The Grid Can't Keep Up With AI's Appetite — And the Problem Is Deeper Than Power
On an October morning in 2024, operators on the Texas grid watched a strange thing ripple through their system: a 23-hertz oscillation with peak-to-peak swings of up to 50 megawatts — roughly the pull of 40,000 homes — pulsing where it had no business pulsing. The fingerprint of the disturbance pointed to a single large electrical load and its power-supply units, whose control firmware had gone subtly, dangerously unstable. Six months later, across the Atlantic in Dublin, a fault on an underground cable triggered protection systems that dropped 387 megawatts of data-center demand in an instant, sending the Irish grid's frequency climbing and its rate-of-change-of-frequency racing to about +0.26 Hz/s — one of the sharpest grid excursions a data center has ever caused.
These two events, catalogued in a sweeping new technological review from a team at Aalborg University, reveal an uncomfortable truth about the AI boom that tends to get lost in the breathless coverage of model releases and chip announcements. AI data centers are not just big consumers of electricity. They are becoming unstable, interactive, grid-scale electrical machines — and the grid was never really designed for the kind of load they represent. The paper, "From Grid to Chip," makes the case that we have been thinking about AI's energy problem in the wrong units. It is not only about how much power AI needs, though that is staggering. It is about the full electrical path from the utility's high-voltage lines down to the sub-1-volt rails that feed a single GPU, and every fragile link in between.
The Science
The authors — Yubo Song, Rui Kong, Takuro Umihara, Pooya Davari, Frede Blaabjerg, and Subham Sahoo, all at Aalborg University's Department of Energy — have written what amounts to a field guide to the electrical frontier where AI meets the power grid. Frede Blaabjerg is one of the most cited power-electronics researchers in the world; the group sits at the intersection of grid engineering and converter design, the exact disciplines now being stress-tested by hyperscale computing.
This is a review paper, but not the sleepy kind. It synthesizes hundreds of prior studies, incident reports, grid-code revisions, and real-world failure events into a single, structured argument: that AI data centers must be understood as grid-interactive computing systems, not as passive loads. The authors trace three interlocking threads. First, they document the brutal realities of getting these facilities connected to the grid at all — the interconnection queue backlogs, the supply-chain delays, the regulatory patchwork emerging in nine different transmission system operator (TSO) regions. Second, they map the evolution of power-delivery architecture, from the medium-voltage transformer at the grid interface down to the on-chip voltage regulator that feeds a single accelerator. Third, they build a three-level stability framework — rack-level DC-bus dynamics, facility-level converter interactions, and system-level grid-coupled behavior — to explain why AI loads destabilize grids and what can be done about it.
The methodology is mostly observational: real incident data, real grid-code documents, real demand statistics. But the synthesis is the contribution. No one had previously connected workload orchestration, power-electronics topology, and grid stability into a single "grid-to-chip co-design" framework. That framing — treating the entire electrical path from utility to silicon as one coupled system that must be designed together — is the paper's central intellectual move.
What They Found
Start with the appetite. The amount of computational resource used to train AI models has doubled roughly every 3.4 months since 2012 — an exponential curve that makes Moore's Law look like a warm-up. The paper reports that ChatGPT's average daily user messages grew from 451 million in June 2024 to 2,627 million by June 2025, a near-sixfold jump in a single year. Global data-center electricity consumption stood at 416 terawatt-hours in 2024 — 106 TWh for enterprise, 144 TWh for colocation and service providers, and 166 TWh for hyperscale operations. Under the International Energy Agency's base-case scenario, that reaches 946 TWh by 2030 and roughly 1,193 TWh by 2035 — nearly tripling in a decade.
Global data center electricity demand is set to nearly triple by 2035
Global data center electricity consumption in terawatt-hours, actual 2024 and IEA base-case projections to 2030 and 2035.
| Label | Value |
|---|---|
| 2024 | 416 |
| 2030 (IEA base) | 946 |
| 2035 (IEA base) | 1,193 |
and
Hyperscale operations lead 2024 data center electricity use
Breakdown of 2024 global data center electricity consumption of 416 TWh by operator type.
| Label | Value |
|---|---|
| Enterprise | 106 |
| Colocation & service providers | 144 |
| Hyperscale | 166 |
sketch these trajectories.
