quarter does not fit that description. The fleet remained 97.9% utilized after the company delivered 850 megawatts of additional capacity.

Its remaining performance obligations also reached $664 billion. Although that figure includes more than AI infrastructure, it indicates extensive contracted demand.

The reversal is important because Oracle spent years as a secondary public-cloud provider. AWS, Azure, and Google Cloud established their positions earlier and built larger developer ecosystems.

Generative AI created another route into the market. Model developers need huge, tightly connected GPU clusters, and capacity availability can matter more than existing cloud share.

Oracle designed large clusters around that opportunity. Its Abilene, Texas, campus became a prominent test of whether the company could translate infrastructure plans into working systems.

During the quarter, Oracle said it delivered 131,000 GPUs at Abilene. Six of eight campus buildings were operational, representing 618 megawatts and 75% of planned customer capacity.

These figures support a delivery-centered interpretation of the business. Oracle already holds contracts, but it recognizes infrastructure revenue as capacity becomes available.

Backlog alone cannot complete that conversion. Oracle needs buildings, electricity, cooling equipment, networking, GPUs, and qualified teams at the right locations.

Any delay can move revenue into a later quarter. A prolonged delay can also damage customer relationships or force contract changes.

This is why the 98% headline should not be read as a universal verdict about AI demand. It describes Oracle’s active fleet during one quarter.

The figure does not measure whether every planned data center has a customer. It also does not reveal the profitability of contracts attached to future sites.

It does show that Oracle’s installed assets were not sitting idle. That distinction matters because installed capacity and announced future capacity carry different risks.

Oracle’s present fleet faces demand risk, operating risk, and pricing risk. Its development pipeline adds construction, financing, power, and supply-chain exposure.

Strong utilization reduces the first risk. It does not automatically solve the others.

Renewal pricing provides another useful signal. Oracle said every GPU contract expiring during the quarter was renewed or resold at a 20% premium.

A premium on four-year-old GPUs suggests continued demand beyond frontier training. It also supports longer useful-life assumptions for at least part of the fleet.

However, one quarter of renewals cannot establish a permanent pricing curve. Competitive capacity, more efficient models, and customer-owned hardware can change future negotiations.

The strongest conclusion is therefore specific. Oracle’s deployed AI infrastructure experienced nearly complete commercial use during the quarter.

That evidence weakens claims that the current buildout has already produced widespread idle capacity. It does not guarantee the same outcome across the industry.

Oracle now has to repeat the performance while adding more capacity. At its current scale, execution matters as much as demand.

What the 98% Figure Does Not Resolve

Near-full GPU usage supports Oracle’s revenue story, but cash flow and capital intensity remain the central challenge.

Oracle recorded $23.1 billion in operating cash flow during the quarter. Capital expenditures reached $28.5 billion, leaving free cash flow at negative $5.4 billion.

The company had already generated negative free cash flow across fiscal 2026. Its spending reflects the cost of data centers, computing equipment, and related infrastructure.

This is the strongest answer to an overly optimistic interpretation of Oracle GPU utilization. Busy assets can still produce disappointing returns when construction and financing costs rise faster than revenue.

Oracle’s revenue expanded quickly, and operating income improved. Yet the company must sustain that growth long enough for operating cash generation to overtake infrastructure investment.

An analysis of cash flow highlighted the contrast between strong AI-related sales and continued negative free cash flow. Oracle carries a different risk profile from cash-rich hyperscalers funding similar construction.

AWS, Microsoft, and Google can finance AI infrastructure with cash generated by larger businesses. Oracle also has mature software revenue, but its balance sheet has less room for repeated mistakes.

Customer financing structures help manage the burden. Prepayments and bring-your-own-hardware agreements move some initial capital requirements away from Oracle.

Those arrangements can also complicate comparisons. A contract funded by a customer does not create the same capital exposure as an Oracle-owned deployment.

Investors need more detail about the mixture of those contracts. They also need to understand margins after electricity, leases, depreciation, maintenance, and financing.

Remaining performance obligations require similar caution. RPO represents contracted business, not cash already collected or revenue already earned.

The $664 billion balance is exceptionally large relative to current quarterly revenue. Its economic value depends on timing, customer concentration, contract terms, and successful delivery.

Oracle said roughly half of the backlog should convert into revenue over 36 months. That conversion would require sustained construction and deployment across multiple sites.

