According to internal Amazon emails obtained by Business Insider, Amazon Web Services (AWS) planned to receive about $5 billion worth of NVIDIA GB300 server racks in August 2026, before they were finally installed in data centers. But some U.S. sites that were meant to host the racks reportedly had not secured the power needed to run them. Meanwhile, Amazon's financial results show that its trailing-twelve-month free cash flow (FCF) turned to an outflow of $7.6 billion, and that total future payments on leases that have been signed but not yet commenced exceeded $137.2 billion as of the end of June 2026. These figures do not represent the value of GPUs sitting idle in warehouses; they show how enormous capital spending and pre-operation contracts are piling up. Read alongside Amazon's financial filings, records of its dispute with a utility, and Microsoft's explanation of how it operates its equipment, the report highlights a challenge that is holding back the expansion of AI infrastructure: the time it takes to secure power and bring equipment into operation.
Early receipt of roughly $5 billion in GB300 racks, with some sites unable to run them for lack of power
According to internal emails reported by Business Insider on October 20, AWS had introduced a new practice called "Early Rack Receipt." Previously, AWS took ownership of completed racks when they arrived at the data center. Under the new approach, it receives racks before they are delivered to their final installation site, allowing payment and asset recognition to proceed earlier. AWS expected to receive about $5 billion worth of NVIDIA GB300 racks in August 2026 and, for the time being, roughly $3 billion worth each month after that. Converted at the Bank of Japan rate as of 5 p.m. on October 21, 2026 (¥158.12 per dollar), the August batch alone comes to about ¥790.6 billion, and each subsequent month to about ¥474.4 billion.
The problem is that racks that have been received cannot necessarily be put into operation right away. According to people Business Insider spoke with, some AWS sites in the U.S. had floor space for additional racks but lacked the power needed to run them. As a result, completed racks were not delivered directly to data centers and were instead held in intermediate warehouses. The internal emails also noted that at least one team had no clear outlook on when the racks it was due to receive could be moved into data centers. However, details such as the number of racks in storage, the sites affected and how long they were stored have not been disclosed.
Purchased GPUs that cannot run because of power shortages or other facility constraints are sometimes called "dark GPUs." For AWS, early receipt has the advantage of securing the latest GPUs, which are in limited supply, ahead of competitors. On the other hand, the equipment generates no revenue until it is up and running and used by customers. With expensive AI servers in particular, the longer the gap between receipt and operation, the more capital is tied up in idle equipment.
An Amazon spokesperson told Business Insider that temporarily storing completed equipment near its installation site is standard practice and "best business practice." The spokesperson also pushed back on the view that the company as a whole lacks visibility into its rack installation plans, saying Amazon has long had systems for tracking infrastructure assets. On the July 30 earnings call, Amazon CEO Andy Jassy also said that servers and networking equipment are typically purchased several months before they go live. Still, buying earlier is different from receiving and storing completed racks at a location other than their final installation site, and the two should be considered separately.
On the same call, Jassy said investments in servers and networking equipment pay back in under three years on average, and that the equipment has a useful life of at least five to six years. The idea is that if there is sufficient demand once equipment goes live, it can keep generating cash flow for years after the investment is recouped. But if power and buildings are not ready, the point at which purchased equipment starts earning revenue is pushed back as well. What matters in this report is not securing GPUs early in itself, but how long the wait between receipt and operation becomes.
Free cash flow turns to a $7.6 billion outflow; leases not yet commenced up 42% in six months
In its second-quarter results, announced on July 30, 2026, Amazon reported trailing-twelve-month free cash flow (FCF) through June 30 of negative $7.6 billion. A year earlier it was an inflow of $18.2 billion, so the figure worsened by nearly $25.8 billion in a year. Trailing-twelve-month FCF fell from $25.9 billion as of the first quarter of 2025 to $1.2 billion as of the first quarter of 2026, then turned negative in the second quarter of 2026. Trailing-twelve-month operating cash flow, meanwhile, reached $161.4 billion, so cash earned from the core business has not declined. Amazon attributed the FCF deterioration mainly to a $66.1 billion year-over-year increase in net spending on property and equipment (purchases less proceeds from sales and incentives), most of which it said stemmed from AI-related investment.
Quarterly capital spending has also surged. Purchases of property and equipment in the second quarter of 2026 totaled $54.2 billion, well above $32.2 billion a year earlier. On the July 30 earnings call, Jassy also raised the company's 2026 cash capital expenditure outlook from roughly $200 billion to roughly $220 billion (about ¥34.8 trillion). He cited rising memory prices as the reason for the upward revision. Investment to meet expanding AI demand has reached a scale that exceeds the cash flow generated by operations.
