PaperlayerView exact revision on arXiv

Source-linked AI summary

From Grid to Chip: Power Architecture, Stability, and Flexibility of AI Data Centers

Yubo Song, Rui Kong, Takuro Umihara, Pooya Davari, Frede Blaabjerg, Subham Sahoo

arXiv:2609.11649v1cs.ETcs.ARcs.SYeess.SY

TL;DR

AI data centers are dynamic electrical systems whose reliability and deployment depend on interactions across computing, power delivery, buffering, and grid operation. This article synthesizes grid integration, power architectures, and stability into a grid-to-chip framework, highlighting deployment constraints and conditional flexibility as central features of scalable infrastructure.

  • Problem

    AI data-center topics are often treated across separate physical and institutional boundaries, although workload, power delivery, buffering, protection, and grid conditions jointly shape electrical behavior and deployment.

    Source: §I · p. 1

  • Method

    The article presents a grid-to-chip perspective linking grid integration, power-electronics architecture, workload flexibility, and a three-level stability framework.

    Source: §I · p. 1, 1) Stability Modeling · p. 16

  • Results

    The synthesis identifies interconnection capacity, equipment lead times, and evolving grid codes as deployment constraints, while workload orchestration, cooling, on-site resources, and energy buffers offer conditional spatio-temporal flexibility.

    Source: §VI · p. 23

  • Takeaways & Limitations

    The article concludes that grid-to-chip co-design offers a pathway for scaling AI infrastructure as a controllable, reliable, and efficient grid-supportive resource.

    Source: §VI · pp. 23–24

Abstract

from arXiv · show

The rapid growth of artificial intelligence (AI) computing is transforming data centers into large, dynamic electrical loads. Their deployment is primarily constrained by energy availability and grid-connection capacity, which is further aggravated by the ability of power-delivery architectures, control systems, and computing workloads to operate reliably during fast grid disturbances. This article presents a technological perspective on AI data centers as grid-interactive computing systems. First, it reviews grid-integration bottlenecks, evolving connection policies, grid-code requirements, which has fostered new technological trends via spatio-temporal flexibility available through workload orchestration, cooling systems, on-site resources, and energy storage. Second, it maps the evolution of power-delivery architectures from medium-voltage grid interfaces to chip-level, discussing higher-voltage DC distribution, solid-state transformers, wide-bandgap devices, advanced chip-level power delivery, and liquid cooling. Third, it establishes a three-level stability framework spanning rack-level DC-bus dynamics, facility-level converter interactions, and system-level grid-coupled behavior. The framework connects dominant instability mechanisms, including constant power load effects, impedance interactions, forced oscillations, and operating-mode transitions, with suitable modeling, assessment, and mitigation approaches. Synthesizing these topics, this article highlights grid-to-chip co-design as a central requirement for scalable AI infrastructure, linking computing workloads, power-delivery systems, energy buffers, and grid operation.

I. INTRODUCTION

AI data centers are dynamic, power-electronics-dominated loads whose deployment faces grid-capacity and infrastructure bottlenecks. The article frames them as grid-interactive systems requiring coordinated attention to demand, flexibility, and reliability across grid-to-chip boundaries.

I. INTRODUCTION

  • Grid-side integration must be assessed alongside computing loads, converters, energy buffers, protection, and the external grid because interactions can produce oscillations and disturbance-driven disconnections or restoration excursions.These dynamics span multiple physical boundaries and timescales, making individual server, UPS, or grid-asset availability insufficient as a reliability assessment.Source: §I · p. 1
  • The article connects grid integration, power-electronics architecture, workload flexibility, and dynamic stability in a unified grid-to-chip perspective.Its review covers grid bottlenecks and flexibility, grid-to-chip power delivery, and a three-level stability framework.Source: §I · p. 1

1) Surging Data Center Demand:

  • Data-center electricity use was 416 TWh in 2024 and is projected to reach 946 TWh by 2030 and 1193 TWh by 2035 under the IEA base-case scenario.The accompanying figure tracks AI-training computational resources, global data-center electricity demand, and U.S. average electricity prices.Source: 1) Surging Data Center Demand · p. 2, Figure 1 · p. 2
  • Operational data centers increased U.S. wholesale electricity prices by 2% to 6% on average between 2021 and 2025, with larger effects in regions hosting major data-center corridors.The passage attributes 2022–2024 inflation mainly to natural-gas price spikes and says data-center connection capacity partly contributed to rate increases since 2024.Source: 1) Surging Data Center Demand · p. 2

