Vacancy in the North American market has declined to a new record low of 2.3 percent, and JLL projects that figure will remain the same or go even further down from now through 2027. For comparison, the vacancy rate stood at 9.8 percent in 2020. As bad as the wait for data center capacity has become, the wait for power is even worse. The average wait time for a grid connection across North America is now four years, according to the report, with power delays representing a significant hurdle in efforts to alleviate the shortage of new colocation capacity. Most of the top markets have doubled or even tripled in size since 2020, with Columbus, Ohio leading the way with 1800% growth, followed by Austin/San Antonio at 500% growth. However, they started from a small base in 2020. In absolute terms, Northern Virginia (+3,975 MW), Dallas (+1,008 MW) and Atlanta (+828 MW) have seen the largest increase in capacity.

The problem is that less than one-eighth of enterprises said that they believed HPE could justify the Juniper deal simply by having HPE salespeople sell Juniper gear. We can take JHPE off the table. But not the Mist influence. Nearly all the enterprises who offered comments on the deal want to see HPE expand Mist to include HPE management, becoming an element in a “full-stack observability” strategy, creating our “Mist4All” name. Three dozen told me that their sales team had suggested this would happen, but interestingly, none who also talked with HPE or Juniper executives said they repeated that promise. HPE’s problem is that only 11 of over 250 enterprises said a move to offer integrated AI observability would even make them consider switching from Cisco to the new company. The problem with vendor-specific observability like Mist4All is that you can’t introduce it incrementally; you need to fork-lift. Only one-fifth of enterprises thought HPE could justify the deal by integrating their own operations tools with Mist. Forget Mist4All, too. Which brings us to HPEJ. I don’t think there’s any question that the biggest driver in how the new venture positions itself is AI. AI is also the most likely near-term driver of incremental data center deployment, both among enterprises and among service providers, something Juniper’s AI-Native positioning aims to exploit and that competitors like Cisco and Extreme are countering. And, of course AI data centers obviously need servers, so HPE has its own interests here. That’s what creates the HPEJ/JHPE tension, but can shareholder value and customer interest be created by simply doubling down on what the companies were doing separately? Three enterprises who use both HPE and Juniper gear enough to get significant account attention from both say that Juniper’s people push AI hosting more than HPE does. Nobody

The Nuclear Regulatory Commission approved construction of the Oak Ridge facility in November 2024, marking the first non-water-cooled reactor permitted in the US in more than 50 years. Kairos Power planned to increase the Hermes 2 reactor’s output from an originally planned 28 megawatts to 50 megawatts, with operations scheduled to begin in 2030. Google’s agreement extended to up to 500 megawatts of Kairos reactors by 2035, the statement added. Enterprise power crunch drives nuclear innovation The technology breakthrough comes at a critical time for enterprises facing explosive electricity demands from artificial intelligence operations. Data center electricity use rose from 58 terawatt-hours in 2014 to 176 TWh in 2023, and could reach between 325 and 580 TWh by 2028, according to Department of Energy projections. For context, a single graphics processing unit used for AI can consume as much electricity as a four-person household uses daily. Unlike renewable energy sources, nuclear reactors provide constant baseload power regardless of weather conditions — critical for data centers that operate around the clock. Google said it needed the nuclear power to support data centers in Montgomery County, Tennessee, and Jackson County, Alabama. The 50-megawatt output equals the electricity consumption of about 36,000 homes. “To power the future, we need to grow the availability of smart, firm energy sources,” Amanda Peterson Corio, Google’s global head of data center energy, said in the statement.

