Managing Grid Stability for High Computing Demands

🛜 Data centres can employ a radial power distribution network based on the Open Compute Project (OCP) architecture proposed by Meta engineers. This system begins with the utility grid (Zg), stepping down voltage through substation transformers (PTX) to medium voltage and then further to low voltage (typically 480 V) via transformers that feed main switchboards (MSB). From the MSBs, power is distributed through several layers of smaller switchboards (SB) and reactive power panels (RPP) down to the server racks containing power shelves and finally the power supply units (PSUs). A key characteristic of this architecture is the direct powering of most server racks from the utility grid through a 480 V distribution network, limiting the use of AC UPS for these racks. The PSUs within a power shelf can be configured either in a star (Y) or a delta (∆) connection.   ⚡ The motivation for studying instability in AC data centres stems from actual resonance events observed in Meta data centres, including both low-frequency (e.g., 11 Hz with harmonics at 49 Hz and 71 Hz) and high-frequency (5-10 kHz) resonances. The root causes of these instabilities are often linked to negative damping in the converter's input impedance, occurring near the fundamental frequency due to DC bus voltage control and at high frequencies due to control delays. In the context of data centres, the "weak grid" environment conducive to instability is frequently created by the impedance of the internal power distribution system rather than the external utility grid (which is typically very strong). The work introduces the concepts of common-mode (CM) resonance, involving the entire distribution network and typically manifesting at low frequencies due to high equivalent source impedance, and differential-mode (DM) resonance, occurring between different groups of PSUs or loads through the network between them and often observed at high frequencies. For Y-connected PSUs, the zero sequence is identified as the weakest link and the first to become unstable. 💡 To ensure system stability, Meta has developed impedance-based design specifications for PSUs and UPS (input impedance). The specifications are based on sufficient conditions derived from the Nyquist stability criterion. These specifications typically comprise three main parts: requirements for the magnitude response below twice the fundamental frequency (2f1) to limit impedance dipping associated with DC bus voltage control; criteria for operation with a defined source impedance, expressed as a minimal Base Impedance Ratio (BIR) to ensure stable operation under realistic network conditions and account for coupled current effects; and requirements for positive damping above 2f1 to mitigate high-frequency resonances related to control delays. A minimal base impedance ratio is also defined for the overall data centre distribution system. #datacenter #ai #infrastructure #powerelectronics #stability #cleanenergy #renewables

The most dangerous moment for the grid is not when a data centre connects. It’s when it suddenly disappears. That is one of the most important signals in NERC’s latest reliability guideline on emerging large loads. The report shows a scenario where multiple generators lose synchronism following a large load trip and rapid recovery. That should fundamentally change how we study large loads. We’ve spent years modelling large loads as demand. But under certain conditions, sudden disconnection does not just create imbalance. It triggers system-wide response. ➤ It changes power flows. ➤ It changes voltage conditions. ➤ It changes what nearby generators have to survive. The system is not most stressed when the load is connected. It can become more stressed when it suddenly disappears. That is the hidden risk. In practice, this is not always a single-site event. It can be the near-simultaneous disconnection of multiple load clusters during relatively minor disturbances, followed by uncoordinated recovery. A sudden data centre trip can leave nearby generation trying to export power through a network that may not have enough transfer capability, system strength, or damping at that moment. This is not a load problem. It is a stability problem. It can drive: • large angle swings • unstable power flows • loss of synchronism • protection operation • cascading risk And this is not a modelling detail. This is not just a stability detail. It can decide whether a project connects, waits, or triggers new system limits. It affects: • contingency definitions • transfer limits • connection assessments • system strength requirements Because a “load” can now behave like a destabilising disturbance. The connection question is no longer: “How many MW?” It is: “How does it behave during and after a disturbance?” We are no longer dealing with passive demand. We are dealing with dynamic, power-electronic, system-interacting assets. My view: We did not design the grid for loads that behave like contingencies. But that is exactly what we are now connecting. Large loads are now part of the stability problem. If designed, modelled, and coordinated properly, they can also become part of the solution. The real question is: Are we testing the load connection, or the system that has to survive it? Are others seeing this scenario appear in large load connection studies? #DataCenters #GridStability #PowerSystems #LargeLoads #NERC #SystemStrength #TransmissionPlanning #IBR #AngularStability

