Flex acquires 800 volt power conversion manufacturer EPC Power in a reported $4.4 billion deal

Flex has agreed to acquire EPC Power, a North American designer and manufacturer of power conversion equipment for data centers, energy storage and microgrids, for a reported 4.4 billion dollars, DCD reported on September 4. The transaction is subject to regulatory approvals and is expected to close sometime in the fourth quarter of this year. Once complete, EPC will operate as a business within Flex's Cloud and Power Infrastructure segment, building on an existing collaboration between the two companies. EPC develops 800 volt data center power architectures including digital rectifiers and solid-state transformers, aimed at higher-density AI infrastructure, and Flex said the acquisition extends its power and thermal management portfolio into what it calls an integrated grid-to-chip offering. Founded in 2021, EPC Power has been backed by Goldman Sachs Alternatives and Cleanhill Partners as controlling shareholders. "What we accomplished over the last four years demonstrates the power of strong partnerships and a shared commitment to innovation," said Jim Fusaro, CEO of EPC, adding that the company "expanded our domestic manufacturing footprint nearly tenfold." Flex, headquartered in Singapore, is a contract manufacturer serving customers across a range of industries including data center power and cooling infrastructure, and it has separately announced plans to spin off the Cloud and Power Infrastructure portfolio under which EPC will sit. Goldman Sachs and J.P. Morgan Securities advised EPC Power and its shareholders, with Vinson & Elkins acting as legal counsel. The practical read for vendors is that the 800 volt architecture now has an industrial owner with global capacity behind it, and that domestic power conversion manufacturing has become a strategic asset rather than a commodity supply line.

Source: Data Center Dynamics (September 4, 2026)

Broadcom posts $29.6 billion quarter and says AI networking revenue will double twice by 2028

Broadcom reported fiscal third quarter revenue of 29.6 billion dollars, up 86 percent year on year, with Semiconductor Solutions at 20.8 billion dollars, up 127 percent, DCD reported on September 4. GAAP net income rose more than 200 percent to roughly 13.1 billion dollars. Infrastructure Software, led by VMware, came in around 8.8 billion dollars, up 29 percent, though its share of group revenue fell from 43 percent to 30 percent. Non-AI semiconductor revenue was 4.2 billion dollars, up 5 percent and flat sequentially. The forward guidance is the part worth reading closely. CEO Hock Tan said AI networking revenue will double in fiscal 2027 to 115 billion dollars and double again in fiscal 2028 to 230 billion dollars. "We have secured the supply to meet this outlook," Tan said, adding that the company is "very much on target to exceed 30 dollars in earnings per share in fiscal 2028." Pressed on whether the demand is real, Tan pointed at physical infrastructure: "this is real demand we believe based on what's available, what data center sites, locations are ready [in] 2028." The named commitments behind that are large. Anthropic is tied to an upcoming 1 gigawatt Ironwood deployment plus a further 5 gigawatts of TPU version 8i in fiscal 2027, with Tan describing "clear line of sight to deliver another incremental 10 gigawatts," and he expects Anthropic to become Broadcom's largest XPU customer in 2027. OpenAI's data movement optimized chip, internally called Jalapeño, is tracking to a 1.3 gigawatt deployment in 2027. On the networking side, the 100 terabit Tomahawk 6 is shipping and a 200 terabit Tomahawk 7 has recently taped out. When a supplier says it has already locked wafers, substrates and high bandwidth memory against 2028 site readiness, that is a schedule signal for every trade downstream of the chip.

