Summary

  • Oracle reported 850 MW of additional AI datacenter capacity and more than 300,000 GPUs delivered in the quarter it calls Q1 FY27, against $664 billion of remaining performance obligations — a record of execution concentrated in very few places.
  • The same call disclosed a fleet being remarketed rather than retired: 97.9% GPU utilisation and a reported 20% premium on renewed or resold capacity that was first installed roughly four years earlier.
  • Delivery was lopsided. Abilene supplied 131,000 GPUs in the quarter and had six of its eight buildings, 618 MW, handed to the customer; three sites of about 1 GW each — Shackelford in New Mexico, Wisconsin and Michigan — delivered nothing.

Oracle has spent two years building one of the largest contracted AI backlogs in the industry. What it began disclosing in its fiscal 2027 first quarter is the behaviour of the fleet that backlog is meant to be served from. The useful distinction is not between announced and unannounced capacity. It is between capacity that has been energised, occupied and priced, and capacity that exists as a site plan, a power interconnection and a capital commitment.

Two decisions the company disclosed, not one

Oracle's Q1 FY27 release, published on 10 September 2026, is a dual record. On the demand side, remaining performance obligations reached $664 billion, up $209 billion year over year, and total quarterly revenue rose 30% to $19.3 billion, which the company attributed to "strong execution in our infrastructure business, with the delivery of 850MW additional datacenter capacity". On the supply side, the same release states that more than 300,000 GPUs were delivered to AI Cloud customers since the end of Q4 FY26, almost triple the prior quarter's volume.

The comparison is easy to misread. They are not the same quantity measured twice. Backlog counts customer commitments; megawatts count infrastructure that can consume power and be invoiced; GPU deliveries count hardware placed into that infrastructure on a schedule. A quarter can improve on all three for different reasons — and it can improve on one while stalling on another.

What the call added was magnitude and concentration. The 850 MW delivered in Q1 FY27 was described as almost three times everything Oracle delivered in Q4 FY26 and 73% of the total capacity it delivered across the prior fiscal year — a cadence that is steeply non-linear, not a steady quarterly drumbeat. The same call put Abilene-specific delivery at 131,000 GPUs in the quarter, 1.9 times the Q4 FY26 figure there (Q1 FY27 earnings call transcript).

What 97.9% utilisation does and does not establish

Utilisation is the metric that decides whether a datacenter campus is a business or a monument. Oracle reported 97.9% GPU utilisation for Q1 FY27. If that figure is read broadly, it says there is very little idle accelerator time in the fleet Oracle has built — which is what a capacity-constrained market should look like when demand outruns supply.

Three cautions attach to it. First, the disclosure does not define the denominator: whether 97.9% is measured across all installed GPUs, across contracted capacity, or across a billable subset is not stated in the retained material. Second, the figure is a company statement; no independent measurement of Oracle's fleet utilisation is retained, and there is no audited operating report behind it. Third, utilisation near the ceiling has a structural consequence that is easy to overlook: it removes the buffer that normally absorbs reconfiguration, node replacement and customer onboarding.

A fleet running at 97.9% has almost no slack for the fault-and-replace cycle that GPU clusters require.

That is not an argument against the number. It is an argument for reading it as a constraint as much as an achievement. When utilisation is already at the ceiling, every additional megawatt of demand must be met by new commissioned capacity rather than by reallocating what exists.

A 20% premium on four-year-old accelerators

The more revealing disclosure is the pricing of renewals. Oracle said capacity that had been renewed or resold earned a 20% premium over the prior contracts, on GPUs first installed roughly four years earlier (Q1 FY27 earnings call transcript).

Four years is roughly a product generation in AI accelerators. The default expectation for hardware bought at the start of a five-year term is that its economics decay: newer chips deliver more throughput per watt, competitors price the newer generation aggressively, and the older fleet gets discounted to stay occupied. A 20% premium at renewal describes the opposite condition — replacement capacity is scarce enough that four-year-old silicon still commands a higher price than it did when it was new.

The caveats matter. The retained material does not say whether the 20% figure applies to all renewals or to a specific subset, whether it is a like-for-like comparison of the same configuration, or what costs — power, cooling, network, support — attach to the renewed contracts. A premium on the contract line is not the same as a premium on the unit economics, particularly in a campus where the power density per rack has changed since the original installation.

Still, the direction of travel is legible. A fleet whose oldest assets are repricing upward is a fleet whose owner has more pricing power at the end of a contract term than at the beginning — at least for as long as the supply of new, energised capacity stays behind demand.

