Oracle delivered 850 megawatts and more than 300,000 GPUs in a single quarter

Oracle reported first quarter fiscal 2027 results on September 10, and the capacity figures are the part worth reading. The company delivered more than 300,000 GPUs and 850 megawatts of capacity to customers during the quarter. Co-chief executive Clay Magouyrk said that is "almost three times what we delivered in all of Q4, and 73 percent of the total capacity delivered last fiscal year." Cloud infrastructure revenue reached 7.4 billion dollars, up 121 percent year on year, accelerating from 93 percent growth the previous quarter. Total revenue was 19.3 billion dollars, up 30 percent. Remaining performance obligations, which is the contracted backlog not yet delivered, stood at 664 billion dollars, up 209 billion year on year and up 26 billion in the quarter alone. Chief financial officer Hilary Maxson said the vast majority of new contracts are structured as prepay or bring your own hardware, and so will not require incremental Oracle capital. Capital expenditure guidance for the fiscal year is 90 to 95 billion dollars, which sits below Amazon at roughly 220 billion and Microsoft at 175 billion. Utilization ran at 97.9 percent and renewing GPUs are being resold at a 20 percent premium. Net income was 4.7 billion dollars and free cash flow was negative 5 billion. Chief executive Mike Sicilia also teased a new agentic AI accelerator to be unveiled at AI World in October, which is notable given Oracle has repeatedly denied custom silicon plans. Shares rose about 6 percent after hours. The number to carry into customer conversations is 850 megawatts in ninety days. Delivery at that rate is a supply chain and commissioning achievement before it is a compute achievement, and the constraint sits with whoever cannot keep up.

Source: Data Center Dynamics (September 11, 2026)

DOE closes a loan of up to $1.9 billion to restart the Duane Arnold nuclear plant for Google

The Department of Energy's Office of Energy Dominance Financing closed a loan of up to 1.9 billion dollars on September 8 to support restarting the 615 megawatt Duane Arnold Energy Center in Linn County, Iowa. The plant entered service in 1975 and shut down in 2020 after damage from that August's derecho. NextEra Energy signed a 25 year power purchase agreement with Google for the output in October 2025, and the restart is targeted for the first quarter of 2029, subject to Nuclear Regulatory Commission approval. NextEra describes thousands of American jobs during construction and refurbishment and more than 400 permanent positions once the plant is running. This is the third federal loan supporting a nuclear restart, following 1 billion dollars for Three Mile Island, now the Crane Clean Energy Center, and 1.52 billion for Palisades. Read it as a template rather than a one off. A hyperscaler underwrites the offtake on a multi decade contract, federal credit support closes the financing gap, and a unit that the market had written off comes back with its entire output already sold. There are a limited number of restartable units left in the United States, which is precisely why the ones that remain are being contracted years before they produce a megawatt hour. For the field, the practical scope is a full plant refurbishment on a 2029 clock: instrumentation and controls, switchyard and transmission work, cooling systems, security and a very long commissioning tail, all under NRC documentation standards rather than commercial construction standards. That is a different qualification bar than a data hall fit out, and the teams that hold those qualifications are about to be in short supply.

Sources: NextEra Energy (September 8, 2026) and Data Center Dynamics (September 9, 2026)

SpaceX signs a fourth compute contract, taking annual leasing income to $41.1 billion

Chief financial officer Bret Johnson told the Goldman Sachs Communacopia conference that SpaceX has signed a new hosting contract worth 13.3 billion dollars a year, or 1.11 billion a month, with an AI customer it did not name, DCD reported on September 11. Revenue begins in December 2026. The deal is the company's fourth compute contract of the year and lifts total annual income from compute leasing to 41.1 billion dollars. Existing customers and their monthly run rates were disclosed as Anthropic at 1.25 billion, Google Cloud at 920 million and Reflection AI at 150 million. SpaceX had roughly 1.4 gigawatts live as of June 30, is targeting 2 gigawatts by the end of this year and close to 10 gigawatts by the end of 2027, anchored on the Colossus campuses in Memphis, Tennessee and a site in Southaven, Mississippi. The company's market capitalisation sits near 2 trillion dollars following its June listing. Set aside the size of the numbers and look at the structure. A company that was not in the data center business two years ago is now one of the larger merchant landlords of AI capacity in the United States, and it is growing that book faster than it is growing megawatts. The gap between 41.1 billion dollars of contracted annual revenue and 1.4 gigawatts of live capacity has to be closed by construction, in the mid South, on a schedule set by contracts that are already signed. That is a demand signal for anyone selling into Tennessee and Mississippi.