Source: Data Center Dynamics (September 4, 2026)

Microsoft discloses Azure revenue for the first time and restructures how it reports

Microsoft has adjusted its financial reporting practices and disclosed Azure-specific revenue for the first time, in an investor presentation filed with the SEC, DCD reported on September 4. Azure brought in 29.42 billion dollars in the most recent quarter, up from 20.7 billion dollars a year earlier, and 101.94 billion dollars across full fiscal 2026, up from 72.6 billion dollars. Microsoft's original late-July earnings had reported only the broader Intelligent Cloud division, which contained Azure, at 39.3 billion dollars. For comparison in the same quarter, AWS reported 42.2 billion dollars and Google Cloud 24.8 billion dollars. Going forward Microsoft will report under two overarching segments, Agents and Infra, covering apps, agents, Microsoft 365 and Azure, and Devices and Consumer, covering search, advertising, Xbox and Windows. The company said investors will have full transparency of quarterly revenue across each of its key businesses from 2027 onwards. The disclosure lands alongside capital spending numbers that matter more directly to the field. Amazon's annual capital expenditure has reached 220 billion dollars, driven by AI and memory costs, and Google has raised its 2026 capital expenditure guidance to between 195 and 205 billion dollars for its AI data center build-out. Better segment reporting is not a small thing for anyone selling into these programs. Quarterly visibility into which business line is growing, and how fast, is the cleanest public proxy available for where the next tranche of capacity gets committed and when.

Source: Data Center Dynamics (September 4, 2026)

Santee Cooper turns to Google Cloud for AI load forecasting as it plans a $10 billion grid expansion

South Carolina's state-owned electric and water utility Santee Cooper has partnered with Google Cloud to deploy Gemini Enterprise and WeatherNext 2 for load forecasting and financial planning, DCD reported on September 4. WeatherNext 2 generates hundreds of ensemble scenarios in under a minute and produces probabilistic forecasts up to 15 days out, which the utility says reduces the risk of over-buying or under-buying power. The cost of getting that wrong is specific. "If we underestimate our power needs, we are forced to buy electricity on the spot market, often at high prices," said Tami Wilson, vice president and CFO of Santee Cooper. "On the coldest winter day, one degree of temperature variance can cost up to 100,000 dollars per hour on the spot market." A financial forecasting model built on Gemini Enterprise is claimed to cut scenario-building time by up to 75 percent, moving work that took weeks or months into days or hours. Those capabilities feed planning around the utility's 10 billion dollar grid expansion budget. "Utilities are drowning in real-time data from weather feeds to asset data, and legacy systems weren't built to process it at today's speed," said Raiford Smith, global director of power and energy at Google Cloud. Santee Cooper is the state's largest power provider, serving around two million people, and last year it enacted a special electricity rate for data centers and other large loads requiring 50 megawatts or more, running four years experimentally before a permanence decision. Two things are worth noting. Utilities are becoming AI customers as well as AI suppliers, and the 50 megawatt large-load threshold in South Carolina is another data point in a national pattern that now includes Ohio, North Carolina, Virginia and, pending the governor's signature, California.

Source: Data Center Dynamics (September 4, 2026)

Meta's $1.2 billion data center in Kuna, Idaho is now serving traffic

Meta's data center in Kuna, Idaho is officially online, DCD reported on September 4. The company first announced plans for the Kuna campus in Ada County in early 2022, originally aiming to invest 800 million dollars in a 960,000 sq ft facility. The site was one of around a dozen paused as part of a design re-scoping to accommodate AI infrastructure and liquid cooling, and it comes back at a larger number. "Meta's Kuna Data Center is officially serving traffic, our second AI-optimized facility in our fleet, now online in Idaho, representing an overall investment of 1.2 billion dollars," wrote Gary Demasi, Meta's vice president of data center development and strategy. The facility uses a closed-loop liquid cooled system that Demasi said "uses zero water for most of the year," and Meta has invested more than 100 million dollars in a dedicated water and wastewater system for the City of Kuna. The company has matched the site's electricity use with 100 percent clean and renewable energy, having added 645 megawatts of new renewable capacity to Ada County. Meta's first AI-optimized data center launched in July of this year in Temple, Texas, another site paused in 2022. Kuna sits in the Boise metropolitan area, a relatively small data center market where Diode Ventures filed for a 620-acre project in February 2025. The pattern to track is the redesign premium. A facility scoped at 800 million dollars for conventional load came back online at 1.2 billion after being re-engineered for liquid cooling and AI density. That 50 percent delta is the clearest public estimate available of what the AI retrofit actually costs on a shell that was already designed and permitted.