AI data centers are becoming grid assets — not just loads. Utilities are tightening requirements faster than developers can adapt. The next wave of hyperscale development will require a hybrid grid-support stack just to achieve rapid interconnection. “The hyperscale campus of the future will bring its own inertia, VAR stability, and ramp control.” ⚡️ The New Grid Reality for Hyperscale AI-scale campuses (100–500 MW, 80–200 kW/rack) no longer behave like traditional IT loads. They generate fast ramps, sub-second variability, harmonics, and voltage sensitivity. In many nodes, this looks less like a “typical customer” and more like a converter-dominated industrial plant. Utilities and TSOs are already responding with stricter technical requirements: • Tighter Power Quality (PQ) limits (harmonics, flicker, voltage deviations) • EMT modelling (sub-cycle electromagnetic transient analysis) • Ramp-rate caps (MW/min load-change limits) • VAR obligations at the Point of Common Coupling (PCC) (reactive-power performance) The bar is rising fast. Here’s how the industry is adapting: 1️⃣ STATCOMs — the Core of Modern VAR & PQ Performance STATCOMs are becoming essential for AI-ready campuses: • Millisecond reactive-power response • Voltage stabilization on weak nodes • Flicker and harmonic mitigation • Dynamic support during rapid load changes Hybrid angle: Many deployments now integrate STATCOM + BESS under one coordinated control layer. 2️⃣ BESS — From Backup System to Ramp-Shaping Engine Battery Energy Storage Systems are evolving into strategic grid assets. They can: • Cap MW/min ramps • Smooth sub-second GPU variability • Support fault-ride-through requirements • Reshape AI load curves for grid compatibility Impact: A 200 MW AI cluster becomes significantly easier for utilities to manage. 3️⃣ Synchronous Condensers — Inertia & Short-Circuit Strength In weak or inverter-dominated grids, synchronous condensers provide: • Real inertia • Higher short-circuit strength (SCR) • Improved transient and angle stability • Reduced FIDVR risk In practice: bringing your own short-circuit power to the PCC. 📌 Implications for Developers & Investors ➡️ Interconnection packages are shifting. Expect utilities to require hybrid systems, especially where SCR is low. ➡️ Faster time-to-energization. Stronger grid-support design reduces system risk, accelerates studies, and improves negotiation leverage. ➡️ Delays are expensive. Months of delay on a 300–500 MW AI campus carry enormous financial consequences. Hybrid VAR, inertia, and ramp-shaping solutions buy time — and time is value. #DataCenters #GridStability #STATCOM #BESS #SynchronousCondenser #Hyperscale #PowerQuality #EnergySystems #AIInfrastructure #Interconnection

For energy, data center, and operations leaders planning gigawatt-scale AI loads, compute is no longer the gating item.  Clean, steady power is.    The play that holds up is two tracks run in parallel.  Build new supply, and get more usable capacity out of the grid you already have.    Supply example: Siemens is partnering with Commonwealth Fusion Systems.  That work spans complex part design, factory automation (PLCs), and plant control where controllers manage extreme currents and cooling to keep the plasma stable. Spark is the first step. Arc is planned in Virginia at 400 MW, sized for an entire data center class footprint.    Grid example: operators are using real-time network models and automation to raise transfer capacity by about 20 percent in some deployments without new wires. They can forecast load, test scenarios like adding 10,000 EVs to a neighborhood, and coordinate buildings to reduce demand briefly (for example, turning down air conditioning) to ride through peaks.    What to do next: tie your AI buildout plan to an energy plan with two tracks and one owner.    Near term  - Instrument demand at the feeder and facility level so you can forecast and shed load intentionally.  - Secure curtailment agreements with clear triggers, durations, and financial ownership.  - Deploy grid-aware controls that act in seconds, not meetings.    Mid to long term  - Align siting and interconnect timing with Arc-class supply coming online.  - Design electrical and cooling systems for high-current, high-cooling regimes from day one.  - Build the operating cadence now, including who approves trade-offs when power is constrained.    Treat clean, firm power as a core dependency, or the AI roadmap becomes fiction. 

