With computer room resources in short supply, are capital players instead starting to be "picky" about data centers?
At the end of September, SB Energy, a SoftBank-backed energy and data center developer, postponed its IPO roadshow. The company, which has signed large-scale data center lease agreements, still needs to respond to additional regulatory inquiries, while investors are reviewing its valuation and its reliance on OpenAI. During the same period, DayOne, which already has operational parks, and Switch continued to prepare for listing.
Also recently, Oracle was reported to have issued a "force majeure" notice to data center developers, hoping to reserve the right to defer payments when projects cannot be put into operation on schedule. Developers are still raising funds for expansion, and cloud vendors as major customers have begun to draw clear boundaries for the delivery risks they undertake.
These changes are taking place against a backdrop of persistently tight data center supply. Data released by CBRE in September shows that in the first half of this year, the vacancy rate in eight major North American data center markets was only 1.4%, and more than 80% of the under-construction capacity has been pre-leased. The expectation of full occupancy has not earned all projects the same financing treatment.
Long-term leases lock in future revenue, but developers must bear several years of construction investment first, and delays in power supply or engineering will further extend the capital advance period. As a result, capital has begun to treat platforms with existing operating revenue differently from projects that rely on long-term contracts from a small number of customers, and customer advance payments and chipmaker guarantees have also been incorporated into financing arrangements. The competition in the data center sector has thus extended to the allocation of construction risks: only those who can secure power supply, organize funds and take delivery responsibilities are qualified to convert computing power orders into server deployments first.
Why Long-Term Contracts From Major Customers Are No Longer Sufficient
SB Energy disclosed in its prospectus updated on September 21 that of the 8.8 GW of signed IT capacity, 0.8 GW is under construction, 8 GW has not yet started, and the operational capacity is zero. The company has operational solar energy and energy storage assets, but its data centers have not yet generated operating rental income. All of this data center capacity belongs to affiliates of SoftBank and OpenAI, of which 8 GW is concentrated in a park in Ohio leased by OpenAI.
To fulfill the contracts, the company estimates that it will need to invest 174 billion U.S. dollars in data center construction in the next six years or so. The 8 GW park is under a 20-year long-term lease, but the rent will be charged in phases by building, on the premise that power supply facilities are available, the building is completed and commissioned. A small number of major customers taking up the entire capacity can reduce the uncertainty of investment promotion, but cannot eliminate the rent-free period from the start of construction to delivery.
Construction payments and interest need to be paid upfront. If the first phase of rent collection is delayed for a long time, subsequent construction will still rely on external financing, and interest expenses will take up the budget for the next phase. Delivery problems of one project may slow down the expansion of the entire park as a result.
According to the listing preparation news disclosed on September 25, DayOne plans to publicly submit its U.S. listing application in mid-October and strive to go public in November; Switch is expected to issue shares later, and the schedule is yet to be officially announced. DayOne already has operational parks in Malaysia and has a presence in multiple markets across Asia Pacific and Europe, so its financing faces a combination of existing businesses and new projects.
Operational assets can provide rental income and operating track records, so investors do not need to pin all their judgments on whether a large park can be completed on schedule. Ke Yan, head of research at Shenton Research, pointed out that investors pay more attention to contracted capacity with secured power supply that is already in operation or close to completion, as well as long-term payment contracts.
Platforms in the construction phase mainly rely on future revenue to support their valuations. When financing costs rise, the same amount of future rent can support less upfront investment. If new funds have to be continuously borrowed to pay for construction costs, the financing terms will in turn determine the scale of the project that can be launched.
On September 30, Accelevation, which provides power distribution, cooling systems and installation services for data centers, went public on NASDAQ at an issue price of 18 U.S. dollars, lower than the previous roadshow range of 20 to 24 U.S. dollars. The IPO window has not closed yet, but enterprises have had to make concessions on pricing. For developers that continue to rely on external funds for expansion, if similar pressures persist, it may mean lower valuations, supplementary guarantees, or phased construction.
Self-Built Power Supplies Also Need to Pass Approval Procedures
The recipient of Oracle's notice is the developer of the Project Jupiter park in New Mexico, including enterprises under Blue Owl. The park is scheduled to be put into operation in 2028, and its supporting power supply facilities are encountering resistance in the approval process. Oracle emphasized that the project is still advancing as planned, and Blue Owl also stated that its existing financial commitments remain unchanged, and the payment liability in case of potential delays has not yet reached a final conclusion.
According to the plan, the park will use Bloom Energy's natural gas fuel cells to generate electricity through electrochemical reactions, reducing reliance on new access capacity of the public grid. The power supply still requires a stable natural gas supply, and the supporting pipeline, which was originally planned to be put into use in September this year, has been postponed to February 2027 due to blocked line permits.
Oracle clarified on September 14 that the data center and the power supply microgrid are adjacent but independent facilities operated by different companies. The judicial procedure involves the air emission permit of the microgrid, not the building permit of the data center. The data center can continue construction, but its power connection depends on the progress of another set of approval procedures.
If tenants can defer payments, the capital pressure during the construction period will fall more on the developers. Lenders therefore review not only the lease term and amount, but also the lease commencement conditions, delay liabilities, and whether the guarantees can cover these risks.
On September 21, the Governor of Texas requested the Texas Commission on Environmental Quality to suspend the issuance of data center-related permits pending grid and water resource reviews. Even if developers find customers and power supplies, new projects still need to meet local power supply reliability and water resource requirements, and sufficient budget cannot directly secure approval.
DayOne is also looking for additional power supplies. The company announced in September that it will cooperate with the power generation subsidiary of Tenaga Nasional Berhad in Malaysia to explore the construction of on-site power generation facilities with a maximum capacity of 1.5 GW in Selangor, equipped with energy storage systems. The plan is still subject to feasibility studies and approval, and existing operational assets do not mean that subsequent expansions will be free from power constraints.
