Summary

  • Oracle reported remaining performance obligations of $664 billion at the end of the first quarter of fiscal 2027 and cloud infrastructure revenue of $7.4 billion, up 121% year over year (Oracle quarterly figure as reported).
  • In the same quarter the company said it delivered 850 megawatts of AI compute capacity, including more than 300,000 GPUs, and that six of eight buildings at its Abilene, Texas campus were running for 618 megawatts (as reported).
  • Fiscal 2027 capital spending guidance is $90 billion to $95 billion gross, with net cash capex not exceeding $70 billion (as reported) — after a fiscal 2026 path that opened with guidance of about $50 billion (as reported).
  • Trade reporting in March 2026 said the expansion of the Abilene campus from about 1.2 gigawatts toward roughly 2 gigawatts had been dropped after unresolved financing talks, with power delays cited (report); OpenAI's compute manager said the additional capacity would be placed elsewhere (quote as reported).
  • The load-bearing constraint in the disclosures is power procurement and financing timing, not accelerator availability, and the company's own language distinguishes capacity it has "secured" from capacity it has energised.

A backlog that converts on a schedule nobody has published

Oracle's disclosure sequence over three quarters is unusually legible, and that is what makes it useful. Remaining performance obligations, the contracted value of work not yet recognised as revenue, ended the third quarter of fiscal 2026 at $553 billion, up 325% year over year and $29 billion sequentially (as reported). One quarter later the same measure stood at $638 billion, up 363% year over year and $85 billion sequentially, alongside cloud revenues up 47% to $9.9 billion, of which infrastructure grew 93% (as reported). By the first quarter of fiscal 2027 it had reached $664 billion, with more than $30 billion of additional AI cloud contracts booked in the quarter (as reported).

RPO is a contracting measure, not a delivery measure. It counts what customers have agreed to buy, not what they can use today. Oracle has not published, in the material available here, a maturity schedule showing how much of the $664 billion is expected to convert in which quarter, nor the cancellation and termination terms attached to the largest AI contracts. That omission is not unusual for a cloud provider, but it is the reason the backlog figure cannot by itself answer whether the build-out is on schedule. A buyer or an investor reading $664 billion as near-term revenue is reading a number that has not been promised in that form.

The delivery side of the same earnings material gives the sharper comparison. Oracle said that since the end of the fourth quarter of fiscal 2026 it had delivered more than 300,000 GPUs to AI cloud customers, and that it delivered 850 megawatts of AI compute capacity in the first quarter of fiscal 2027, including 131,000 GPUs at the Abilene campus (as reported). Delivered capacity is a company-reported operating metric. It is not independently metered, and the disclosure does not state what share of that megawattage was billing at full contracted rates during the quarter rather than being commissioned, tested or held in reserve for a customer's later start.

At Abilene, the expansion that did not happen

The most informative single event in this period is a decision not to build. In March 2026, multiple reports said Oracle and OpenAI had scrapped plans to expand the Abilene, Texas campus from about 1.2 gigawatts toward roughly 2 gigawatts, after months of unresolved financing talks and shifting capacity needs, with power delays cited as holding the current expansion back (report). OpenAI's compute manager, Sachin Katti, was reported as confirming the decision and explaining that the additional capacity would be placed in other locations (quote as reported).

That framing matters. It describes a change of location rather than a reduction in the total capacity the customer intends to buy, and the reported pressure was financing and power timing rather than a shortage of demand. Six months later, the same site was reported at 618 megawatts across six of eight delivered buildings — roughly three-quarters of total capacity — with earlier reporting having described a two-building campus and six further buildings scheduled to come online in the year, for an estimated 1.2 gigawatts (as reported). The two figures come from different dates and different documents; they are reported here as dated observations rather than reconciled into a single trajectory.

Read together, the March reversal and the September delivery count describe a programme that is advancing building by building while its growth path is being renegotiated site by site. For a customer waiting on contracted capacity, the practical question is which of those two motions is faster.

Power procurement is the pacing item

Oracle's own framing puts electricity at the centre. The company stated that it had secured more than 10 gigawatts of power and data capacity scheduled to come online over the following three years, while industry reporting projected that 30% to 50% of large data-centre projects slated for 2026 would be delayed or cancelled (as reported). The industry-wide projection is a modelled estimate, not a measurement, and it should not be read as an Oracle-specific outcome. It is still the context in which the company's "secured" megawatts should be read: securing power is a contractual and permitting milestone, and it precedes interconnection, turbine delivery, commissioning and billing.

The programme that sets the scale is Project Stargate, announced as a multi-partner initiative targeting up to $500 billion of investment by 2029 and 10 gigawatts of compute capacity, with Oracle as a technology partner (as reported). Programme targets of that kind are announcements by several parties at once; they are not commitments any single one of them has undertaken to fund alone. Oracle's technical claims run in the same direction: the Abilene supercluster is described as powered by its Zettascale10 system, which the company says scales to 131,072 NVIDIA GPUs, with the first NVIDIA Vera Rubin systems due to customers in the following quarter (as reported). Those are supplier and vendor performance claims, corroborated here only by the parties' own public positions.

Capital intensity, and who is funding it

The spending numbers explain why the gap is worth watching rather than merely noting. Fiscal 2026 guidance opened at about $50 billion of capital expenditure against roughly $67 billion of expected revenue (as reported). Fiscal 2027 guidance is $90 billion to $95 billion of gross spending, with net cash capex not exceeding $70 billion and described as unchanged from prior guidance (as reported). The $20 billion to $25 billion between gross and net is the interesting part: it implies that a material share of the build is being funded by customers, partners or financing arrangements that the disclosure does not itemise. Oracle has also stated that a significant portion of its capital expenditure is for revenue-generating equipment tied to committed customer spend.

That structure is not improper, and it is increasingly common. It does mean that the balance-sheet picture and the physical build-out can move at different speeds, and that a reader tracking only net cash capex will see less than the whole programme.

What would settle the question

The gap between contracted backlog and energised capacity is a hypothesis, not a conclusion, and it is falsifiable on published evidence. A disclosure of RPO conversion by maturity, with the cancellation terms attached to the largest AI contracts, would either confirm or dissolve the concern. An executed power purchase agreement or interconnection filing for a named campus would confirm capacity that currently exists only as "secured" in company language. A site-by-site megawatt and billing-start schedule would collapse the gap entirely if the numbers reconciled.

Until one of those appears, the honest reading of Oracle's AI build-out is that contracted demand is running ahead of delivered supply, that the pacing constraint is electrical and financial rather than silicon, and that the next hard numbers — not the next announcements — will decide whether the difference is a phase or a problem.

Oracle Cloud's directory entry is at btw.media.