Date: November 3, 2026
Author: Saf Malik, Head of Content & Insights
Estimated Reading Time: 9 minutes
Executive Overview
The artificial intelligence boom has pushed the global digital infrastructure ecosystem to a dramatic inflection point. As next-generation foundational models and dense graphics processing unit (GPU) clusters become heavier, faster, and more power-hungry, traditional data centre engineering is running up hard against fundamental physical laws.
According to Steve Carlini, Chief AI Advocate at Schneider Electric, the industry is hurtling past conventional paradigms. In an extensive interview with Capacity, Carlini outlined the radical infrastructural shifts required to sustain the coming wave of AI hardware. By 2027, with the commercial arrival of platforms like NVIDIA’s Rubin Ultra, standard AI racks are projected to leap from approximately 400 kilowatts (kW) to a staggering 1 megawatt (MW) or more.
To safely feed these power-dense systems, data centre architecture must abandon legacy alternating current (AC) distribution in favour of 800-volt direct current (800V DC) power frameworks. Furthermore, grid constraints are forcing hyperscalers and neoclouds to look behind the meter, establishing autonomous energy parks powered by natural gas turbines and renewable integration. Concurrently, hotter liquid cooling systems are poised to dramatically slash data centre water consumption, while surging community pushback is permanently ending the era of stealthy, secretive facility development.
Detailed Chronology & Technological Evolution
The breathless pace of AI hardware innovation—driven heavily by NVIDIA’s aggressive release cycles—has fundamentally altered how infrastructure operators plan, design, and construct facilities. Carlini notes that Schneider Electric works hand-in-hand with hardware giants, publishing reference designs for each new GPU cluster roughly six months ahead of commercial launch to give operators adequate runway for facility-level engineering.
The 2026–2027 Horizon: The Shift to 800V DC
The physical limits of conventional AC power distribution have finally been reached. Traditional copper cabling simply lacks the volumetric efficiency required to feed individual servers operating at ultra-high densities.
To overcome this bottleneck, the power supply units (PSUs) are being physically extracted from the servers entirely and relocated to dedicated sidecar enclosures—such as Schneider’s NetShelter power racks. These auxiliary racks integrate localized batteries and capacitors designed to smooth out the violent, high-frequency transients and rapid power draws characteristic of massive GPU arrays. Meanwhile, the core compute racks house exclusively server boards and GPUs, enabling unprecedented hardware densification.
[Traditional AC Architecture]
Grid -> Heavy Copper Cabling -> In-Server PSUs -> Dense GPUs (Bottlenecked by physical space)
[Next-Gen 800V DC Architecture]
Grid -> High-Voltage DC -> External Sidecar Racks (NetShelter with integrated batteries/capacitors) -> Ultra-Dense GPUsCrucially, Carlini emphasizes that a single universal standard will not emerge. While NVIDIA’s architecture spans from 0V to 800V, alternative collaborative frameworks like "Mt Diablo"—developed alongside Google and the Open Compute Project—rely on a balanced plus-and-minus 400V DC topology. Battery and capacitor configurations will similarly vary across ecosystems. Hyperscalers will pioneer these deployments upon receiving the initial allocations of Rubin Ultra systems, with colocation providers and neoclouds expected to follow suit in subsequent years.
Supporting Context & Metrics: The Behind-the-Meter Power Crunch
Grid connectivity delays have transformed how developers approach site selection, particularly across North America. For gigascale developments, relying solely on traditional utility grid infrastructure is increasingly viewed as an exception rather than the rule.
The Rise of Behind-the-Meter Energy Parks
Data centre campuses are increasingly sourcing 50% to 100% of their total energy load on-site or from adjacent, dedicated generation facilities. Because ultra-large generation assets face catastrophic interconnection delays, developers are leveraging smaller, readily available 50MW natural gas turbines, paralleling them into cohesive 200MW generation blocks. While Schneider Electric does not manufacture turbines, the company supplies the specialized switchgear, uninterruptible power supplies (UPS), busways, and orchestration software required to stabilize and control these localized microgrids.
