AI load no longer resembles a classic industrial load

Three structural differentiators separate the load of an AI campus from that of a traditional data centre.

Power density per rack

NVIDIA Blackwell Ultra dissipates 1.4 kW per GPU. A GB300 NVL72 rack dissipates between 125 and 140 kW. The 2026–2028 frontier is 200 kW per rack, with a 1 MW horizon. This density has no precedent in classic industrial grid connection.

Dynamic time profile

AI training cycles aggregate and release tens of MW on sub-second time scales. Distributed inference modulates the load in the 0.1–10 Hz range, within the flicker and power-quality disturbance band monitored by DSO analysers.

Geographical concentration

An AI campus aggregates 50–200 MW in a footprint of a few hectares, something historically seen only in heavy industrial plants with a dedicated HV connection. The 2025–2026 transition to bipolar 800 V DC architectures, driven by NVIDIA and leading power-infrastructure manufacturers, concentrates primary AC/DC conversion at a single point.

Four measurable effects at the point of common coupling

These are the effects captured by power-quality analysers at the head of an MV feeder with an active AI campus downstream.

The mitigation chain: from the rack to the PCC

Four coordinated layers with distributed allocation of the dynamic response. A poorly sized layer degrades the technical contract of the whole.

Regulatory framework: the DSO perspective

European grid codes are tightening the real-time technical connection contract.

ENTSO-E NC RfG 2.0, Phase II

Current response <10 ms · damping ≥5% on electromechanical oscillations · full ride-through of phase-angle jumps · compliance with the mechanical starting time method for virtual synchronous systems. Threshold: any BESS or renewable plant >1 MW. Its transposition to the data centre as a grid-interactive load is an open area of regulatory discussion; CIRED 2026 (Dublin, June 2026) was one of the forums where that discussion was being formalised among European operators.

German inertia market (28 January 2026)

First published European inertia price, with four product categories and mandatory VDE FNN certification. A data centre with a bidirectional BESS and a compliant PCS is technically in a position to bid for those contracts, turning mitigation infrastructure into a monetisable asset that improves the business case of the connection.

UK Pathfinder: Blackhillock

First commercial Virtual Synchronous Machine plant in operation, demonstrating grid-forming performance under real grid conditions. It is the benchmark for what an NC RfG 2.0-class system must deliver commercially, not only in laboratory certification.

Iberian blackout (28 April 2025)

The most widely read case study in Europe on cascades caused by low residual inertia. Direct evidence of the consequences of an under-specified dynamic response at points of connection, including large loads with passive PCS. The event has accelerated DSO requirements for documentation of dynamic response in new connection agreements.

The Premium SA proposition

Over 40 years of mission-critical industrial power electronics, mapped to the grid interface of the hyperscale market.

ODX/OVX-6000

Rugged inverters for critical microgrids: the base on which bidirectional PCS with Virtual Synchronous Machine (VSM) and sub-second response are developed. Main adaptation: migration of the control to an NC RfG 2.0-compliant architecture, with an auditable dossier for the DSO generated from the D2W gate.

EPS line

DC UPS with sub-millisecond response and fast transfer certified in critical railway applications. Rescaled to the industrial data centre sidecar, it covers the BBU/CBU layer without structural platform modification; it only requires additional qualification gates for EN 50600 and, where applicable, OCP.

CHB topology

Cascaded H-Bridge: pure modularity · no capacitor-balancing issues · fault tolerance through cell bypass · natural isolation per cell. It is the canonical topology for the grid-scale BESS + STATCOM segment on data centre campuses, where the inertia or voltage-support contract becomes bankable before the DSO or the ancillary services market.

D4Co methodology

Design for Compliance: from the first D2W gate, the dossier that the operator will submit to the DSO or the regulator so that the connection is approved and the grid-support services provided are remunerable. Derived from over 40 years of railway RAMS, derating to 80% of T_j,max and post-deployment FRACAS. It turns a technically good PCS into a contractually bankable PCS.

The boundary between data centre and grid is not a cable: it is a well-designed PCS

When a European hyperscaler requests a new 50 or 100 MW connection from a DSO, the technical debate is no longer settled by sizing an additional transformer. It is settled by articulating a quality contract at the PCC: THD limits, ride-through profile, dynamic reactive power response and behaviour during phase-angle jumps. That contract requires equipment at the point of connection that behaves like the grid, not like a load.

Premium SA provides the industrial asset — ODX/OVX-6000, EPS, CHB topology — and the D4Co methodology that turns the operator’s technical specification into an auditable compliance dossier. When the difference between award and rejection is measured by the quality of the technical dossier, that discipline is decisive.