Server CPU shipments were “uninteresting,” as he put it, on a sequential basis, with neither supplier seeing much growth on quarter after last quarter’s atypical increase. On-year, the server market was up significantly but that’s because a year ago, the segment was near its cyclical lows and dealing with inventory adjustments. “Intel was able to sustain volumes in total server unit shipments by moving shipments to non-data center products, such as Xeon D in networking/storage servers, which they noted in their earnings call. That comes at a price; those products have much lower ASPs, so lower revenues, which is why Intel’s DCAI revenues were lower when units were flat,” McCarron told Network World. “Nothing really moves that fast in servers, and in general a ‘freefall’ can’t really happen outside of some systemic demand collapse event like 2008 was, as the rest of the industry realistically can’t absorb market share at an unlimited rate due to supply chain considerations,” he added. AMD’s server revenues hit a record high, but most of the revenue gains was from selling a higher mix of its new Turin core CPUs, and unit shipment growth was very modest. However, even with a 0.1-point increase in share, that means a new record high in AMD server sales. It now has 37.2% market share. Excluding IoT/SoC embedded products from consideration, Intel’s shipments slightly outgrew AMDs in the quarter resulting in Intel having a modest sequential share increase thanks in part to mobile CPU shipments, where Intel has solid products. AMD made a slight gain in desktops, where it is particularly strong. As for Arm, it showed strength in the server market thanks to Nvidia’s GB200 processors ramping up volumes. On the client side, Apple had slightly higher shipments in the second but that was offset by weakness in

AI infrastructure ahead of plan and growing Cisco reported more than $800 million in AI infrastructure orders from webscale customers in the fourth quarter, bringing the total for 2025 to more than $2 billion, which is over 2x what the company originally projected. This is a mix of its own Nexus switches, optics, AI PODs, UCS servers and Silicon One. Success here is critical for Cisco, as at one time, the company had next to no business with the hyperscalers. The development of Silicon One was pivotal in Cisco’s success with this audience, as it’s given them market leading price performance. Cisco has also cultivated a partnership with Nvidia and is the only company to have its silicon integrated into the GPU maker’s Spectrum-X product. There is another wave of business coming for Cisco in this area selling AI infrastructure to non-hyperscalers. Robbins talked about this: “The Cisco Secure AI Factory with Nvidia provides a blueprint for building AI-ready data centers for enterprises, sovereign cloud providers and newly emerging neocloud providers. We expect the sovereign AI opportunity to build momentum in the second half of fiscal year ’26.” AI will drive campus upgrades Most of the focus of network growth in AI has been in the data center, as that’s where the growth has been. However, the traffic agentic AI creates will drive campus upgrades as well. On the call, Cisco showed a chart of traffic generated by chatbots pre and post agentic, and it shows Cisco is expecting agentic to drive a consistent level of traffic that most networks are not able to handle. Robbin explained: “Network traffic will not only increase beyond the peaks of current chatbot interaction but will remain consistently high with agents in constant interaction.” The impact of this is twofold. The bump in traffic

More Nvidia news from SIGGRAPH The RTX Pro server wasn’t the only news at the show. Nvidia also announced two new models of its Nemotron model family – the Nemotron Nano 2 and Llama Nemotron Super 1.5 – with advanced reasoning capabilities for building smarter AI agents. Nemotron is a family of enterprise-ready, open large language models (LLM) designed to enhance agentic AI for tasks requiring sophisticated reasoning, instruction following, coding, tool use, and multimodal (text+vision) understanding. These models deliver high accuracy for their relative size in areas such as scientific reasoning, coding, tool use, instruction following and chat, according to Nvidia. They are designed to imbue AI agents with deeper cognitive abilities and help AI systems explore options, weigh decisions and deliver results within defined constraints. Nvidia claims Nemotron Nano 2 achieves up to six times higher token generation throughput compared to other models its size. Llama Nemotron Super 1.5 offers top-tier performance and leads in reasoning accuracy, making it suitable for handling complex enterprise tasks. Also, Nvidia is empowering robotics and machines to “see” and react to what they see with new AI models that can ingest visual information and think about said information. The vendor just announced Cosmos Reason, a new open, customizable 7 billion-parameter reasoning Vision Language Models, or VLMs. VLMs allow robots and vision agents to think about what they see, just like a human. Up to now, robots have had the ability to “see,” but their reaction to what they saw was extremely limited. A VLM provides robotics with the ability to think about their actions.