The consequences are already visible in residential bills. Operational data centers have raised US wholesale electricity prices by 2% to 6% on average nationwide between 2021 and 2025, with far larger regional effects where data-center corridors cluster.
![Fig. 1: Demand increase trend of data centers. (a) Computational resources used by each AI model training [1], (b) Global electricity demand growth of data centers [13], (c) Rise of average electricity price in the U.S. [15].](https://arxiv.org/html/2609.11649v1/demand.png)
shows the compounding picture: compute demand rising exponentially, grid consumption rising, and US average residential rates climbing since 2022 — first on the back of natural-gas price spikes after Russia's invasion of Ukraine, then increasingly driven by data-center connection demand.
But the more arresting findings are about bottlenecks and instability. Amsterdam and Tokyo now carry the longest interconnection lead times in the world — ten years from planning to integration. Amsterdam has paused new data-center projects entirely due to insufficient grid capacity. Construction costs run $6–18 million per megawatt depending on the region. The paper's comparison of major corridors is stark: Virginia, the world's densest data-center region, has 11,300 MW operational and a staggering 37,500 MW planned, with a seven-year lead time (
Virginia dwarfs every other data center corridor on operational capacity
Operational data center capacity in megawatts across major global corridors, showing Virginia's dominance.
| Label | Value |
|---|---|
| Amsterdam | 850 |
| Tokyo | 1,200 |
| Frankfurt | 800 |
| London | 1,290 |
| Virginia | 11,300 |
| Atlanta | 1,800 |
| Singapore | 1,000 |
| Sydney | 750 |
). The root of the delay problem: while a data center can be built in 18 to 36 months, grid operations take 5 to 10 years. The bottleneck is almost never construction — it's the grid.
Then there's the instability catalogue — the incidents that reveal how fragile this coupling really is. In June 2022, Dominion Energy's system saw a 10–11 Hz UPS control mode destabilize into a sustained 14.7 Hz oscillation. In 2017, Meta's data centers experienced an 11 Hz oscillation driven by impedance resonance between aggregated server power supplies and the facility's distribution network — a case of "negative damping" where the converters themselves were injecting the instability. In 2023, a Midwest forced-ringdown event showed data-center power electronics unintentionally cycling on a one-second schedule, repeatedly exciting an 11 Hz natural grid mode — an external "forcing" that the grid could not shake off. And the 2024 ERCOT event in Texas registered 23 Hz oscillation with 50 MW peak-to-peak power swings, traced to a poorly damped mode in the power supply's power-factor-correction and DC-link voltage control, aggravated by legacy firmware.
Perhaps the most eloquent illustration of the mismatch is in the numbers themselves. The GPU power curves in the paper (
![Fig. 3: Needs of flexibility of data center. (a) Typical computing power flow of GPU [58], (b) Needs of power smoothing by the data center components [6], (c) Needs of peak cut by shifting time, or allocating workload to another data center.](https://arxiv.org/html/2609.11649v1/flexibility.png)
) show consumption beginning to fluctuate within about 15 seconds of operation. And per NVIDIA, thousands of GPUs run in lockstep during AI training — performing the same computation on different data simultaneously — producing grid-level power fluctuations that would be unacceptable without smoothing. An AI data center's power demand is not a slow, gentle tide; it is a synchronized, spiky waveform that, left unchecked, resonates with the very infrastructure meant to serve it.
shows the practical answer: shaving load by reallocating workloads between 6:00 and 8:00 and discharging on-site battery storage between 18:00 and 21:00.
Why This Changes Things
The paper's most consequential finding might be a single, counterintuitive number buried in its bottleneck analysis: an analysis of the US power system concludes that by factoring in flexibility capable of curtailing just 0.25% of annual peak demand, fully 76 gigawatts of new load — equivalent to 10% of the nation's current aggregate peak — could be integrated into the grid without new transmission. The catch? The conventional calculation of available connection capacity compares a project's peak output to "maximum output during annual peak demand" and ignores demand-side flexibility entirely. In other words, the grid has more room than the paperwork admits — but the connection rules were written before anyone imagined a load that could shift its work across time and geography.