Source: Data Center Dynamics (September 11, 2026)

Federal money moves further into private AI infrastructure

Two items on September 11 point the same direction. The Defense Information Systems Agency opened bidding on the Joint Warfighting Cloud Capability Unified Cloud Marketplace, with a ceiling of 21.6 billion dollars and bids due October 6. The previous JWCC award, made in December 2022, carried a 9 billion dollar ceiling, so the government has more than doubled the envelope in under four years. Separately, and still unconfirmed, the Pentagon is reported to be in talks to lend Fluidstack 5 billion dollars through its Office of Strategic Capital, with the money earmarked for domestic manufacturing of data center components. The report originated with the Wall Street Journal and no agreement has been announced, so treat it as a live discussion rather than a signed deal. Taken together with the Duane Arnold loan, the pattern is a federal balance sheet extending into generation, components and cloud capacity at the same time. Two implications for vendors. Federal money brings federal terms, which means domestic content requirements, cost accounting standards, security clearances and audit obligations that most commercial data center suppliers have never had to satisfy. Firms that already hold those qualifications are about to find them commercially valuable in a market that has been almost entirely private. And a 21.6 billion dollar marketplace with an October deadline will pull integrator and cloud provider attention through the fourth quarter, which is worth knowing if you are trying to get decisions out of those same accounts before year end.

Sources: Data Center Dynamics (September 11, 2026) and Data Center Dynamics (September 11, 2026)

Aligned breaks ground on a 2 GW campus on a retired coal site in western Pennsylvania

Aligned Data Centers broke ground on September 10 on Project Phoenix, a three building, two gigawatt campus at the Shippingport Industrial Park in Beaver County, Pennsylvania. The campus sits alongside the former Shippingport power station, a retired coal plant now being redeveloped into a gas energy campus, in a borough that Mayor John Erickson describes as the birthplace of commercial nuclear energy. Aligned describes the campus as anchoring roughly 10 billion dollars of regional investment, supporting about 3,000 construction jobs and around 640 full time roles. Holder Construction and Kokosing are the construction partners. Three details deserve attention. The campus will be served by a dedicated on site natural gas energy campus rather than by grid supply, and Aligned says it expects to return surplus capacity to the country's largest grid operator rather than draw from it. Cooling is closed loop deionized with zero water consumption. And Aligned has come out in support of Pennsylvania's GRID standards and is a signatory of the state's Ratepayer Protection Pledge, which is a deliberate answer to the political argument that has stalled projects elsewhere in the Northeast and Mid Atlantic. This is the clearest example yet of the model that defined the week. A legacy energy site brings existing transmission, cooling water rights, industrial zoning and a community already familiar with heavy generation, and the developer supplies its own power rather than joining an interconnection queue. Expect more of these. The scope implications are substantial, because a project like this carries a full power plant package alongside the data halls, including turbines, fuel handling, emissions control, electrical balance of plant and the commissioning regime that goes with all of it.

Source: GlobeNewswire (September 10, 2026)

Kearney, Nebraska approves a 368 MW campus and writes a data center ordinance the same night

The Kearney City Council voted unanimously on September 8 to approve the conditional use permit and redevelopment plan for Project Horizon, a phased data center of roughly 368 megawatts on the northeast side of the city. Local coverage names SDV and MARA Holdings as the developers, redeveloping a site that operates today as a bitcoin mine. The eventual tenant has not been disclosed, and city officials have said they do not know who it will be, so treat any characterisation of the workload as speculation. Cooling is closed loop, with water use projected at two to five million gallons a year. Construction is targeted for November or December of this year, with the first phase operational in spring 2028 and full build out around 2030. No investment figure has been disclosed. What makes this worth reading is what the council did alongside the approval. Kearney adopted one of Nebraska's first dedicated data center zoning ordinances, covering noise, water and energy, and then held Project Horizon to it. Conditions include sound walls that are expected to leave the site quieter than the crypto mine operating there today. The story runs against the regional trend, because a number of Nebraska communities have been pulling back from data center development this year. Kearney went the other way, and did it by writing the rules first and approving the project second. That sequencing is worth noting if you are advising clients on entitlement strategy. A community that has a standard can approve against it. A community that has no standard tends to litigate every project individually, which is how eight hour hearings happen.