2) Bottlenecks of Data Center Integration:

  • Grid operations typically take 5 to 10 years compared with 18 to 36 months for data-center construction; Amsterdam and Tokyo have 10-year lead times, and Amsterdam paused projects over capacity.Power-cable and large-transformer lead times approximately doubled from 2021 to 2024, and infrastructure can lag a completed data center.Source: 2) Bottlenecks of Data Center… · pp. 2–3
  • Some operators have replaced first-come, first-served queues with first-ready, first-served prioritization based on project feasibility.The queue process can extend lead times, while early applications under the former approach encouraged operators to submit before plans were ready.Source: 2) Bottlenecks of Data Center… · p. 3
  • Accounting for flexibility that curtails 0.25% of annual peak demand could integrate 76 GW of new U.S. load, equivalent to 10% of current aggregate peak demand.Capacity calculations based on maximum annual peak demand omit demand-side flexibility; grid expansion also requires time for land, materials, and impact analyses.Source: 2) Bottlenecks of Data Center… · p. 3

B. Power Systems Challenges and Opportunities on Data Center Integration.

Grid-connection rules and policies are changing as regions manage large-load applications and technical requirements. The reviewed approaches include queue reforms, demand response, and connection requirements that vary across systems.

1) Grid Code Regulations:

  • Nine representative transmission system operators are compared, and the status and specific content of grid-code revisions vary by region.The comparison covers large-capacity load connections across North America, Europe, and Japan.Source: 1) Grid Code Regulations · p. 3
  • First-ready, first-served policies prioritize feasible projects over application order, while the Netherlands prioritizes connections by social importance.The passage reports first-ready policies in the UK, Germany, and Texas, and a Dutch rule covering objectives such as housing, hospitals, schools, and national defense.Source: 1) Grid Code Regulations · pp. 3–4
  • Screening can identify multiple applications that congest connection queues, while Japan applies a penalty to demand over 30 MW when contracted capacity is not finalized as planned.The penalty passage describes review of contracted capacity after a customer continues for a certain period without setting the final contracted amount.Source: 1) Grid Code Regulations · p. 4
  • PJM has planned Connect and Manage options for projects bringing generation and for large demand projects without new power plants; EirGrid requires on-site or proximate generation for data centers of 10 MVA or more.The PJM options are described as ways to connect without waiting for grid upgrades in the specified cases.Source: 1) Grid Code Regulations · p. 4
  • Grid codes address fault ride-through, power quality, and ramping, with AESO limiting ramping to 10 MW/min and Energinet allowing up to 60 MW/min.Requirements vary with power-system capacity and other conditions.Source: 1) Grid Code Regulations · p. 4
  • Regions have introduced or are considering demand response through mandatory curtailment for non-firm connections or voluntary program participation.The passage says such mechanisms exist or are under consideration in all regions compared.Source: 1) Grid Code Regulations · p. 5

2) Unlocking Flexibility Capability of Data Centers by Components:

Data-center flexibility can come from workload scheduling and physical assets, but its availability depends on task requirements and service-level agreements. The reviewed mechanisms address power fluctuations, congestion, and shifting demand across time or locations.

2) Unlocking Flexibility Capability of Data Centers by Components:

  • Workload orchestration can shift deferrable tasks across time or allocate them to geographically distant data centers, subject to task type and service-level agreements.Higher SLA requirements limit flexibility in maintaining service levels.Source: 2) Unlocking Flexibility Capability of… · p. 5
  • Thousands of GPUs operating in lockstep can create grid-level power fluctuations, so data centers need power smoothing and may shift or geographically reallocate workloads during congestion.The passage identifies later execution as temporal flexibility and transfer to another data center as spatial flexibility.Source: 2) Unlocking Flexibility Capability of… · p. 5
  • Flexibility assets include on-site generation, cooling, IT loads, and UPS batteries; thermal energy storage can shift capacity over hours, while UPS batteries provide instantaneous backup power.Thermal energy storage responds over minutes to hours and is therefore not suitable for frequency-regulation services.Source: 2) Unlocking Flexibility Capability of… · pp. 5–6
  • Workload orchestration on a 130 kW production-scale GPU cluster supported real-time dispatch, sustained curtailment, carbon-aware computation, and geographic load shifting while maintaining service levels for priority AI jobs.The reported use cases included AI priority training and inference jobs.Source: 2) Unlocking Flexibility Capability of… · p. 6
  • An illustrative 60 MW profile reduces demand through workload allocation from 6:00 to 8:00 and through on-site BESS discharge from 18:00 to 21:00.The suitable flexibility source varies with event timing, SLA constraints, task importance, and the availability of other data centers.Source: 2) Unlocking Flexibility Capability of… · p. 6, Figure 4 · p. 6