That is the real shift this paper documents. Connection policies are being rewritten across the developed world, and they are converging. The old "first-come, first-served" queue — which rewarded speculative applications that rushed to claim grid priority — is giving way to "first-ready, first-served," adopted in the UK in 2025 and in Germany and Texas in 2026. The Netherlands, since January 2026, prioritizes connections on social importance — residential construction, hospitals, schools, national defense. Grid codes are tightening around fault ride-through, power quality, and ramping limits; a key split is visible in how strictly TSOs meter how fast a data center can change its load — Canada's AESO limits ramping to 10 MW/min while Denmark's Energinet allows up to 60 MW/min. In Ireland, EirGrid began requiring data centers of 10 MVA or more to bring on-site or proximate generation from December 2025. And at the time of the paper's submission, Canada, Denmark, and the Netherlands had all paused acceptance of new connection applications.
None of this is anti-AI. It is the opposite: it is the machinery of making AI connectable. The paper's deeper argument is architectural. The power-delivery path from grid to chip has evolved from a simple passive transformer to a cascade of active power-electronics converters — medium-voltage solid-state transformers stepping down to 800-volt DC distribution, then to 48-volt rack rails, then through intermediate bus converters to the 12-volt level, then finally to on-chip voltage regulators delivering sub-1-volt to the accelerator itself. Each stage is a DC-DC or AC-DC converter, each has control dynamics, each is a potential source of instability, and each increasingly sits inside a data center that also packs liquid cooling directly onto the chips. The paper identifies constant-power-load effects, impedance interactions, forced oscillations, and operating-mode transitions as the four dominant instability mechanisms across the three stability levels. At rack level, the DC bus must survive sub-millisecond load steps from compute phase transitions. At facility level, hundreds of converters must not excite each other's resonances. At system level, the whole facility must maintain fault ride-through when the grid hiccups — and, critically, must ride through without disconnecting and dumping its load back onto an already-stressed grid.
The grid-to-chip co-design idea means that workloads, power-delivery architecture, energy buffering, and grid operation can no longer be treated as separate disciplines. A data center that can shift its training tasks by an hour, discharge its UPS batteries (which the paper notes have historically been used for less than 0.1% of their operational lifespan), pre-cool during cheap hours, and present a smooth, damped electrical profile to the grid is not just a tenant — it is a grid asset.
What's Next
The paper's forward-looking sections are speculative but concrete. One concept — "computing on power infrastructure" — imagines data centers harnessing the intrinsic dynamics of the power grid itself to perform computations, capitalizing on converged power-information systems (
). Another plot is immediately useful: carbon intensity of electricity across major data-center hub cities varies dramatically by time of day, and
![Fig. 17: Comparison of carbon intensity of electricity in major data center hub cities by Coordinated Universal Time (UTC), August 12th, 2026, obtained from Electricity Maps [230].](https://arxiv.org/html/2609.11649v1/Carbon.png)
shows those curves — meaning that workload scheduling across geographies is also, in effect, carbon scheduling. A task run in a low-carbon city at a given hour measurably reduces emissions, and spatial flexibility can route around dirty grids in near-real-time.
The honest caveats deserve airing. A data center's flexibility is bounded by service-level agreements — you cannot postpone an interactive chatbot request the way you can delay a batch training job. Pre-training workloads are particularly inelastic. And while 76 GW of "hidden" headroom sounds like a gift, capturing it requires rewriting connection rules, building control systems that can participate in demand response, and — critically — proving to grid operators that a facility's load-shaping promises are enforceable, not speculative. The regulatory convergence the paper documents is real, but it is uneven; Japan, for instance, has only begun discussing grid-code revisions precisely because it has integrated so few data centers so far.
What this paper gives the world is a vocabulary for the era we have already entered. We are no longer asking whether AI infrastructure will strain the grid — we are living through the era when it does. The question the Aalborg group poses is subtler: whether we can design the glass pipeline from the power station to the transistor well enough that AI's electricity problem becomes a co-design opportunity. The Texas oscillation, the Dublin frequency spike, the ten-year waitlists in Amsterdam — these are not footnotes to the AI story. They are the story. The grid and the chip have become one machine, and we are only beginning to learn how to operate it.