In September, Morgan Stanley raised its forecast for new power demand from U.S. data centers from 2026 to 2028 to 97 GW. After accounting for under-construction capacity, grid supply, on-site power generation and other measures to accelerate power supply, it is estimated that there will still be a cumulative gap of about 33 GW by 2028, and alternative power supply solutions cannot cover all construction needs.
GE Vernova, a power equipment manufacturer, disclosed on September 16 that it has sold multiple gas turbine production slots for delivery in 2032. The company also mentioned that some projects hope to connect to the grid five to seven years later, but there are no definite commitments yet. On-site power generation may last for many years, and equipment delivery lead times and fuel security have also become long-term constraints.
On October 1, Japanese power generation enterprise JERA, Dell, and AI infrastructure developer RHAELM announced the signing of a memorandum of understanding for cooperation, planning to use the existing power plant in Chiba to deploy AI infrastructure with a maximum power supply capacity of 400 MW, targeting operation around 2028. The project plans to adopt behind-the-meter power supply, relying on existing power generation facilities to reduce the waiting time for new grid connection. Dell provides standardized rack solutions, and RHAELM is responsible for development and delivery.
JERA provides the power plant and site, and Dell participates in the computing equipment solution before the data center is completed. The two sides hope to reduce the time for separate design and docking of power, cooling and servers. Locations with better power supply conditions have the opportunity to attract customers and equipment partners in advance, while parks that are still waiting for grid connection or gas turbine delivery need to bear a longer capital occupation period.
Customers and Chipmakers Share Capital Pressure
Oracle disclosed that as of September, 75% of the total capacity of its Abilene data center in Texas has been delivered. While operational parks and to-be-built projects are advancing in parallel, cloud vendors are also requiring customers to share upfront investment.
Oracle's financial results announced on September 10 show that its cash capital expenditure in the latest fiscal quarter was about 28.5 billion U.S. dollars, while it received about 11.4 billion U.S. dollars in customer advance payments with significant financing components. The management stated that most of the new large AI contracts adopt arrangements where customers pay in advance or bring their own hardware, and these new contracts will not affect capital expenditure and revenue until fiscal 2028 or later.
Advance payments reduce Oracle's capital advance when purchasing and deploying equipment, while bringing one's own hardware transfers part of the GPU procurement responsibility to customers. Therefore, changes in the cloud vendor's own capital expenditure cannot be directly equated to fewer chips purchased for the entire project. Customers still need to raise funds for advance payments and hardware, and their financial strength will continue to affect the performance of contracts.
GPU cloud service provider CoreWeave disclosed on September 17 that its asset financing is supported by long-term "take-or-pay" customer contracts, GPUs and other equipment, data center leases and other assets, with funds drawn in installments as costs are incurred and infrastructure is deployed, and then repaid with contract proceeds. The company plans to complete the repayment before the contracts expire. Customer credit and equipment value together determine the financing base. For the same number of GPUs, different tenants and contracts behind them may lead to different capital conditions.
CoreWeave divides data center development and GPU installation and testing into two phases. The first phase includes buildings, grid connections and electromechanical cooling facilities. Chip procurement needs to match data center delivery: arriving too early increases capital occupation, while arriving too late affects service contracts, and phased financing needs to align with equipment deployment.
According to SB Energy's prospectus, NVIDIA has provided residual value guarantees for the initial approximately 4.25 GW of IT capacity at its Ohio park. In case of agreed circumstances such as tenant bankruptcy or payment default without remedy, NVIDIA can take over the lease, or require re-leasing, and after the re-leasing attempt fails, it can require the sale of relevant assets. The guarantee has a cap, and compensation must meet delivery conditions, so it cannot cover all construction risks.
SoftBank and OpenAI are both investors and tenants of SB Energy, and NVIDIA provides credit support. The overlapping roles of these enterprises are conducive to organizing funds and orders, but also make their risks interrelated. The prospectus warns that the financial deterioration of related parties may simultaneously affect tenant payments, guarantee performance and cooperative relations. As a result, chipmakers assume responsibilities beyond sales, and multiple purchasing entities do not necessarily represent mutually independent demands.
On September 30, AIB, a data center developer and operator, announced that it has signed a binding agreement with AI cloud service provider Nebius to provide 50 MW of IT capacity. AIB previously signed a 15-year power supply agreement for this project in the southeastern United States, corresponding to a grid load of 65 MW, and stated that no large-scale new power infrastructure upgrade is required. Nebius also listed power supply conditions as an important reason for site selection.
AIB expects that customer advance payments under the initial 12-year contract, together with project loans and preferred stock financing, will support most of the initial development costs, reducing the need for common stock financing. The power supply agreement helped the company secure long-term contracts, which are then used to organize construction funds. Power supply affects customer choices first, and then the financing structure, and it is not just an operating cost after the project is put into operation.
For GPU, HBM and network chip suppliers, projects with secured power supply and funds can arrange procurement around the delivery date; parks that still rely on subsequent financing are more likely to adjust batches or postpone deployment. Even if customer demand remains unchanged, equipment orders originally planned for the same period may be scattered for delivery in different quarters.
Conclusion
The data center shortage creates demand for expansion, but does not eliminate the capital and delivery risks during the construction period. The gaps widened by power supply conditions, customer credit and financing capabilities will eventually be reflected in who can deliver computing power first and who can continuously purchase chips.
This article is from the WeChat Official Account "Semiconductor Industry Panorama" (ID: ICViews), written by Jun Xi, republished with authorization from 36Kr.