3) Unlocking Spatial Flexibility Capability of Data Centers over Regions:

The section connects data-center workload flexibility to grid needs across regions and task types, while relating these opportunities to policy and power-delivery constraints. It also maps the grid-to-chip voltage hierarchy and the architectural shifts responding to rising rack power.

3) Unlocking Spatial Flexibility Capability of Data Centers over Regions:

  • Spatial workload allocation across synchronous systems can ease supply-demand imbalances that are difficult to address within a single system, but latency and privacy constraints limit which tasks can move.The section discusses virtual interconnectors that route computing work through fiber networks rather than transferring electricity over power lines.Source: 3) Unlocking Spatial Flexibility… · pp. 6–7

4) Incentivization of Flexibility Utilization from Data Center:

  • Data-center flexibility varies by task: interruptible pre-training has less than one-tenth the VoCL of on-demand SLA-bound tasks, supporting differentiated incentives such as price hedging by VoCL tier.PJM and European TSOs are considering connect-and-manage approaches for large demands.Source: 4) Incentivization of Flexibility… · pp. 7–8

C. Observations and Recommendations

  • Grid-integration responses include longer power-system operating timelines than data-center construction, evolving grid codes, and demand adjustment through non-firm curtailment or voluntary demand response.The cited regional trends include greater hosting-capacity transparency in Europe and discussion of data centers bringing on-site power in North America.Source: C. Observations and Recommendations · p. 8
  • Data centers can smooth instantaneous fluctuations, shift workloads over time, and reallocate them geographically, with flexibility shaped by task priorities, service agreements, and time of day.On-site resources, cooling systems, and IT workloads are identified as flexibility resources; spatial flexibility may address both local congestion and imbalances across synchronous systems.Source: C. Observations and Recommendations · p. 8

A. From Grid Boundary to Chips: A Cross-Scale Hierarchy

  • The power path steps voltage down from medium- or low-voltage AC toward extra-low-voltage rack buses, while rising rack power makes conventional 48 V distribution increasingly challenging.An example cited for 800 V DC rather than 415 V AC reports 45% lower copper requirements; the shift also raises the conversion ratio relative to 48 V systems.Source: A. From Grid Boundary to Chips: A… · pp. 8–9

B. Architectures of Power Electronics Converters

Power-electronic converter stages perform conversion, voltage stepping, and distribution throughout the data-center power path. Their architectures balance voltage and current stress, modularity, bidirectional operation, and the needs of backup systems.

B. Architectures of Power Electronics Converters

  • AC-DC stages commonly use two-level bridges or Vienna rectifiers, while DAB or resonant DC-DC converters are common for PDU and UPS battery interfaces where bidirectional flow is preferred.MMC and CHB structures cascade lower-voltage modules at medium-voltage boundaries; multi-winding transformers can distribute load-side power.Source: B. Architectures of Power Electronics… · p. 9, Figure 6 · p. 10, Figure 7 · p. 10
  • Converter cells distribute voltage stress in series at medium-voltage buses and current stress in parallel at low-voltage buses, enabling modular structures but adding insulation, maintenance, and space costs.The modularization penalty arises because each isolation barrier may still face the full operating-voltage rating.Source: B. Architectures of Power Electronics… · p. 9
  • Backup battery units interface with AC or DC server buses through power-electronic UPS systems, and bidirectional UPSs can buffer interactions between utility grids and AI servers.The passage cautions against relying simply on oversized storage and points instead to flexibility from xPU load variation between near-TDP and near-idle phases.Source: Figure 8 · p. 10, B. Architectures of Power Electronics… · p. 10
  • The architecture map extends observations across power-supply hierarchy levels, including emerging 800 V DC-class solutions, rather than describing only one architecture.The comparison is intended as a reference for the relevant hierarchy levels and converter stages.Source: B. Architectures of Power Electronics… · p. 10

C. Evolution of Power Electronics Devices

Power-device choices and packaging support higher switching performance and power density, while chip-level delivery and cooling address constraints as AI rack power grows. The section presents these choices as interdependent parts of data-center design.