Source: Nebraska TV (September 9, 2026)

Fleet asks Nevada regulators to let it build more than 360 MW of its own methane generation

Fleet Data Centers, launched by Tract Capital Management in January 2025, has asked the Public Utilities Commission of Nevada for approval to build two methane gas fired generating plants serving its Peru Ridge and South Valley campuses at the Tahoe Reno Industrial Center in Storey County, about thirty miles from Reno, DCD reported on September 8. Combined capacity would be 360 megawatts. Fleet has described the generation as a bridge of two to three years until NV Energy can deliver permanent utility service. A decision was expected within days of that report and we have not been able to confirm an outcome, so treat the application as pending. If approved, it would be the first methane gas supply cleared specifically for data center use in Nevada. Opponents argue the structure effectively routes around the state's renewable portfolio standard, which requires 50 percent renewable generation by 2030. Fleet's position is that it is not subject to the standard because it is not an energy provider. That argument is going to be made in a lot of states over the next eighteen months, and how Nevada resolves it will be cited elsewhere. For vendors, watch the outcome rather than the filing. A clean approval here makes behind the meter gas a default option across the Mountain West and pulls forward demand for turbines, fuel systems, emissions control and the electrical infrastructure that ties private generation to a campus.

Source: Data Center Dynamics (September 8, 2026)

Oracle goes to market for 2 GW of renewables in New Mexico after the state blocked a gas pipeline

Oracle issued a request for proposals on September 8 for two gigawatts of new solar, wind, geothermal and other renewable capacity in New Mexico, with delivery between 2027 and 2031, and named Mahesh Thiagarajan, executive vice president of Oracle Cloud Infrastructure, on the announcement. Oracle is targeting 100 percent carbon free matching by 2031. Per DCD's coverage the following day, the power is intended for Project Jupiter, the 2.5 gigawatt campus in Doña Ana County being developed across roughly 1,400 acres and four buildings with Stack and BorderPlex Digital Assets. DCD also reports Oracle committing up to 1 million dollars to research carbon capture on the Bloom Energy fuel cells serving the site, under a solid oxide fuel cell deal expanded from 1.8 to 2.8 gigawatts in April that replaced planned gas turbines and diesel generators. The context matters more than the megawatts. New Mexico regulators rejected a gas pipeline extension to the campus in July, and this RFP is the answer to that decision. Fuel cells and a two gigawatt renewable procurement are what a 2.5 gigawatt campus looks like when the pipeline does not get built. Two gigawatts of new generation solicited by an end user rather than a utility is a large procurement by any standard, and it lands in a state with good solar and wind resource but limited transmission. Developers responding to this will need firming, storage and interconnection work, and the campus itself needs the infrastructure to accept generation from multiple sources on different schedules.

Sources: Oracle (September 8, 2026) and Data Center Dynamics (September 9, 2026)

SoftBank-backed SE Cosmos puts about $1 billion into a 50 MW AI hardware lab at the old 3M Austin campus

SE Cosmos, indirectly owned by SoftBank backed SB Energy, is converting the former 3M campus in Austin, Texas into the Cosmos Technology Campus, an AI hardware research and development facility with 50 megawatts of critical IT load, DCD reported on September 9 based on disclosures in SB Energy's SEC filing. The campus was built in the 1980s and sold by 3M in 2018. Roughly 1 billion dollars in financing has been secured. The anchor tenant is Silver Bands 3 (US) Corp., another SoftBank affiliate, on a fifteen year lease expected to generate about 2.5 billion dollars in rent, commencing in stages across 2026 and 2027. Austin Energy will upgrade its existing substation to serve the campus. Per an S&P Global ratings note, the build comprises rack development halls, silicon development benches and high density powered lab halls. This is not a colocation facility and should not be modelled as one. Hardware development space carries denser and more varied power delivery than a production data hall, more instrumentation, more frequent reconfiguration and tighter environmental control on the bench side. The wider context is that SB Energy filed for a 5 to 7 billion dollar IPO the week before at a valuation near 50 billion, Nvidia committed 1.5 billion in August to its Pike County, Ohio campus, and SB Energy is building OpenAI's 1.2 gigawatt Stargate site in Milam County, Texas. A brownfield conversion of an existing industrial building is also a reminder that not every megawatt added this cycle arrives as greenfield construction.

