> ## Content Index > Fetch the complete content index at: https://www.aixenergy.io/llms.txt > Use this file to discover other available public pages before exploring further. # The Millisecond Menace - URL: https://www.aixenergy.io/the-millisecond-menace/ - Published: 2026-06-10T14:04:08.000Z - Updated: 2026-07-29T19:42:34.000Z - Description: The 800-volt DC data center promises faster, leaner AI infrastructure. But synchronized GPU power swings are turning electrical design into a millisecond-scale control problem. - Author: Morgan Bazilian - Tags: Shadow Grid by [Morgan Bazilian](https://www.linkedin.com/in/morganbazilian/?ref=aixenergy.io) and [Denis Kouroussis](https://www.linkedin.com/in/denis-kouroussis/?ref=aixenergy.io) Global data center electricity consumption reached an estimated [415 terawatt-hours in 2024](https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai?ref=aixenergy.io), roughly 1.5 percent of all electricity generated worldwide. The International Energy Agency (IEA) projects that figure will nearly double to [945 terawatt-hours by 2030](https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary?ref=aixenergy.io), growing at around 15 percent per year, more than four times the rate of every other sector combined. In the United States, [Lawrence Berkeley National Laboratory](https://newscenter.lbl.gov/2025/01/15/berkeley-lab-report-evaluates-increase-in-electricity-demand-from-data-centers/?ref=aixenergy.io) estimates that data centers consumed 176 terawatt-hours in 2023, representing 4.4 percent of total national electricity use, with projections ranging from 325 to 580 terawatt-hours by 2028\. The forces behind this expansion are familiar: larger models, faster chips, and seemingly relentless scale-out. What receives far less attention is what happens once all that power arrives at the site. 0:00 /0:35 1× The answer the industry seems to be converging on, in its more forward-looking configurations, is 800 V direct current (DC) distribution fed by solid-state transformers (SST) that convert medium-voltage (MV) alternating current (AC) directly to the DC bus in a single step. [NVIDIA announced 800 V DC as its target rack architecture at Computex 2025](https://developer.nvidia.com/blog/nvidia-800-v-hvdc-architecture-will-power-the-next-generation-of-ai-factories/?ref=aixenergy.io), specifying it for Kyber rack-scale systems scheduled for full production in 2027\. But, beneath the efficiency gains lies a control-systems problem the industry has yet to fully confront: the interaction between millisecond-scale Graphics Processing Unit (GPU) power transients, cascaded power electronics, and the distributed impedances of the DC bus. This is an engineering problem that will determine whether these architectures deliver on their promise. ## The Case Against Conventional Architecture 💡 Conventional data center power architecture moves electricity through multiple conversion stages before it reaches GPUs, adding cost, long equipment lead times, and 6–8% energy losses—enough to waste 6–8 MW in a 100 MW facility. The emerging 800 V DC architecture, using solid-state transformers to feed racks directly from medium-voltage utility service, could collapse that equipment stack and improve efficiency, but it introduces a more complex control-systems challenge that must still be proven at scale. The traditional path from a medium-voltage utility feed to a GPU involves four discrete conversion steps: a medium-voltage transformer, low-voltage switchgear, an uninterruptible power supply, and a power distribution unit. Each adds losses, cost, and a procurement lead time of its own. Together they dissipate 6 - 8% of input energy before it reaches the compute tray. At 100 MW of grid input, a conventional 480 V Alternating Current (AC) architecture wastes 6 to 8 MW in distribution hardware alone, capacity that could otherwise run GPUs. The supply-chain burden of this equipment stack has become severe. [Wood Mackenzie's second-quarter 2025 survey](https://www.woodmac.com/news/opinion/mind-the-gap-tackling-supply-chain-challenges-in-the-electric-td-sector/?ref=aixenergy.io) found standard power transformers averaging 128 weeks for delivery and generator step-up units running to 144 weeks, with switchgear at 44 weeks. The [North American Electric Reliability Corporation](https://www.theinvadingsea.com/2025/12/05/power-grid-electricity-supply-chain-data-centers-transformers-circuit-breakers-utilities/?ref=aixenergy.io) recorded lead times crossing 120 weeks for power transformers in 2024, with larger transformers stretching to 210 weeks. Demand for high-voltage power transformers has grown 116 percent since 2019, according to Wood Mackenzie, and generator step-up units by 274 percent, faster than domestic manufacturing capacity has been able to respond. For a data hall that requires four MV transformers, multiple switchgear assemblies, Uninterruptible Power Supply (UPS) systems, and dozens of Power