C. Evolution of Power Electronics Devices

  • Infineon’s .XT diffusion-soldering interconnect is reported to reduce junction-to-case thermal resistance by up to 25%, supporting improved thermal performance and higher power density.Low-inductance packaging is particularly significant for WBG and UWBG devices with high di/dt and dv/dt.Source: C. Evolution of Power Electronics… · p. 11

D. Chip-Level Power Supply Innovations and Trends

  • BPD moves power interconnects to the wafer backside to free front-side routing and reduce IR drop, while VPD places converters beneath the die to shorten power paths and free board area.BPD and VPD address different segments and can be combined; BPD still requires external converters, and embedded passive components remain a challenge for VPD.Source: D. Chip-Level Power Supply Innovations… · pp. 11–12

E. Development of Cooling Design

  • Cooling is a co-design variable because it affects chip and power-electronics performance and reliability, and its substantial, flexible facility load can complement workload and storage balancing.Cooling may consume 10–20% of total system power; liquid cooling and workload scheduling are discussed as potential demand-response resources.Source: E. Development of Cooling Design · p. 12

IV. STABILITY OF DATA CENTER SYSTEMS

Data-center stability spans internal power-delivery dynamics and interactions with the external grid, with reported oscillations and load changes illustrating the stakes. The section organizes assessment around instability categories and three analytical levels linked by bidirectional dynamics.

IV. STABILITY OF DATA CENTER SYSTEMS

  • Reported events include oscillations at 14.7 Hz, approximately 23 Hz, and around 11 Hz, while a 387 MW load reduction in Ireland raised frequency to 50.341 Hz with approximately +0.26 Hz/s RoCoF.These events motivate data-center stability models, control-interaction studies, and fault ride-through coordination.Source: §IV · pp. 12–13, Table VI · p. 13

1) Scope of Instability Issues:

  • The framework groups issues into small-signal instability, large-signal or transient instability, forced oscillations, and grid-side impacts from aggregate load transitions.Fig. 10 identifies small- and large-signal stability, forced oscillation, and grid-interactive impacts as the principal categories.Source: 1) Scope of Instability Issues · p. 13, Figure 10 · p. 14

2) Stability-Oriented Three-Level Framework:

  • Three analytical tiers retain different interactions: rack-level source-load impedance and constant-power-like behavior, facility-level converter-network interactions, and system-level grid coupling at the PCC.The appropriate tier depends on disturbance origin, dominant interaction, and the variables used to judge stability.Source: 2) Stability-Oriented Three-Level… · pp. 13–14
  • The tiers are study boundaries rather than fixed architectures; in DC-distributed systems, direct connection to a common facility bus can make rack- and facility-level analyses overlap.Workload changes propagate outward from racks, while grid disturbances propagate back into data centers.Source: 2) Stability-Oriented Three-Level… · p. 14

B. Instability Mechanisms

Rack stability depends on interactions among regulated IT loads, finite stored energy, and supply controls; at facility scale, converters coupled through shared networks can create oscillations and disturbance-driven voltage excursions.

1) Rack-Level DC-Bus Stability:

  • Within downstream converters’ regulating bandwidth, constant-power-like IT loads draw more current as bus voltage falls, reinforcing voltage depression and reducing damping.Above that bandwidth, terminal behavior depends on the PoL converter’s internal dynamics.Source: 1) Rack-Level DC-Bus Stability · p. 14
  • Large workload steps or severe voltage disturbances can drive a locally small-signal-stable rack into deep DC-link voltage excursions or collapse when source current limits prevent rapid energy replenishment.The DC-link capacitor temporarily supplies the power imbalance.Source: 1) Rack-Level DC-Bus Stability · p. 14
  • Correlated workload transitions can trigger large bus excursions or sustained forced oscillations, and fluctuations across many racks can propagate through PSUs into facility distribution.Slow workload changes instead shift the operating point and stability margin.Source: 1) Rack-Level DC-Bus Stability · p. 15

2) Facility-Level Distribution Stability:

  • At facility scale, unfavorable source–load impedance magnitude and phase near network or filter resonances can cause oscillation; grid dips, workload changes, and UPS transfers can also produce bus excursions and transient oscillations.Low-frequency modes often involve DC-voltage regulation or constant-power behavior; higher-frequency resonances can arise from input filters or digital delays.Source: 2) Facility-Level Distribution Stability · p. 15
  • A shared bus couples rack converters: branch disturbances can feed back through other converters, while AC phase imbalance or DC cross-rack coupling can create modes hidden by aggregate or balanced models.DC distribution reduces the AC phase-loading effects but strengthens direct interactions among racks on a common DC network.Source: 2) Facility-Level Distribution Stability · p. 15

3) System-Level Interactive Stability:

Data-center stability extends to bidirectional coupling with the external grid, while modeling must be matched to the phenomenon and timescale. The paper relates workload and electrical representations to small- and large-signal models, forced-oscillation analysis, and grid-interactive models.