Abilene shows what a commissioning cadence looks like

The clearest evidence of how Oracle's rollout actually proceeds is campus-level, not company-level. On the Q1 FY27 call, the company described the Abilene, Texas campus as having delivered six of its eight buildings to the customer, representing 618 MW — described as 75% of total campus capacity — alongside $30 billion of AI contract bookings in the quarter that management said were structured so as not to require incremental Oracle capital, and multicloud database revenue growing 353% year over year (Q1 FY27 earnings call transcript, second copy).

Two things follow. The first is that Oracle's headline megawatts are assembled building by building, and a campus can be substantially complete — six of eight buildings, 75% of capacity — without being complete. The second is that the marginal unit of Oracle's AI business is a commissioned building, not a signed contract. Buildings have permitting, utility interconnection, cooling, network fabric and staffing dependencies that contracts do not.

The concentration is also a risk in its own right. If one campus carries the majority of a quarter's GPU deliveries while three others carry none, then quarterly delivery figures are hostage to the construction schedule of individual sites rather than to a portfolio average.

Three gigawatt-scale sites that delivered nothing

The most concrete negative datapoint in the quarter is the status of the greenfield pipeline. Oracle described Shackelford in New Mexico, a Wisconsin site and a Michigan site as roughly 1 GW each, and stated that none of the three was delivered in Q1 FY27 (Q1 FY27 earnings call transcript).

That is not an accusation of failure; large sites take years. It is a statement about the shape of the risk. Between the announcement of a roughly 1 GW site and its first billed megawatt sit land, power, water, equipment lead times and local approvals, and none of those are visible in RPO. Three such sites simultaneously outside the delivery window means the gap between Oracle's contracted obligations and its operating footprint is being managed by a small number of very large, very slow projects.

The counterweight is the company's own account of the pipeline behind them. In March 2026 Oracle stated that it had secured more than 10 GW of power and datacenter capacity through partners, coming online over the following three years, and that more than 90% of that capacity was fully funded. The same post reported more than 400 MW delivered to customers in Q3 FY26, with 90% of committed capacity delivered on or ahead of schedule (From the Q3 Earnings Call).

Read side by side with the Q1 FY27 numbers, those statements establish a delivery history that is real but lumpy: 400 MW in one quarter, 850 MW two quarters later, with a small number of sites responsible for the swing. "Fully funded" is a statement about who pays, not about whether the substation is energised, and the retained material does not define what the threshold measures.

Who is actually carrying the hardware

The funding structure is where the fleet's risk is really allocated. Oracle's fiscal 2026 year-end release put total revenue at $67.4 billion, up 17%, cloud revenue at $34.0 billion, up 39%, and remaining performance obligations at $638 billion, up 363% year over year. It also disclosed that prepaid and customer-supplied hardware portions of large AI contracts totalled $75 billion (Oracle FY26 results).

That $75 billion is the single most important structural detail in the disclosure set, because it moves a large share of hardware ownership away from Oracle. When a customer prepays for GPUs or supplies them outright, the customer carries the capital and, depending on terms that are not public, more of the obsolescence risk. Oracle carries the obligation to build and operate the site that houses them.

The trade-press reading of the quarter adds a conversion frame: OCI revenue of $7.4 billion, up 121% year over year, with Oracle expecting roughly half of the $664 billion backlog to convert over the following 36 months and most new AI contracts using customer prepayments or bring-your-own-hardware structures that require no incremental Oracle capital (Converge Digest). If that conversion framing is accurate, the next three years are less about selling capacity than about commissioning it.

The operating question this leaves open

None of this makes Oracle's position weak. A fleet at 97.9% utilisation whose oldest assets reprice upward is a fleet that is not stranded, and 300,000 GPUs delivered in a single quarter is a real placement of hardware into real buildings. The composition of the disclosure is what deserves attention: the metrics that describe how well the existing fleet is working are strong, and the metrics that describe how fast the next fleet arrives are lumpy and concentrated.

What remains unknown is the specific thing a buyer or an investor would want most. How many of the 618 MW delivered at Abilene, and of the 850 MW delivered in the quarter, sit under contracts whose initial term has already been renewed, and at what price? What commissioning dates can be attached to the roughly 1 GW sites that delivered nothing? And does the 20% renewal premium reflect a broad repricing of installed accelerators or a narrow set of contracts whose comparison base was unusually low?

Without independent measurement, all of the operating figures remain company statements. They are specific, period-labelled and internally consistent — and they are still the seller's own account of its own fleet. The reporting that would settle the question is the one Oracle has not published: a schedule of energised megawatts by site and quarter, alongside renewal pricing. Until that exists, the most defensible reading of Q1 FY27 is that Oracle's AI capacity is scarce, its oldest capacity is still earning, and its future capacity depends on a handful of very large sites finishing.