Distribution Units (PDUs), these timelines dominate the construction schedule in ways that site preparation and civil works cannot offset. The 800 V DC approach collapses this equipment stack. By connecting a solid-state transformer directly to the 34.5 kV utility feed and delivering 800 V DC to the rack, the intermediate conversion chain disappears. Companies including [Heron Power Electronics](https://www.heronpower.com/?ref=aixenergy.io), [Delta Electronics](https://www.deltapowersolutions.com/en/mcis/news-2025-delta-demonstrate-hvdc-power-cooling-networking-to-drive-ai-data-center-evolution-at-ocp-global-summit.php?ref=aixenergy.io), and [Enphase Energy](https://investor.enphase.com/news-releases/news-release-details/enphase-energy-announces-development-iq-solid-state-transformer?ref=aixenergy.io) are developing Solid-State Transformer (SST) platforms claiming conversion efficiencies above 98 percent, against the 93 to 94 percent typical of the conventional chain. A [2025 simulation study by Xu et al.](https://arxiv.org/pdf/2601.16502?ref=aixenergy.io) published on arXiv provides one of the more rigorous technical evaluations of this architecture, demonstrating stable SST-based 800 V DC operation under realistic Artificial Intelligence (AI) workload profiles in real-time digital simulation. It also documents the complexity of coordinating the multiple nested control loops that SST operation requires, a complexity the industry tends to understate. Heron Power Electronics recently published a [technical blueprint](https://www.heronpower.com/?ref=aixenergy.io) proposing a 12 MW data hall building block powered by four 4.2 MW solid-state transformers, each connecting to a 34.5 kV feed and delivering 800 V DC to forty compute racks in a 4-to-make-3 redundant topology. The company claims that eliminating the conventional MV transformer, switchgear, UPS, and PDU cuts equipment and installation costs by 65 to 90 percent, reduces construction timelines by roughly three months, and eliminates the dedicated electrical room. Whether those figures hold across diverse project conditions remains to be demonstrated at scale. ## When GPUs Breathe Together 💡 AI training clusters behave unlike conventional data center loads: thousands of GPUs spike, pause, and surge together, creating synchronized power swings that can propagate from racks to the utility interconnect and reach tens or hundreds of megawatts. Current stabilization methods can reduce those swings, but they often burn energy simply to hold power steady, meaning some of the efficiency gains promised by 800 V DC architectures may be offset by the need to control GPU-driven transients. Modern GPU architectures bear little resemblance to traditional electrical loads. A data center chiller ramps over seconds; conventional server demand shifts gradually with workload. AI training clusters behave differently. During matrix computations, GPU power spikes; during data transfers and synchronization, it drops sharply. Checkpointing operations can briefly reduce load to near zero before snapping back to full demand. A 2025 paper from Microsoft, OpenAI, and NVIDIA, ["Power Stabilization for AI Training Datacenters" by Choukse et al.](https://arxiv.org/abs/2508.14318?ref=aixenergy.io), documents that these fluctuations span 30 to 100 percent of thermal design power (TDP) within windows as short as 0.2 to 2 seconds, with intra-batch variations occurring at millisecond granularity. Distributed AI training synchronizes thousands of GPUs through collective operations such as AllReduce; every GPU in a cluster steps through the same iterations in lockstep. The facility's power draw does not fluctuate randomly in ways that average out. [Oracle's cloud infrastructure team](https://blogs.oracle.com/cloud-infrastructure/behind-the-scenes-gpu-power-smoothing-ai-training?ref=aixenergy.io) has documented seeing this signature propagate from the rack level to the data center level and ultimately to the utility interconnect, where swings can reach tens or even hundreds of megawatts. The Choukse paper notes that the frequency spectrum of these oscillations, concentrated between 0.2 and 3 Hz, falls close to known resonant modes of turbine-generator shafts and long transmission lines, raising the prospect of mechanical stress in upstream generation equipment. [SemiAnalysis has reported](https://newsletter.semianalysis.com/p/ai-training-load-fluctuations-at-gigawatt-scale-risk-of-power-grid-blackout?ref=aixenergy.io) that multiple utility providers have recorded harmonic impacts from synchronized computing loads, with amplitude growing proportionally to cluster size. The Choukse paper documents two mitigation strategies currently deployed in production, and their energy costs are instructive. The first, called Firefly, injects controlled secondary workloads during low-power communication phases to sustain a more uniform draw. It works, but as the authors acknowledge, energy burned