Source: Data Center Dynamics (September 4, 2026)

Bitdeer pays $100 million for 200 acres in Milam County, Texas, next to its Rockdale campus

Bitdeer Technologies Group has paid 100 million dollars for a 200-acre parcel in Milam County, Texas, carved out of the former Alcoa aluminum smelting plant and adjacent to its existing Rockdale data center, DCD reported on September 4. Across the two sites Bitdeer now holds 563 megawatts of interconnected grid capacity in Texas, expandable to 742 megawatts, against a company-wide power pipeline it puts at 3 gigawatts. The reasoning behind buying rather than leasing was stated directly. "Securing outright ownership of land in Milam County reflects the disciplined approach we're taking to expanding powered land under our control," said Michael G. Potter, Bitdeer's CFO. "Owning this land gives us greater long-term certainty over our development plans, eliminates the renewal risk associated with a lease, and provides flexibility to maintain our existing Bitcoin mining operations while we develop the newly acquired property." Bitdeer manages 175,000 Bitcoin mining machines and runs its AI neocloud through subsidiary Bitdeer AI, with operating sites including 570 megawatts in Ohio, 37 megawatts in Tennessee, 13 megawatts in Washington State and 175 megawatts at Tydal in Norway. The Washington and Tennessee sites are converting to AI data centers. Context matters here. The Alcoa site spans roughly 33,000 acres, smelting ended in 2008 and the plant fully closed in 2017, and SoftBank was reported in June to have bought 4,000 acres there, with its SB Energy unit building a Milam County data center to be leased by OpenAI. Texas is also currently operating under a pause on new data center developments after Governor Greg Abbott ordered the state utility provider to audit its interconnection queue. Milam County is quietly becoming one of the densest concentrations of powered land in the country, and land control there is now being bought at a premium.

Source: Data Center Dynamics (September 4, 2026)

T5 sells its operations business to Salute and spins its construction arm out as EverOn

T5 Data Centers has split its services business in two, DCD reported on September 4. Salute has entered a definitive agreement to acquire T5 Operations, with terms undisclosed. On closing, the acquisition will expand Salute's operational footprint to more than 15 gigawatts of data center capacity under management across roughly 150 markets globally, and take its worldwide headcount above 3,800 employees. "This latest deal is a significant step in our growth journey, one that's fueled by demand for AI/HPC facilities across the world," said Erich Sanchack, CEO of Salute. David Mettler, EVP of T5 Operations, joins the Salute executive leadership team, and Pete Marin, CEO of T5 Data Centers, joins the Salute board. Founded in 2013, Salute provides facility operations and management, quality assurance, technical commissioning and retrofitting services, and has been majority owned by New Mountain Capital since 2024, having acquired Keysource, Advanced Data Center Consulting Group and Northshore since. At the same time, T5 Construction has rebranded as EverOn Data Center Services, an independent data center general contractor. The growth numbers behind that spin-out are striking: revenue rose from 87 million dollars in 2021 to 1.7 billion dollars in 2025, with 83 percent coming from third-party construction projects, across more than 260 completed projects and over 12 million sq ft of delivered data center space in the United States. "While our name is changing, what customers rely on is not," said Tom Mertz, now president and CEO of EverOn. Two structural changes for the field in one announcement. The operations side of the market is consolidating into a small number of very large platforms, and one of the country's larger data center general contractors is now independent and free to bid work it previously could not.