🔥 #AI #datacenters are being treated like “just another big load.” That’s a dangerous planning assumption. Most of the power they draw isn’t flexible by default - it’s reliability-driven and must stay on to keep compute running. Backup systems aren’t demand response, they are continuity systems. And GPUs don’t pull smooth power - they fluctuate in ways grids were not designed for. But here’s where the story has potential to shift 👇 📌 Batteries and energy storage aren’t just backup anymore - they can make large power users behave like flexible grid assets. With the right controls, storage can charge when the grid is abundant and discharge when it’s stressed, helping balance supply and demand and support frequency and stability, all while keeping compute running. This is backed by recent grid research on dispatch and optimal BESS use. 📌 Growing work on grid-interactive UPS and storage systems shows that data centers can participate in ancillary markets and provide services like fast frequency response and other flexibility if designed and governed with that intent. In #Europe, this is already moving from theory to planning reality ⚡ Reports show that grid congestion and connection constraints are now influencing where data centres are built, with utilities reassessing connection rules and flexibility incentives as grid capacity becomes a decisive factor in investment decisions. So the real shift isn’t debating whether AI loads are “flexible” - it’s about engineering them to be grid-interactive assets, not inflexible liabilities. 👉 If we plan for them as firm loads PLUS intentional, contracted flexibility, we unlock new options for reliability, carbon goals, and grid stability. This isn’t future talk - credible research and emerging deployments are already pointing toward hybrid storage, smarter dispatch, and real grid value. In our work where we help design control centers of the future for utilities and system operators, this needs to be part of the discussion. https://lnkd.in/dXDvf3BE https://lnkd.in/dnMXjKzh ⚡ #GridPlanning #AIInfrastructure #DataCenters #EnergyStorage #BESS #GridFlexibility #EnergyTransition photo: Interactive map of data centre hubs alongside associated power and digital infrastructure // IEA's Energy and AI Observatory

One of the most important messages in North American Electric Reliability Corporation (NERC) new guideline on emerging large loads is that reliability risks are no longer coming solely from the supply side of the grid. For decades, system planners assumed that load was relatively predictable and passive. That assumption is breaking down. AI data centres, hyperscale computing facilities and other large power-electronic loads can add gigawatts of demand in a short timeframe, respond differently to disturbances, and create system impacts that traditional planning processes were never designed to address. What stands out in the guideline is the emphasis on: - Early engagement between load developers and grid operators - Better dynamic modelling of load behaviour - Enhanced system strength and stability assessments - Clear operational and communication protocols - Continuous monitoring after connection The lesson is broader than data centres. As grids become increasingly dominated by inverter-based resources on both the supply and demand sides, reliability will depend less on simply adding capacity and more on understanding system behaviour. The future grid needs both megawatts and physics. Ignoring either one is a reliability risk. #electricity #energy #energypolicy #energytransition

One of the least understood aspects of the AI-data-center boom is not the size of the load … it’s the volatility of the load. Utilities have historically treated data centers as large but relatively flat demand sources — closer to steady industrial load than to highly dynamic systems. AI changes that. Why? Because frontier AI clusters may involve tens of thousands of GPUs operating in synchronized computational cycles. Instead of millions of independent computing tasks smoothing each other out, you increasingly get giant clusters behaving almost like a single machine. That means power demand can ramp sharply — and quickly. And the volatility doesn’t stop with the chips. When GPU utilization spikes, heat spikes, cooling systems ramp, pumps and chillers respond, and power electronics react. At very large scale, those coupled swings can become significant grid events. A 1 GW AI campus experiencing a rapid 10% load swing means a 100 MW change in demand. That is utility-scale generation territory. And unlike traditional utility planning assumptions, these changes may occur in seconds — or subseconds — rather than over hours. This matters because the grid was largely designed around gradual load ramps, predictable industrial demand and hourly planning models. AI infrastructure may require a different architecture: 1) onsite batteries for power smoothing; 2) advanced inverter systems; 3) sophisticated reactive power management; 4) grid-aware workload scheduling and 5) new interconnection standards. Ironically, the future AI campus may look less like a passive customer and more like a miniature grid operator. The next era of grid planning may not just be about adding more power. It may be about managing a fundamentally different kind of load.