on artificial computations that produce no useful output is simply wasted. The second uses NVIDIA's GB200 Minimum Power Floor feature, which prevents GPU power from dropping below a specified fraction of thermal design power even when no useful work is running. At a floor set to 90 percent of TDP, the paper reports a total energy overhead of 10.5 percent. For a 100 MW training cluster, that represents 10.5 MW of continuous power consumption that produces no computational output, burned purely to prevent the consequences of load transients. They are in production because unmitigated swings at critical grid frequencies pose genuine risks to interconnection agreements and, at sufficient scale, to grid stability. A system that promises 50 percent lower distribution losses but requires 10 percent energy overhead for load stabilization has surrendered a significant portion of its claimed advantage, before accounting for the additional complexity of coordinating GPU-level controls with SST-level and battery-level controls across a distributed DC bus. ## The Architecture of Stability 💡 In an 800 V DC architecture, high-C-rate batteries integrated at the solid-state transformer can smooth the data center’s grid-facing load, but they may not respond fast enough to clamp the sharpest voltage excursions at distant racks. That means stability likely requires a coordinated hierarchy: local capacitance near the racks for sub-millisecond events, bulk storage at the SST for larger slower swings, and carefully tuned control loops so the battery, SST, and rack converters do not fight each other. The conventional response to power transients in data centers has been to place a battery buffer behind the UPS, centralizing storage at the conversion stage. In that configuration, the storage sits at a low-impedance node close to the source of regulation, and fast transients are absorbed before they propagate. The 800 V DC architecture offers a different option: integrating a high-C-rate battery directly onto the DC bus at the SST. Heron's blueprint co-locates a battery buffer unit with each Heron Link, coupling it directly to the 800 V DC bus without an additional DC-to-DC converter, so it can respond to load fluctuations faster than any mechanical switching device and present a smooth draw to the grid. But, a transient absorber is only as effective as the impedance path between it and the transient source allows. Storage close to the load, with low series inductance and resistance, can respond fast and clamp hard. Storage at the SST, separated from each rack by the full length of the DC distribution cable runs, sees a higher-impedance path. By the time charge traverses that network, the voltage excursion may have already occurred at the rack. The practical result is that SST-integrated bulk storage smooths the load profile as seen from the grid while doing comparatively less to clamp the fastest voltage excursions at the rack itself. The two storage roles may therefore be complements rather than substitutes: modest local capacitance at the rack handles sub-millisecond events that distribution impedance would otherwise allow to propagate, while bulk energy storage at the SST handles the slower, larger swings that constitute the bulk of grid-visible fluctuation. The [Xu et al. simulation study](https://arxiv.org/pdf/2601.16502?ref=aixenergy.io) addresses SST control architecture in detail. The paper models a cascaded H-bridge front-end rectifier converting medium-voltage AC to intermediate DC buses, with dual-active-bridge (DAB) converters providing galvanic isolation and regulating the 800 V DC output. The control scheme is structured hierarchically: an outer voltage loop generates current references for an inner decoupled DQ current controller in the rectifier stage, while the DAB employs a simpler single-loop phase-shift controller. When the integrated battery and the SST's conversion stage share regulation of the same 800 V DC bus, a further complication emerges. Two control systems acting simultaneously on one node can, if designed without explicit coordination, interpret each other's corrections as disturbances and produce the oscillatory instability the storage was meant to prevent. This is a well-documented pathology in cascaded power electronics, traceable to the foundational work of [Middlebrook and Cuk in 1976](https://premiermag.com/wp-content/uploads/2022/01/H2PToday1911%5Fdesign%5FTexasInstruments%5Fpart10.pdf?ref=aixenergy.io), who showed that two individually stable subsystems can become unstable when cascaded if the output impedance of the upstream stage is not significantly lower than the input impedance of the downstream stage across all relevant frequencies. In the SST context, the minor-loop gain, defined as the ratio of the SST's output impedance to the load's input impedance, must be verified across the full operating envelope. As rack power demand pulses, the effective input impedance changes, and