Source: Data Center Dynamics (September 4, 2026)

Jane Street reportedly signs a $13 billion, five-year cloud contract with Crusoe

Quantitative trading firm Jane Street has reportedly signed a 13 billion dollar cloud computing contract with Crusoe, giving it GPU access through Crusoe's cloud platform, DCD reported on September 4. Bloomberg reported the deal on September 3, citing people with knowledge of the matter, and put the term at five years. Crusoe declined to comment and the hosting location was not stated. Crusoe has data centers operating or under development in Texas and Wyoming and has been linked to projects in Missouri and Virginia, and it also leases third-party capacity. Its website claims 13.6 million sq ft and 4.9 gigawatts, though how much of that is operational is unclear. This is Crusoe's largest cloud contract to date, and it is a different kind of deal from its prior work with Oracle, Microsoft, which took a 700 megawatt lease at Abilene, Texas, and Meta, all of which were for physical data center space rather than platform consumption. Bloomberg had previously reported Crusoe seeking a 3 billion dollar funding round at a 30 billion dollar valuation. Jane Street has been expanding compute aggressively, signing a 6 billion dollar AI cloud agreement with CoreWeave in April 2026 that included Nvidia Vera Rubin hardware, and it was reported in June to be exploring building its own 100 to 200 megawatt data center. Other quantitative firms including IMC and Hudson River Trading have signed notable cloud contracts recently. Financial trading firms are now buying compute at a scale that used to be reserved for AI labs, and unlike the labs they tend to want capacity near existing low-latency footprints in the Northeast and Chicago rather than wherever power is cheapest. That is a demand pocket worth tracking separately.

Source: Data Center Dynamics (September 4, 2026)

Quantum leaves the lab: IBM's Nighthawk r2 and the facility requirements nobody has priced yet

IBM announced on September 4 that its Nighthawk r2 quantum processor is broadly available through the cloud, describing it as the most complex quantum processor the company has ever put into production. It executes more than 100,000 quantum circuits per second. Nighthawk was first unveiled in November 2025, and the architecture carries 120 qubits, 218 next-generation tunable couplers arranged in a square lattice, and 120 reset elements, for 458 total physical quantum elements. IBM says r2 maintains the scale of its r1 predecessor while targeting improvements in quality and speed, claiming computations 25 times faster than IBM Quantum Heron and accurate observable estimation on circuits containing more than 7,500 gates. Future Nighthawk iterations are expected to reach 10,000 gates in 2027.

The bottleneck IBM attacked is worth understanding, because it is a physical one rather than an algorithmic one. Qubits must return to their ground state between circuit executions, and the wait for that reset is dead time. Adding dedicated reset elements to the lattice is an infrastructure fix to a throughput problem, which is a very familiar shape of engineering to anyone who has worked on a data hall.

The commercial trajectory behind it is no longer speculative. IBM reported one billion dollars in cumulative quantum bookings covering the first quarter of 2017 through the fourth quarter of 2024, committed more than 10 billion dollars to quantum over five years in June 2026, and closed its acquisition of HRL Laboratories, the quantum R&D firm jointly owned by Boeing and General Motors, in late August. Nighthawk r2 will power the next generation of IBM quantum systems including an as-yet unannounced deployment.

Why it matters for North American builds: quantum systems are being installed in real buildings, and the requirements do not resemble anything in a standard AI fit-out. Superconducting processors run in dilution refrigerators at millikelvin temperatures, which means helium management, vibration isolation, tight electromagnetic shielding, very clean power and floor loading calculations that have nothing in common with a GPU rack. The power draw per system is modest but the tolerance for interference is not, which makes siting one inside an operating, densely populated AI hall a genuine engineering problem. As national labs, universities, defense-adjacent programs and now automotive and aerospace R&D groups move from cloud access to on-premise systems, somebody has to build those rooms. That work is small in megawatts and high in specification, which is exactly the profile that rewards contractors and vendors with mission-critical discipline over those competing on volume. Worth putting on the radar now, because the specifications for the first wave of these rooms are being written by people who have never had to write them before.

Source: Data Center Dynamics (September 4, 2026)