3) System-Level Interactive Stability:

  • Grid impedance and aggregated data-center input behavior can couple facility controls to external grid modes, while periodic workloads can excite and selectively amplify weakly damped facility or grid modes.During grid disturbances, current limits, persistent constant-power demand, insufficient stored energy, or unsuitable mode transitions can contribute to voltage excursions and ride-through failure; trips and synchronized restoration can also disturb grid frequency and voltage.Source: 3) System-Level Interactive Stability · pp. 15–16

C. Stability Modeling and Assessment Methods

1) Stability Modeling:

  • Workload traces or step, pulsed, and periodic profiles represent computational demand, while electrical-load models range from ideal constant-power loads to reduced or averaged converter models.Use averaged converter models when internal control loops materially affect the oscillation of interest; after linearization, state-space and terminal-impedance formulations support small-signal analysis.Source: 1) Stability Modeling · p. 16
  • Large-signal studies use nonlinear averaged models, extended with piecewise or hybrid mode logic when current limits or operating transitions matter.Forced-oscillation studies represent both workload excitation and its transmission path; PCC-oriented models capture continuous power dynamics and discrete grid-disturbance transitions.Source: 1) Stability Modeling · pp. 16–17
  • Detailed EMT or switching models resolve high-frequency effects but limit device count and simulation duration, whereas reduced-order positive-sequence models enable larger phasor-domain studies and should be checked against detailed EMT models.Measurement-based models can fill proprietary-parameter gaps, but their conclusions depend on represented operating conditions and they complement rather than replace physics-based models.Source: 1) Stability Modeling · p. 17

2) Stability Assessment Methods:

Stability assessment methods address different mechanisms and disturbance scales, and enhancement strategies act at workload and converter levels. The discussion emphasizes matching the method or intervention to the instability mechanism.

2) Stability Assessment Methods:

  • Small-signal eigenvalue analysis identifies modal frequency, damping, and participation, while impedance-based methods assess source–load coupling at an electrical interface; both are local to the linearization operating point.Impedance methods are modular, but cannot determine recovery after sufficiently large disturbances.Source: 2) Stability Assessment Methods · p. 17
  • Large-signal trajectory and Lyapunov methods assess whether post-disturbance states remain within a recoverable region, though useful Lyapunov functions are difficult to construct for high-order systems.These analytical methods are most useful for reduced rack- or converter-level models, with detailed EMT simulation often used for facility verification.Source: 2) Stability Assessment Methods · p. 17
  • Forced-oscillation assessment separates workload excitation from the dynamic path that amplifies it, combining workload-to-output frequency response with waveform spectra or modal identification when available.Measurement alone cannot distinguish forced oscillations from small-signal oscillations.Source: 2) Stability Assessment Methods · p. 17
  • Grid-side assessment simulates facility responses such as connection, storage transfer, or disconnection and evaluates resulting grid frequency, voltage, and power-flow metrics.Sudden disconnection may raise frequency and local voltage, whereas rapid reconnection may depress both.Source: 2) Stability Assessment Methods · pp. 17–18

D. Stability Enhancement Strategies

  • A credible assessment combines event-reproducing time-domain simulation with analytical methods that trace behavior to its dominant physical mechanism.Fig. 13 organizes stability-enhancement measures and typical application scenarios across system tiers.Source: 2) Stability Assessment Methods · p. 18, Figure 13 · p. 18

1) Workload-Side Power Smoothing:

  • Workload smoothing uses power capping, processor-frequency adjustment, job scheduling, and temporal staggering to reduce demand ramps and correlated power oscillations; its scope is schedulable variation, not fast converter resonances or grid-fault imbalance.Closed-loop schemes can use facility power, workload status, and grid requests while respecting service requirements.Source: 1) Workload-Side Power Smoothing · p. 18

2) Converter-Oriented Stabilization:

  • Converter stabilization uses active damping to counter negative damping near resonance, with frequency-selective virtual impedance and adaptive tuning addressing trade-offs and shifting resonance frequencies.Nonlinear controllers can shape post-disturbance trajectories to enlarge the region of attraction and help avoid collapse.Source: 2) Converter-Oriented Stabilization · pp. 18–19

3) Energy Buffering and Coordination:

Energy buffers and workload flexibility can be coordinated across timescales, while ride-through and grid-support controls shape the facility’s response at the PCC. The outlook connects evolving computational loads and data-center infrastructure with grid integration.