stability margins satisfied at the nominal design point may be violated at another. ## Cables, Bus Bars, and Hidden Resonances 💡 In 800 V DC data centers, cables, bus bars, and rack converters form a lightly damped resonant network where SST control actions meant to stabilize voltage can instead excite hidden oscillations that AC engineering intuition may miss. The move to 800 V DC changes the electrical character of the distribution network in ways that are not always intuitive to engineers trained on AC systems. DC cables and bus bars present primarily resistive and inductive impedance. A 50-meter cable run to a remote rack may present 20 to 30 microhenries of series inductance, modest in isolation but consequential in aggregate. Multiplied across dozens of parallel paths and combined with the input capacitance of rack-level DC-DC converters, the result is a distributed LC network with multiple resonant modes. These resonances matter because the SST's control loops operate at frequencies that can excite them. A voltage control loop with 200 Hz bandwidth, responding to a load step, injects energy from DC through several hundred hertz. If a cable-capacitor resonance falls within this range, the correction can excite the bus rather than stabilize it. The problem is compounded by the relative absence of damping in DC systems. AC distribution benefits from the inherent damping of line-frequency transformers, skin-effect losses, and resistive load components. DC systems optimized for efficiency minimize resistance everywhere, producing high-Q resonances that persist longer and ring to higher amplitudes. The tools and intuitions developed for 60 Hz AC distribution do not transfer directly to a DC network driven by fast-switching power electronics. ## The Regulatory Environment 💡 As AI data centers grow into hundreds-of-megawatt grid loads, regulators are beginning to treat them less like passive customers and more like active power-electronics systems that must prove ride-through capability, dynamic stability, and grid-compatible performance through commissioning tests conventional facilities were never designed to meet. These technical challenges do not exist in isolation from the regulatory environment. As AI campus loads grow to hundreds of megawatts on single feeders, grid operators are raising interconnection requirements. ERCOT, which operates the Texas grid and is processing more large-load interconnection requests than any other US operator, has become a practical reference point. [ERCOT's June 2025 Market Notice M-B062325-01](https://www.ercot.com/services/comm/mkt%5Fnotices/M-B062325-01?ref=aixenergy.io) formally requested voltage ride-through capability data from all data center loads of 75 MW or greater, citing reliability concerns about large electronic loads during grid disturbances. The December 2025 implementation of PGRR115 established mandatory interconnection requirements for such loads, including dynamic modeling standards. [ERCOT's December 2025 constraints report](https://www.ercot.com/files/docs/2025/12/23/2025-Report-on-Existing-and-Potential-Electric-System-Constraints-and-Needs.pdf?ref=aixenergy.io) notes that the operator is tracking more than three times as many large-load interconnection requests as in 2024, with the total queue pointing toward 181 GW by 2030, more than double the grid's current peak demand of 85 GW. IEEE 2800, the interconnection standard for inverter-based resources, is increasingly referenced as the model for what regulators will require of active power electronics at the grid interface. Facilities built on conventional passive-load architecture cannot meet these requirements without significant additional equipment. An SST with integrated battery storage, when properly designed, addresses them from the outset. The SuperBBU architecture in Heron's blueprint is designed to absorb AI training transients directly on the 800 V DC bus and provide sufficient hold-up time for backup generator startup during grid disturbances. But native capability and demonstrated capability are different things. The latter requires commissioning tests that conventional electrical infrastructure has never been asked to perform. ## Monitoring and the Operational Shift 💡 800 V DC AI data centers will require continuous, high-speed waveform monitoring and real-time control tuning because their electrical “normal” is a shifting dynamic envelope where instabilities can emerge faster than conventional power monitoring—or human operators—can respond. If 800 V DC SST architectures cannot be designed once and left alone, they must be continuously monitored and periodically retuned. This is a genuine departure from how electrical infrastructure has historically been operated. Traditional power monitoring focuses on steady-state quantities, voltage, current, power factor, harmonic content, sampled at low rates and compared against static thresholds. AI data center power systems violate