AI Data Center Cables — Power Distribution, Data and Fire Integrity

01How the density shift reorganised the cable specification

The change in AI data centres is not incremental. A traditional enterprise rack at 5 to 10 kW could be fed from a couple of circuits and a standard whip. An AI training rack built around current-generation accelerators draws 100 to 250 kW, some designs already exceed 500 kW per rack with liquid cooling integrated into the rack itself, and leading-edge GPU training clusters are approaching 1 MW per rack. Facility totals have crossed 150 MW. Applying the cable schedule from a previous enterprise or colocation build to that is not a conservative shortcut; it is a category error, because the numbers do not scale linearly. Cross-sections balloon, tray space evaporates, derating becomes a first-order design input, and fire classification moves from preference to code requirement.

This page is part of cable solutions by industry, the cross-industry overview that compares the application families Sorivo supplies and the standards governing each.

Before any of that, it helps to state where each cable type sits. Power enters a data centre through a short sequence of stages, and the cable requirement changes character at each one — from fault-current withstand at the utility interface, through circuit integrity on the critical path, to space and flexibility at the rack.

Stages two to four and stage six are where the density shift bites hardest, and they are what the rest of this page covers. Stage one is a medium-voltage problem that NEC 2026 has just rewritten the rulebook for, and stage five is dominated by connector temperature ratings rather than by cable construction.

02Sizing a high-density rack, and the derating trap underneath it

Cable sizing starts from apparent power rather than active power, because the UPS and PDU have to deliver the total volt-ampere demand. At a typical data centre power factor of 0.9, a 100 kW rack corresponds to approximately 111 kVA. That figure converts to roughly 308 A per phase at 208 V three-phase and roughly 154 A per phase at 415 V — a halving of current from a voltage change alone, which is the whole reason hyperscale designs have moved to 415 V distribution for AI halls.

The 415 V column is the practical argument for the voltage choice. A 100 kW rack needs 70 mm² at 415 V against 150 mm² at 208 V, which is a large difference in tray space once multiplied across a hall. Specifying 208 V distribution for a new AI build locks in the higher cable cost and the congestion for the life of the facility, and 415 V three-phase is already standard in much of the world outside North America.

02.1 Where the derating trap closes

Ampacity tables assume a 30 °C ambient with a cable clipped direct. A hot aisle in an AI cluster runs at 35–45 °C, and the cables in that aisle are usually grouped in trays and often run through conduit on the way. The corrections stack multiplicatively, and the result is not a marginal adjustment.

A conductor read off the table as capable of 179 A ends up rated at around 106 A once it is routed through a hot aisle alongside other cables and sleeved in conduit — barely enough for a 35 kW rack, against the 100 kW the original calculation assumed. The practical rule that follows is to size against the derated condition rather than the table value: take the tabulated ampacity for the conductor's temperature class, apply a blanket factor of about 0.70 for temperature, grouping and conduit, and work from that figure. It is conservative, and it is the difference between a feeder that holds at peak summer load and one that does not. Our derating guide covering temperature, altitude and grouping sets out the individual correction factors, and the ampacity calculator applies them without manual table lookups.

02.2 Why the conductor is copper

Aluminium is a legitimate cost reduction in much industrial work, but it is effectively absent from the data centre critical path, and the reason is the constraint that actually binds. Copper carries roughly 1.3 times the current of an equivalent aluminium conductor at the same cross-section and insulation class, so matching copper's ampacity in aluminium needs about 1.6 times the cross-section — a bulkier cable in a tray that is already congested. Voltage drop compounds the argument, since power runs from UPS to rack row commonly span 30 to 100 m and aluminium's higher resistivity works against the three per cent limit. Termination behaviour decides it in the end: aluminium connections loosen under thermal cycling, in a facility where rack loads fluctuate daily and racks are reconfigured regularly. Several design standards recommend copper explicitly for power distribution inside the IT space, and the vibration from liquid cooling pumps and fans adds a fatigue consideration that aluminium handles less well. Our comparison of copper and aluminium sets out where each material is the right answer, which depends on whether the binding constraint is metal cost or space.