Beyond this rise in spending, assets and contracts that Amazon has not yet begun to use are also swelling in its financial filings. The quarterly report (Form 10-Q) as of June 30, 2026 includes a line item for "Leases not yet commenced." It totals the amounts Amazon expects to pay in the future on leases that have been signed but for which use of the underlying asset has not yet begun. The total reached $137.214 billion (about ¥21.7 trillion). Compared with $96.373 billion at the end of 2025, that is a 42.4% increase in just six months.
"Construction in progress," which represents assets under construction or preparation, has also grown. According to the annual report (Form 10-K) for fiscal 2025, construction in progress rose from $46.636 billion at the end of 2024 to $71.745 billion at the end of 2025. With gross property and equipment at the end of 2025 at $534.098 billion, construction in progress accounted for 13.4% of the total. That is up about 1.6 percentage points from 11.8% at the end of 2024, so the share of pre-operation assets was rising as well.
However, neither construction in progress nor leases not yet commenced directly shows the value of GPUs held in warehouses. Construction in progress includes buildings, equipment and other assets in preparation, and Amazon does not break out the portions attributable to AWS or GPU servers. Leases not yet commenced may also include contracts for buildings and land, so the figure cannot be treated as GPU purchases. Moreover, the year-end 2025 construction-in-progress figure predates the August 2026 rack receipt plan now being reported. The June 2026 10-Q has no note with the same breakdown, so it is also unclear how far construction in progress grew in 2026.
Still, these figures show that Amazon is putting enormous sums into infrastructure that is not yet operating. The widening FCF outflow, the increase in leases not yet commenced and the buildup of construction in progress are all consistent with a time lag between building out facilities and actually monetizing them. How much of the increase is accounted for by GPUs that cannot run because of power shortages cannot be determined from public materials.
Amazon's power constraints surfaced in 2024, with a dispute with a utility in Oregon
The problem of power shortages constraining AWS's expansion of its facilities is not being reported for the first time. In October 2024, Business Insider reported on internal documents saying that power supply, permitting and labor availability were obstacles to business expansion across AWS's Americas regions. In 2025, public records then revealed that Amazon was in a dispute with a utility over power supply to its data centers. Including the August 2026 early rack receipt plan, reports have now indicated for at least two years that the buildout of power and related facilities is failing to keep pace with the expansion of AI infrastructure.
The main developments, in chronological order, are as follows.
| Date | Event | Source |
|---|---|---|
| October 2024 | Business Insider reports internal documents citing constraints on power, permitting, labor and more in AWS's Americas regions | Press report |
| October 2025 | Amazon Data Services files a complaint against PacifiCorp with the Oregon Public Utility Commission over power supply (UM 2410) | Public dispute records and press reports |
| December 19, 2025 | PacifiCorp files its answer to the complaint | Primary source |
| August 2026 | Internal email describes a plan for AWS to receive about $5 billion worth of GB300 racks early | Press report based on internal emails |
| October 2026 | Business Insider reports on early receipt and storage caused by power shortages | Press report |
The information on the 2024 internal documents and 2026 emails is based on Business Insider's reporting, and the original texts have not been made public. For the Oregon power dispute, on the other hand, the answer filed by the utility is available. Amazon Data Services (ADS), which operates Amazon's data centers, filed a complaint with the state utility commission against PacifiCorp over power supply. The issues in dispute include the timing of the power supply needed to develop the data centers and how costs are to be shared.
In the answer PacifiCorp filed on December 19, 2025, four data center sites are named: Specialized, Litespeed, Pivot and Gray. The combined additional power capacity the four sites require is redacted, but the answer says building the necessary transmission facilities would cost billions of dollars. PacifiCorp acknowledged that it was not supplying power to Litespeed at the time Amazon filed its complaint. It has since supplied interim power until full-scale service begins, but says the outlook for formal commencement of operations is later than the target date in the contract the two companies signed in 2023.
PacifiCorp, however, rejects the view that responsibility for the delay lies with it alone. It argues that Amazon's failure to complete the procedures needed for construction and energization on schedule also contributed to the delay. For Specialized, it says it is already supplying enough power to meet Amazon's current demand, and that Amazon itself is not fully using the amount of power available to it under the contract. The dispute thus shows that there are delays in power supply, but it does not establish that the utility is wholly to blame. It involves a complex interplay of facility development plans, contract terms and transmission construction schedules.
Across the U.S., building out power infrastructure is also taking time. According to Lawrence Berkeley National Laboratory's (LBNL) "Queued Up: 2026 Edition", the total capacity of generation and storage projects waiting to connect to the U.S. grid stood at about 2,061 GW at the end of 2025. That is down 10% from the previous year but still large. Of the capacity that applied for interconnection between 2000 and 2020, only 13% had reached commercial operation by the end of 2025, and 75% of applications had been withdrawn.