Sixty data centers. 1.5GW. Gone in seconds. In July 2024, a lightning arrestor failed on a 230-kV transmission line in #NorthernVirginia. Home to the world’s largest concentration of data centers. In milliseconds, 60 hyperscale facilities dropped off the grid and switched to backup diesel and batteries. PJM, the regional grid operator, had to throttle power plants across two states to prevent a blackout. It worked. Barely. But it exposed something deeper: the U.S. grid wasn’t built for the physics of AI-scale demand. • Many data centers share the same transmission backbone. • Their protection systems act in unison. • And much of the infrastructure dates back to the 1970s. Ironically, the systems built to preserve uptime nearly collapsed the grid that powers them. For investors, this changed everything: Reliability is no longer a utility problem it’s a valuation problem. The smartest operators are already adapting: • Building private substations and dual-feed connections • Adding behind-the-meter microgrids and batteries • Using grid-interactive UPS systems that earn money supporting voltage stability AI’s next bottleneck isn’t compute it’s power. And in the market that powers intelligence, reliability is alpha. Read the full breakdown below #datacenters

DC grids exist to distribute power. In data center, industrial, aerospace, and marine applications, many connected loads impose a constant power rather than a constant current. This has major implications for DC grid voltage behavior and stability. A positive, constant power load (CPL) exhibits a negative incremental (small-signal) resistance. As load power increases, this effective resistance decreases, amplifying current excursions in response to voltage perturbations. When the DC grid voltage is regulated using a voltage-mode PI controller, the negative incremental resistance of the load directly counteracts the proportional action. The impact can be captured using an equivalent circuit representation: the PI controller can be modeled as a resistor (proportional term) in parallel with an inductor (integral term). The CPL’s negative resistance appears in parallel with the proportional resistor, reducing the net conductance, and hence the net loop gain. As a result, the closed-loop voltage dynamics becomes load dependent, with overshoots increasing with load power and oscillations slower to damp down. An alternative is to regulate stored energy rather than voltage. In an energy-based framework, the PI controller controls the amount of stored energy, issuing power commands to achieve its objective. This approach inherently accommodates constant power loads and significantly reduces the dependence of grid dynamics on load power. Energy-based control does introduce a division as power commands must be divided by the grid voltage to obtain current references. In practice, however, this computational burden is modest and can often be mitigated through approximations. For DC grids with significant constant power loading, controlling energy instead of voltage is often the more robust choice. #DC, #Microgrids, #800V, #SystemDesign, #DataCenters

The air-cooling era of the data center is officially over. A traditional enterprise server rack runs at about 8kW. We are now entering the era of the 1MW rack: A concentration of power and heat that fundamentally breaks the laws of traditional infrastructure. Average power density per rack doubled from 8kW to 17kW in twenty-four months and is projected to hit 30kW by 2027. But that is just the beginning. The individual GPU levels are scaling from 150W to 1,500W, with some projections exceeding 10,000W per unit. This requires a transition from 100kW IT racks to 1MW configurations. How do you deliver and dissipate 1MW in a single IT rack without catastrophic failure? To reach this density, the industry is moving toward disaggregated systems and high-voltage direct current (HVDC) architectures. By bringing 480V AC into the rack and converting it to +/- 400V or even 800V DC, you can distribute power directly to the IT gear through HVDC bus bars. This configuration, based on the OCP Diablo spec, allows for 110kW to 180kW power shelves interspersed with battery or capacitive backup units. The thermodynamic reality is even steeper. As rack power increases 20x, we see a parallel 20x increase in heat generation. This requires in-row Coolant Distribution Units (CDUs) with a 1.8MW capacity, operating at flow rates of 1.5 liters per minute per kilowatt. Liquid cooling is a mandatory physical constraint to prevent clock-speed throttling across clusters of 72 GPUs acting as a single logical chip. But where does a 1GW site find this level of stable, always-on power? Traditional grids cannot handle the 1GW "flat load" profile of a massive AI cluster. In France, our nuclear baseload provides the sovereign, carbon-free energy needed for 99.9% frequency stability. If the grid frequency deviates, you risk losing an entire training epoch during checkpointing. Most companies simply cannot build this vertical integration.