each of these assumptions. Load changes occur faster than conventional monitoring can capture. Normal is not a fixed operating point but a dynamic envelope that shifts with workload, ambient temperature, and component aging. Instabilities can develop within the bandwidth of the control loops driving them, far faster than any human can intervene. What is needed is continuous system identification: waveform capture at millisecond or sub-millisecond resolution, synchronized across the distribution network, feeding real-time analytics capable of identifying emerging instabilities before they become hardware failures. A voltage transient at one rack correlated with a current spike at the SST tells a different story than either measurement alone. AI data center power infrastructure must be continuously engineered while operating, with feedback loops from monitoring to control tuning to system modification. The boundary between commissioning and ongoing operation needs to be substantially narrowed. One platform that addresses this monitoring requirement directly is Volta Insite. The platform captures full waveforms at 20 kHz and above at every feed and branch circuit simultaneously, providing the sub-millisecond resolution across the distribution network that SST-based stability analysis requires. Domain-trained AI agents perform continuous predictive analysis using LLM-based reasoning, correlating power quality anomalies, voltage ripples at the SST output, current transients at the rack, across the full system before individual events compound into hardware failures. ## A Practical Path Forward 💡 800 V DC data center architectures can deliver real efficiency gains, but only if the industry treats them as dynamic control systems requiring impedance-based stability analysis, coordinated storage controls, GPU-like commissioning tests, and continuous high-speed monitoring from rack to grid. The 800 V DC solid-state transformer architecture is not inherently unstable. It is inherently dynamic and demanding of a level of control-systems rigor that the data center industry has not historically applied to electrical infrastructure. Closing that gap requires several changes in practice. ![](https://storage.ghost.io/c/e3/c9/e3c9e740-20f7-4e9c-9a31-131fc1166819/content/images/2026/06/ChatGPT-Image-Jun-10--2026--10_25_09-AM.png) **Figure 1**: This diagram compares the conventional 480 V AC data center power chain with an SST-based 800 V DC architecture. The 800 V DC path can reduce conversion stages, losses, and equipment footprint, but it also turns AI power delivery into a dynamic control problem as synchronized GPU load swings interact with cables, converters, batteries, and monitoring systems. Impedance-based stability analysis must become a standard design deliverable. Every interface, SST to battery, SST to distribution bus, bus to rack, needs to be characterized across frequency and verified against Middlebrook criteria or more sophisticated equivalents across the full operating envelope. Control loop interactions need to be co-simulated with realistic models of communication latency and sensor noise. Distributed energy storage requires coordinated control from the start, droop sharing, virtual impedance shaping, or explicit bandwidth hierarchy, rather than as a corrective measure applied after integration problems surface. Commissioning protocols need to include dynamic testing. Static load tests confirm that a system can deliver rated power. They say nothing about whether it remains stable while doing so. Step loads, pulse trains, and synthetic GPU-like load signatures should be standard elements of acceptance testing, alongside millisecond-resolution waveform monitoring synchronized across the full distribution network. The efficiency gains of 800 V DC distribution are real and material. But, the transition asks something the industry has not fully acknowledged: that data center electrical infrastructure is no longer a static distribution problem but a dynamic control problem, operating at power levels and response times where the margin for error is measured in megawatts and the timescale of failure is measured in milliseconds. The GPUs driving AI's computational expansion breathe in synchronized rhythms that propagate through power electronics, cables, and bus bars to the utility grid itself. Managing that breathing requires a synthesis of power electronics, control theory, and operational intelligence that goes substantially beyond what has historically been asked of the people who build and run data centers. The infrastructure that rises to that challenge will underpin the next decade of AI development. ## Sign up for AIxEnergy AIxEnergy helps leaders understand the physical economy of artificial intelligence by connecting energy, infrastructure, markets, policy, and governance into clear strategic insight. Subscribe Email sent! Check your inbox to complete your signup. No spam. Unsubscribe anytime.