What 6 GW of GPUs Really Means The 6 GW of accelerator load envisioned under the OpenAI–AMD partnership will be distributed across multiple hyperscale AI factory campuses. If OpenAI begins with 1 GW of deployment in 2026, subsequent phases will likely be spread regionally to balance supply chains, latency zones, and power procurement risk. Importantly, this represents entirely new investment in both power infrastructure and GPU capacity. OpenAI and its partners have already outlined multi-GW ambitions under the broader Stargate program; this new initiative adds another major tranche to that roadmap. Designing for the AI Factory Era These upcoming facilities are being purpose-built for next-generation AI factories, where MI450-class clusters could drive rack densities exceeding 100 kW. That level of compute concentration makes advanced power and cooling architectures mandatory, not optional. Expected solutions include: Warm-water liquid cooling (manifold, rear-door, and CDU variants) as standard practice. Facility-scale water loops and heat-reuse systems—including potential district-heating partnerships where feasible. Medium-voltage distribution within buildings, emphasizing busway-first designs and expanded fault-current engineering. While AMD has not yet disclosed thermal design power (TDP) specifications for the MI450, a 1 GW campus target implies tens of thousands of accelerators. That scale assumes liquid cooling, ultra-dense racks, and minimal network latency footprints, pushing architectures decisively toward an “AI-first” orientation. Design considerations for these AI factories will likely include: Liquid-to-liquid cooling plants engineered for step-function capacity adders (200–400 MW blocks). Optics-friendly white space layouts with short-reach topologies, fiber raceways, and aisles optimized for module swaps. Substation adjacency and on-site generation envelopes negotiated during early land-banking phases. Networking, Memory, and Power Integration As compute density scales, networking and memory bottlenecks will define infrastructure design. Expect fat-tree and dragonfly network topologies, 800 G–1.6 T interconnects, and aggressive optical-module roadmaps to minimize collective-operation latency, aligning with recent disclosures from major networking vendors.

In a new report spanning 2022 through 2024, the Union of Concerned Scientists (UCS) identifies a significant regulatory gap in the PJM Interconnection’s planning and rate-making process—one that allows most high-voltage (“transmission-level”) interconnection costs for large, especially AI-scale, data centers to be socialized across all utility customers. The result, UCS argues, is a multi-billion-dollar pass-through that is poised to grow as more data center projects move forward, because these assets are routinely classified as ordinary transmission infrastructure rather than customer-specific hookups. According to the report, between 2022 and 2024, utilities initiated more than 150 local transmission projects across seven PJM states specifically to serve data center connections. In 2024 alone, 130 projects were approved with total costs of approximately $4.36 billion. Virginia accounted for nearly half that total—just under $2 billion—followed by Ohio ($1.3 billion) and Pennsylvania ($492 million) in data-center-related interconnection spending. Yet only six of those 130 projects, about 5 percent, were reported as directly paid for by the requesting customer. The remaining 95 percent, representing more than $4 billion in 2024 connection costs, were rolled into general transmission charges and ultimately recovered from all retail ratepayers. How Does This Happen? When data center project costs are discussed, the focus is usually on the price of the power consumed, or megawatts multiplied by rate. What the UCS report isolates, however, is something different: the cost of physically delivering that power: the substations, transmission lines, and related infrastructure needed to connect hyperscale facilities to the grid. So why aren’t these substantial consumer-borne costs more visible? The report identifies several structural reasons for what effectively functions as a regulatory loophole in how development expenses are reported and allocated: Jurisdictional split. High-voltage facilities fall under the Federal Energy Regulatory Commission (FERC), while retail electricity rates are governed by state public utility