Despite soaring capital expenditures, demand shows no signs of decelerating. Major hyperscalers, neoclouds, and tier-one colocation providers have aggressively locked down their supply chains for the next two to three years. If localized overbuilding occurs anywhere globally, Carlini suggests it will likely be in the United States, given the sheer volume of capital and steel currently being deployed relative to other global markets.
This phase marks a pivotal transition year for enterprise AI. While training workloads and scientific research continue to consume massive tranches of capacity, the market is shifting toward inference workloads—the point at which enterprises begin actively monetizing AI investments by moving models from experimental pilots into live production.
Official Statements & Community Accountability: The End of Secrecy
The unprecedented physical scale of modern AI campuses has triggered intense public scrutiny. In regions like Texas and parts of Europe, local communities have pushed back aggressively against data centre developments over concerns regarding escalating utility bills, traffic congestion, and intensive water consumption.
Carlini candidly admits that the industry has made missteps. In the past, many data centre projects were initiated under clandestine code names and shell companies. While developers frequently argued that these masking tactics were necessary to protect intellectual property and strategic intent from rival tech firms, the practical consequence was identical: local residents often remained entirely unaware of a massive industrial facility planned for their neighbourhood until heavy excavation equipment arrived.
"It’s not acceptable to tap into the town’s water supplies," Carlini stated, emphasizing that data centres can no longer justify driving up utility costs for local residents.
That era of corporate opacity is officially over. Moving forward, operators must transparently disclose noise emissions, exact power sourcing, and water consumption metrics up front, securing genuine local social license before municipal permits are issued. Furthermore, innovative behind-the-meter generation strategies are increasingly being designed to generate structural power surpluses that can eventually be fed back into local municipal grids to stabilize regional power pricing.
Future Outlook: Liquid Cooling, Prefabrication, and the Token Economy
Looking toward the latter half of the decade, several transformative technologies are transitioning from niche engineering concepts to mainstream mandates.
1. Hotter Chips and Dramatic Water Savings
NVIDIA’s forthcoming Rubin systems are engineered to operate liquid-cooling loops at elevated temperatures reaching 45°C (113°F). This higher thermal delta enables new facilities to rely almost exclusively on dry coolers year-round, entirely eliminating the need for energy-intensive mechanical chillers, evaporative cooling towers, and water-spray assists in many climates. Furthermore, operators are increasingly bypassing potable municipal drinking water supplies entirely, opting instead to recycle and treat industrial wastewater for closed-loop server cooling.
2. The Dominance of Prefabricated Infrastructure
Faced with rapidly shifting rack densities that outpace traditional on-site construction timelines, prefabrication has evolved into an essential deployment strategy. Factory-built power and cooling modules—including pre-engineered IT pods designed specifically for NVIDIA and AMD Helios clusters—can be manufactured, tested, and dropped onto job sites in flat-packed, multi-storey configurations while civil engineering and foundational work occur in parallel.
3. Digital Twins and the Simulation Bottleneck
While platforms like NVIDIA Omniverse (utilizing Schneider-owned AVEVA for cooling and ETAP for power simulation) represent the ultimate blueprinting tool for next-generation data centres, their widespread adoption hinges on the availability of "SimReady" digital assets. Once a comprehensive library of physically accurate component models becomes commercially available, operators will be able to simulate entire facilities and instantly generate precise bills of materials.
4. Redefining Capacity: From Megawatts to Tokens
As enterprises move toward right-sized IT stacks and specialized models tailored to specific operational tasks, the very metric used to measure data centre scale is evolving. Carlini forecasts that tokens—the foundational units into which AI models parse text, imagery, and audio—will soon sit alongside megawatts as a primary benchmark for sizing and commercializing data centre capacity.
"It could be a one gigawatt data centre," Carlini notes, "but how many tokens is it good for?"
As the industry approaches critical regulatory and utility interconnection tipping points toward the end of 2027, the marriage of high-voltage DC power, decentralized microgrids, and intelligent thermal management will dictate which operators successfully navigate the AI era—and which ones stall out against the physical grid.