3) Energy Buffering and Coordination:

  • Local capacitors or backup batteries can absorb sharp transients near racks, while centralized UPS batteries or storage serve more loads over longer intervals, with response shaped by converters and network impedance.UPS batteries should preserve supply continuity rather than smooth sustained deviations that could compromise backup capability.Source: 3) Energy Buffering and Coordination · p. 19
  • Workload scheduling or migration can reshape adjustable fluctuations while UPS storage compensates the residual, subject to state-of-charge and cycling constraints.Recovery charging should be ramped and distributed to avoid a secondary power step; parallel converters also need coordinated current limiting and mode transfers.Source: 3) Energy Buffering and Coordination · p. 19

4) Ride-Through and Grid-Supportive Control:

  • Ride-through control must maintain supply to critical IT loads while avoiding abrupt demand loss at the grid interface; emerging requirements favor remaining connected when equipment and converter limits permit.Conventional UPS protection may transfer to backup after shallow voltage depressions, potentially removing substantial load during a normally cleared fault.Source: 4) Ride-Through and Grid-Supportive… · p. 19
  • During severe sags, controlled active-power reduction, reactive support, and local storage can manage converter current limits; recovery should be staged with DC-link restoration and storage.After ride-through, rectifier reactive-current control and UPS battery modulation can support PCC voltage and provide fast frequency response.Source: 4) Ride-Through and Grid-Supportive… · p. 19

V. OUTLOOK

  • The outlook identifies evolving computational loads, grid architecture, power-information infrastructure convergence, and grid-aligned sustainability as areas for development.It notes that grid-to-data-center consolidation remains to be fully exploited.Source: §V · p. 20

A. Evolving Computational Loads

  • The computational-load discussion focuses on scaling pressures, power supplies for emerging computing, and grid-supportive potential from flexibility and energy storage.It frames these issues as infrastructural impacts of load evolution across data-center hierarchies.Source: A. Evolving Computational Loads · p. 20

1) Scaling of computational loads:

  • Hyperscale data centers had lower PUE values, and between 2010 and 2018 computing power rose 550% while global data-center power consumption rose 6%.The text also points to multi-data-center training as a response to physical and power-system constraints on large-scale facilities.Source: 1) Scaling of computational loads · p. 20

2) Power supply for leading-edge and emerging computing:

Power supply for emerging computing must accommodate unusual operating conditions while grid-interactive data centers balance flexibility with service and stability requirements. The section connects evolving grid arrangements and fault-ride-through rules to these needs.

2) Power supply for leading-edge and emerging computing:

  • Quantum computers’ energy requirements remain unconfirmed, while operation below 1 K and stringent noise and electromagnetic conditions challenge power-supply design.Other emerging computing paradigms may also impose electromagnetic, mechanical, or thermal demands.Source: 2) Power supply for leading-edge and… · p. 20

3) Grid-interactive data centers:

  • Coordinated workload flexibility, cooling thermal inertia, and storage could let data centers regulate PCC power like virtual power plants, provided service and stability are protected.Converter-embedded sensing could support real-time adaptive damping or power support, subject to reliable measurement, cybersecurity, and standardized service rules.Source: 3) Grid-interactive data centers · p. 20

B. Evolving Grid Architectures for Data Centers

  • As grids move toward distributed operation, lower inertia, and more varied interactions, data centers face evolving functional and standardization requirements.The SAFEr Grid and fault-ride-through requirements are examples of this evolution.Source: B. Evolving Grid Architectures for… · p. 20

1) The SAFEr Grid infrastructure:

  • Data centers’ power-electronics-based, frequency-agnostic supply chains and MV-SST boundaries with backup batteries make them a near-ready proving ground for SAFEr Grid energy routing.Realization still requires internal hierarchies to reconcile data-center load functions with societal energy-routing resilience commitments.Source: 1) The SAFEr Grid infrastructure · pp. 20–21, Figure 14 · p. 21