With the conclusion of the 2025 OCP Global Summit, William G. Wong, Senior Content Director at DCF’s sister publications Electronic Design and Microwaves & RF, published a comprehensive roundup of standout technologies unveiled at the event. For Data Center Frontier readers, we’ve revisited those innovations through the lens of data center impact, focusing on how they reshape infrastructure design and operational strategy. This year’s OCP Summit marked a decisive shift toward denser GPU racks, standardized direct-to-chip liquid cooling, 800-V DC power distribution, high-speed in-rack fabrics, and “crypto-agile” platform security. Collectively, these advances aim to accelerate time-to-capacity, reduce power-distribution losses at megawatt rack scales, simplify retrofits in legacy halls, and fortify data center platforms against post-quantum threats. Rack Design and Cooling: From Ad-Hoc to Production-Grade Liquid Cooling NVIDIA’s Vera Rubin compute tray, newly offered to OCP for standardization, packages Rubin-generation GPUs with an integrated liquid-cooling manifold and PCB midplane. Compared with the GB300 tray, Vera Rubin represents a production-ready module delivering four times the memory and three times the memory bandwidth: a 7.5× performance factor at rack scale, with 150 TB of memory at 1.7 PB/s per rack. The system implements 45 °C liquid cooling, a 5,000-amp liquid-cooled busbar, and on-tray energy storage with power-resilience features such as flexible 100-amp whips and automatic-transfer power-supply units. NVIDIA also previewed a Kyber rack generation targeted for 2027, pivoting from 415/480 VAC to 800 V DC to support up to 576 Rubin Ultra GPUs, potentially eliminating the 200-kg copper busbars typical today. These refinements are aimed at both copper reduction and aisle-level manageability. Wiwynn’s announcements filled in the practicalities of deploying such densities. The company showcased rack- and system-level designs across NVIDIA GB300 NVL72 (72 Blackwell Ultra GPUs with 800 Gb/s ConnectX-8 SuperNICs) for large-scale inference and reasoning, and HGX B300 (eight GPUs /

AI deployment uses multiple storage layers, and each one has different requirements, says Dell’Oro’s Fung. For storing massive amounts of unstructured, raw data, cold storage on HDDs makes more sense, he says. SSDs make sense for warm storage, such as for pre-processing data and for post-training and inference. “There’s a place for each type of storage,” he says. Planning ahead According to Constellation’s Mehta, data center managers and other storage buyers should prepare by treating SSD procurement like they do GPUs. “Multi-source, lock in lanes early, and engineer to standards so vendor swaps don’t break your data path.” He recommends qualifying at least two vendors for both QLC and TLC and starting early. TrendForce’s Ao agrees. “It is better to build inventory now,” he says. “It is difficult to lock-in long term deals with suppliers now due to tight supply in 2026.” Based on suppliers’ availability, Kioxia, SanDisk, and Micron are in the best position to support 128-terabyte QLC enterprise SSD solutions, Ao says. “But in the longer term, some module houses may be able to provide similar solutions at a lower cost,” Ao adds. “We are seeing more module houses, such as Phison and Pure Storage, supporting these solutions.” And it’s not just SSD for fast storage and HDD for slow storage. Memory solutions are becoming more complex in the AI era, says Ao. “For enterprise players with smaller-scale business models, it is important to keep an eye on Z-NAND and XL-Flash for AI inference demand,” he says. These are memory technologies that sit somewhere between the SSDs and the RAM working memory. “These solutions will be more cost-effective compared to HBM or even HBF [high bandwidth flash],” he says.

Data center information management covers a lot of ground. Here are five of the most important things you need to know about DCIM.

Meanwhile, in a release issued last week, CoreWeave stated, “it has been unequivocal — to Core Scientific and publicly — that we will not modify our offer. Our offer is best and final.” Alvin Nguyen, senior analyst at Forrester Research, said what happens next with the overall data center market “depends on when AI demand slows down (when the AI bubble bursts).” He added, “if AI demand continues, prices continue to go up, and data centers change in terms of preferred locations (cooler climates, access to water, lots of space, more remote), use of microgrids/energy production, expect [major] players to continue to dominate.” However, said Nguyen, “if that slowdown is soon, then prices will drop, and the key players will need to either unload property or hold onto them until AI demand builds back up.” Generational shift occurring Asked what the overall effect of AI will be on CIOs in need of data center capacity, he said, “the new AI mega-factories alter data center placement: you don’t put them near existing communities because they demand too much power, water, land, you build them somewhere remote, and communities will pop up around them.” Smaller data centers, said Nguyen, “will still consume power and water in contention with their neighbors (industrial, commercial, and residential), potential limiting their access or causing costs to rise. CIOs and Network World readers should evaluate the trade offs/ROI of not just competing for data center services, but also for being located near a new data center.”