Source: Data Center Dynamics (September 9, 2026)

Alpha Compute finances 85 percent of a Pennsylvania land purchase with a seller note

Alpha Compute Corp., listed on Nasdaq as ALP, announced on September 10 that it has secured 47 million dollars in seller financing toward a 55 million dollar land purchase in Pennsylvania for a planned AI data center campus. The campus parcels are roughly 155 and 88 acres in Tioga County in the north of the state, alongside about 107 acres in Beaver County in western Pennsylvania held for future development, carrying around 350 surface and pore space acres in aggregate along with approximately 1,800 net unleased Marcellus mineral acres at full net revenue interest. The company will pay 8 million dollars in cash at closing, made up of a 3 million dollar deposit already credited and 5 million more, with the remaining 47 million carried by the sellers as a non recourse note at 6.00 percent fixed, interest only, with a five year balloon. Principal is to be repaid in part from half of gas production proceeds, and the data center, power and compute assets are carved out of the collateral. The plan calls for 200 megawatts initially, expandable to a gigawatt, powered behind the meter by gas at an estimated 5.85 cents per kilowatt hour against PJM commercial and industrial rates of roughly 8 to 10 cents. The caveats belong in the same breath as the plan, and to its credit the company states them plainly. This is greenfield land with no power and no data center capacity on it today, the 200 megawatts is planned rather than permitted, and neither closing nor county approval is assured. The reason to track it is the financing structure. Sellers who hold mineral rights are effectively taking equity like exposure to a data center thesis, and a developer with no operating asset is acquiring a gigawatt of optionality for 8 million dollars of cash. Expect imitators, and expect a share of them not to reach construction.

Source: GlobeNewswire (September 10, 2026)

Designing for GPUs per megawatt: NVIDIA's DSX MaxLPS and the case for warmer water

NVIDIA has been extending DSX from an AI factory reference design into a broader infrastructure platform, and Data Center Frontier published a useful synthesis of where it now sits on September 10. The pieces include reference designs, a simulation environment called DSX Sim, a partner marketplace in DSX Exchange, a grid interactive workload control layer called DSX Flex, an Omniverse digital twin blueprint, and the element most relevant to anyone specifying mechanical systems, DSX MaxLPS.

MaxLPS pairs GPU level power optimisation with 45 degree Celsius liquid cooling. NVIDIA's claim, modelled on a hypothetical 100 megawatt AI factory, is that the combination allows up to 40 percent more Rubin GPUs inside the same fixed power envelope. The mechanism is straightforward once stated. Warmer supply water means less energy spent on heat rejection, and in a facility where the power feed is fixed and fully subscribed, every kilowatt not spent on cooling is a kilowatt available for compute. Partner reference designs from Trane and Eaton claim up to 15 percent better energy efficiency, up to 30 percent lower installation cost and as much as 80 percent less copper than conventional low voltage designs. Those are vendor projections and should be treated as such until someone publishes measured results.

The complementary insight came from CoolIT the same day. Chief technology officer Kamal Mostafavi argued the industry has been optimising against the wrong number. Thermal design power is effectively a solved problem, and CoolIT has built a coldplate capable of cooling a hypothetical 15 kilowatt chip, well beyond its commercial hardware and not intended for volume sale, as a deliberate demonstration of headroom. The variable that actually binds is heat flux measured in watts per square centimetre, and Mostafavi puts NVIDIA's Rubin generation at 160 to 170, held there by package warpage, material selection and yield limits. Whether Rubin Ultra lands nearer 400 to 500 watts remains an open question.

Why it matters for North American builds. Put those two together and you get a defensible answer to the question every operator commissioning capacity this year is asking privately, which is whether the facility has to be gutted in three years. If heat flux is flattening while package area grows, single phase direct to chip has considerably more runway than the assumption that immersion is inevitable, and warmer water is the lever that converts that runway into deliverable capacity. That matters most precisely where this week's builds are landing. Aligned is self generating in Pennsylvania, Fleet is asking to self generate in Nevada, and Oracle is buying two gigawatts of renewables in New Mexico because a pipeline was refused. When the power is that hard to obtain, and NVIDIA notes more than 200 gigawatts sitting in United States interconnection queues, the return on a mechanical design that frees up 10 or 15 percent of the power budget is measured in months of schedule rather than in operating cost. Design for output per available megawatt, and specify accordingly.

Sources: Data Center Frontier (September 10, 2026) and Data Center Dynamics (September 10, 2026)