2) Grid specifications of data centers:

  • Data-center grid specifications should define permissible load reduction, current priority, storage-supported duration, and post-fault recovery while accounting for facility architecture.The requirements should preserve facility security without letting transfer or recovery logic amplify the initiating disturbance.Source: 2) Grid specifications of data centers · p. 21

3) Requirements on FRT capability:

  • Energinet, EirGrid, and AESO require connection maintenance for 150 ms at 0% voltage; RTE specifies 150 ms at 5% voltage.FRT requirements are being applied or proposed across several transmission operators, with other codes specifying stepped connection-voltage requirements over time.Source: 3) Requirements on FRT capability · p. 21, Figure 15 · p. 21

4) Responsible parties for data center integration:

Data-center integration requires transparent capacity information, updated grid requirements, and operator attention to congestion and carbon-aware operations. The section also considers how power and information infrastructures can share channels, while creating new cybersecurity concerns.

4) Responsible parties for data center integration:

  • A Dublin single-phase fault disconnected 480 MW, or 52% of the country’s data-center demand, illustrating the stability risk from FRT-incompatible demand.Grid operators are also urged to disclose hosting capacity and improve queue management; developers and operators need to address congestion and carbon-aware operations.Source: 4) Responsible parties for data center… · p. 21

C. Convergent Power-Information Infrastructures

1) Cybersecurity:

  • Cyberattacks on grid or converter interfaces can propagate at rack and system levels, potentially crossing stability boundaries; abnormal data-center behavior can also threaten local-grid resilience.Source: 1) Cybersecurity · p. 22

2) Power line data transmission:

  • Power-line data transfer can use the already instrumented and digitally controlled power chain as a low-incremental-cost information channel, adding resilience to communication-network outages.Source: 2) Power line data transmission · p. 22

3) Local power-over-fiber/photonic delivery:

  • Power-over-fiber addresses copper-volume and electromagnetic-isolation challenges, while photonic interconnects target electrical-link latency and bandwidth limits.Both remain maturing candidates for settings where power or information density is a first-order design requirement.Source: 3) Local power-over-fiber/photonic… · p. 22

4) Physical/analog computing concept:

The article surveys physical/analog computing and sustainability directions alongside carbon-aware workload allocation. It concludes that grid-to-chip co-design should coordinate power behavior, reliability, flexibility, and carbon awareness as AI infrastructure scales.

4) Physical/analog computing concept:

  • Power-electronics network dynamics can perform affine transformations on encoded perturbations alongside power flow, suggesting a physical computing substrate.The concept could support computing-load outsourcing or workload and power-delivery coordination, with proposed efficiency and sustainability benefits.Source: 4) Physical/analog computing concept · p. 22

D. Sustainable Data Centers in Power Grids

  • Data-center sustainability efforts include carbon-aware computing and equipment measures, while European operators have committed to carbon-free energy, water conservation, and waste-heat reuse.The European initiative targets carbon neutrality by 2030 and includes more than 100 operators and trade associations.Source: D. Sustainable Data Centers in Power… · pp. 22–23

1) Carbon footprint of computing tasks and carbon-aware computing:

  • In simulations across three AI data-center parks, scheduling reduced operating costs by at least 1.69% and carbon emissions by 4.5% versus four alternative schemes.The scheduling considered continuity-constrained workloads and maintained computational tractability; Google’s CICS reduced average cluster power by 1–2% during the highest-carbon hour.Source: 1) Carbon footprint of computing tasks… · p. 23
  • Carbon intensity varies across time and location; shifting nighttime workloads to regions experiencing daytime can reduce computing-related emissions.Source: 1) Carbon footprint of computing tasks… · p. 23

2) Cost allocation for carbon-aware computing:

  • Carbon-aware cost allocation differs by data-center type: cloud and enterprise owners can manage emissions, while colocation operators may allocate costs through tenant fees or voluntary bidding.Source: 2) Cost allocation for carbon-aware… · p. 23

VI. CONCLUSION

  • Stability spans rack DC-bus workload and constant-power-load effects, facility converter interactions and mode transitions, and grid-level ride-through, transfer, disconnection, and restoration.The conclusion recommends coordinating compute scheduling, energy buffering, converter control, and reliability management to define a predictable power envelope at the PCC.Source: §VI · pp. 23–24
Loading 2609.11649v1…