Customer concentration remains another concern. Large AI infrastructure contracts can make quarterly progress depend on a limited number of model developers.

Prepayments reduce credit and financing exposure, but they do not remove reliance on those customers’ long-term plans. Model economics may change before every contract reaches maturity.

Technical change introduces a separate risk. Four-year-old GPUs commanded higher renewal prices this quarter, but future chips can alter performance per watt and per unit of computing.

AI software is also becoming more efficient. Smaller models, quantization, sparse architectures, and improved inference systems can reduce computing requirements for specific tasks.

Efficiency does not always reduce total infrastructure demand. Lower costs often encourage more usage, new products, and larger workloads.

Still, Oracle should not treat current scarcity as permanent. Capacity entering the market from hyperscalers and specialized providers could weaken renewal pricing.

An independent equity assessment identified on-time capacity delivery as the key link between Oracle’s backlog and revenue growth. Project delays or supply bottlenecks remain clear downside risks.

The 97.9% figure itself also needs consistent disclosure. Investors cannot evaluate a trend if Oracle changes the calculation or stops publishing it.

Useful future reporting would separate utilization by chip generation, workload type, and ownership model. Oracle has not provided that level of detail publicly.

Margin disclosure would add even more value. High usage at weak pricing can produce revenue without sufficient returns, although the reported renewal premium currently points elsewhere.

For now, Oracle has established that demand exists for deployed capacity. It has not established the ultimate return on its full investment program.

That is why the quarter challenges bubble fears without ending them. The utilization data answers an operating question, while the cash-flow statement raises a financial one.

Three Signals Will Test Oracle’s AI Infrastructure Case

The next phase depends on utilization stability, backlog conversion, and a credible path back to positive free cash flow.

The first signal is Oracle’s next reported GPU utilization rate. Investors should compare it directly with 97.9% rather than treating the current figure as permanent.

Stable utilization during another large capacity increase would strengthen Oracle’s demand case. A meaningful decline would indicate that deliveries are beginning to outpace customer deployment.

Renewal pricing belongs in the same test. Continued premiums for older GPUs would support Oracle’s claims about useful life and residual value.

Discounting would tell a different story. It could mean that new supply, better chips, or more efficient models are weakening demand for earlier hardware.

The second signal is conversion of the $664 billion backlog into reported OCI revenue. Contracted demand matters only when Oracle delivers capacity and begins recognizing sales.

OCI revenue grew 121% during the latest quarter. Future results must show that this acceleration can continue as new sites enter service.

Delivery progress at major campuses will provide an early indication. Power availability, construction schedules, and equipment installation can all constrain recognized revenue.

Investors should also watch how much new contracting comes from customers beyond Oracle’s largest AI partners. Broader demand would reduce concentration risk.

The third signal is free cash flow. Oracle’s negative $5.4 billion result demonstrates the cost of expanding before all associated revenue arrives.

A path toward positive free cash flow would strengthen the argument that high utilization supports an economically durable business. Continued deterioration would keep financing risk at the center of the debate.

Capital expenditures should be read together with operating cash flow, customer prepayments, and hardware ownership. A single spending figure cannot capture the full structure.

Oracle’s financial statement gives readers the essential starting point. The company generated strong operating cash, but infrastructure spending remained even larger.

Enterprise buyers should watch these indicators for practical reasons. Capacity scarcity affects contract terms, deployment timing, and access to newer accelerators.

Developers should track whether older GPUs remain available for inference and fine-tuning. Those systems can expand access even when frontier clusters stay fully reserved.

Knowledge workers will experience the outcome indirectly through AI product availability, response speed, and usage limits. Infrastructure economics eventually reach software customers.

Teams evaluating AI vendors should preserve their own deployment evidence, contract assumptions, and workload results. A searchable knowledge base can keep those decisions tied to actual technical records.

Oracle GPU utilization has supplied a valuable data point in a debate often dominated by forecasts. Its installed fleet was nearly full, and customers paid more to renew older capacity.

That evidence contradicts the claim that Oracle’s current GPU inventory lacks real demand. It does not prove that every planned facility will earn an attractive return.

The next few quarters will reveal whether 97.9% marks a repeatable operating pattern or a peak created by temporary scarcity. Watch utilization, backlog conversion, and free cash flow together.