Note, however, that the LBNL statistics count generation and storage projects, which supply power, waiting for grid connection. They do not directly show the connection queue for facilities that consume power, such as data centers, and Amazon's sites are not part of this queue. Even so, the long time it takes to connect new sources of power to the grid can be a constraint in meeting rising data center demand. Rushing to procure AI servers does not guarantee that the power infrastructure to run them will be built at the same pace.
Microsoft cuts GPU dock-to-live time by about 50%
Behind Amazon's continued heavy capital spending is AI demand that still exceeds supply capacity. On the second-quarter 2026 earnings call, Jassy indicated that even with roughly $220 billion in annual capital spending, the company would not secure enough capacity to meet all demand in 2026. He said the same situation could continue in 2027, and that demand toward 2028 is already large. However, the capacity shortfall he refers to can include multiple constraints such as servers, data centers and power, so not all of it is necessarily due to power shortages.
The same excess of demand persists at rival Microsoft. On the earnings call for the fourth quarter of fiscal 2026, held July 29, CEO Satya Nadella explained that the company is shortening the time between GPUs arriving and going live. The company calls this period "dock-to-live," and said it had cut it by about 50% over the past fiscal year in its largest regions. It also added about 1 GW of data center capacity in the quarter, and said its plan to roughly double total capacity over the next two years is on track. CFO Amy Hood also said customer demand continues to exceed available capacity.
What Microsoft emphasizes is not just buying GPUs, but getting them into a state where customers can use them as soon as possible after arrival. The AWS early receipt reported by Business Insider is a mechanism for securing completed racks before they are delivered to data centers. Microsoft's dock-to-live, by contrast, measures the time from GPUs arriving to going live. The two cover different stages of the process, so Amazon's early receipt and Microsoft's roughly 50% reduction cannot be compared directly.
The improvement Microsoft disclosed is also limited to its largest regions, and the same effect has not necessarily been achieved at all data centers. Figures for other regions, and whether the company has GPUs in storage in warehouses, have not been disclosed. There is therefore not enough information to judge which company runs its GPUs more efficiently. What they share, though, is that shortening not only the volume of GPUs procured but also the time from receipt through installation and energization to going live has become an important management issue for cloud providers.
Doubling power capacity by the end of 2027; the focus is how long early-received GPUs wait
To address power shortages, Amazon is pursuing a plan to double its power capacity from 2025 levels by the end of 2027. On the July 30 earnings call, Jassy said this expansion is proceeding as planned. If realized, it could greatly increase the scale of servers the company can run in its data centers. But when the GB300 racks already received can begin operating depends on when power supply and construction are completed at each site. In addition to power, the readiness of the buildings that house the racks, along with cooling and power distribution equipment, also affects when operations can begin.
The gap between facility plans and actual start of operations has also been observed in Japan. In its "Demand Forecast for the Whole Country and Each Supply Area", published on January 21, 2026, the Organization for Cross-regional Coordination of Transmission Operators, Japan (OCCTO) projected the increase in peak electricity demand from new and expanded data centers and semiconductor plants at 830,000 kW in fiscal 2026, 4.2 million kW in fiscal 2030 and 7.62 million kW in fiscal 2035. In the fiscal 2035 breakdown, data centers alone account for 6.61 million kW. The figures show that rising power demand from the expansion of AI infrastructure is also a long-term challenge in Japan.
Compared with the previous demand forecast, the document explains that start-up dates and the timing of reaching final demand levels were pushed back, mainly because of postponed or delayed data center construction and design changes. This does not indicate that purchased GPUs are sitting in warehouses for lack of power, as in Amazon's case. But it shares the problem that facility construction plans and the actual start of operations do not line up. Even as investment in data centers grows, it does not necessarily show up immediately as an increase in computing capacity or power demand.
Going forward, one thing to watch for Amazon is whether total future payments on leases not yet commenced increase further in the next quarterly report. How the dispute with PacifiCorp over power supply develops will also be a gauge of the state of its data center buildout. What matters is how well the expected GB300 rack receipts reported by Business Insider, about $5 billion in August and roughly $3 billion a month thereafter, mesh with Amazon's plan to double power capacity by the end of 2027.
If power and facility readiness keep pace with the rate of rack receipts, early receipt can be an effective procurement strategy for securing scarce GPUs and putting them into operation right after a data center is completed. If facility preparation lags, on the other hand, the period during which purchased equipment generates no revenue lengthens, and the financial burden may grow. What determines the return on AI infrastructure investment is not only how many GPUs a company secures, but how quickly it can bring them online and turn them into revenue. The report once again shows that in expanding computing resources for AI, the pace at which power and data centers are built out, and not just GPU supply, is an important constraint.