03Fire performance and circuit integrity on the critical path

Data centres occupy an unusual position in fire codes because the cost of a shutdown dwarfs the cost of most fires. A short outage at a hyperscale AI facility destroys work in progress, and a training run measured in weeks can lose days of compute from an interrupted checkpoint. That economics is why the fire requirement in an AI facility is about keeping the critical load alive long enough for an orderly response, not only about evacuation.

In the United Kingdom and much of Europe, the governing framework for fire-resistant cable in data centres is BS 8519, which defines three circuit-integrity categories. In North America, NFPA 75 sets the baseline for IT equipment spaces, and local codes frequently push beyond it.

For an AI facility the logic points past the minimum. PH30 or PH60 cable buys the time needed to evacuate a building; it does not buy the time needed to bring a GPU cluster down in good order. PH120 gives a facility team two hours to execute a controlled power-down, bring generators online, or let the fire service reach the seat of the fire without destroying the compute investment. BS 8519 Category 3 is the sensible baseline for the UPS-output-to-PDU path in mission-critical AI infrastructure, and where the cost difference between PH60 and PH120 is small, standardising on the higher category across the facility removes a whole class of specification error.

03.1 LSZH is no longer a premium option

Low-smoke zero-halogen sheathing moved from upgrade to baseline over the last decade, and several independent paths now require it. BS 6724 covers armoured LSZH cable with halogen-free performance per IEC 60754 and low smoke per IEC 61034. NFPA 75 requires cables in IT equipment spaces not to contribute to smoke and toxic gas hazards, for which LSZH is the compliance route. EU CPR requires LSZH-sheathed cable rated Dca or higher as the baseline for general IT spaces, with B2ca or Cca possible for main risers and fire-critical circuits depending on building height and fire strategy. The ANSI/BICSI 002 data centre design standard recommends LSZH for all intra-building power and data cabling.

The practical consequence is that sheath material in a data centre should read LSZH across the board — armoured power feeders, control cable and data cabling alike. The premium over a PVC equivalent is modest against a liability that can shut a facility. Our comparison of XLPE and LSZH constructions covers where the two differ as material systems rather than as acronyms, and the difference between fire resistance and flame retardance is worth stating before either term enters a specification.

04Data cabling: where copper still holds the rack

The move to 800G and 1.6T backbones for AI clusters has changed what gets deployed between racks, but it has not removed copper from inside the rack. Cat6A remains the workhorse for server management and 10G connections at reaches up to 100 m, while Cat8 constructions are seeing growing use for short-reach 25G and 40G GPU interconnects within about 30 m. Above that distance the answer is optical: multimode fibre covers spine–leaf and inter-row runs at 100G, and single-mode fibre carries the long-haul and data centre interconnect traffic beyond that. The LSZH requirement applies to every copper data cable running through the IT space.

One published item covers the copper end of that table: Sorivo's Cat6A and Cat8 S/FTP shielded data cable, rated CPR Eca/Dca with IEC 60332-1-2 flame retardance, which suits general IT space structured cabling. Where a route requires B2ca or Cca and runs in a fire-critical path, that is a containment and fire-stopping question rather than a cable-jacket question alone, and the matched system needs specifying as a whole. The distinction that matters for verification is between a cable's own CPR class and the performance of the installed pathway around it.

05EMI, separation distances and the three-zone cabling model

The failure mechanism that does not show up on a thermal camera is electromagnetic interference. An AI training hall is an electrically hostile environment: hundreds of GPU nodes switching at high frequency, dozens of power supplies generating harmonic currents, and high-speed links running at 400G or 800G between nodes. Every power cable in that space acts as both an antenna and a receptor. The symptoms are recognisable once you know them — training jobs that hang or retrain because high-frequency noise coupled into an interconnect and corrupted packets, telemetry that becomes unreliable when monitoring signals pick up switching noise, and intelligent PDUs losing communication with the monitoring system because of common-mode noise on the power line.

The standards position is not ambiguous. TIA-942 and ISO/IEC 24764 both specify a minimum separation of approximately 30 cm between power and data cabling in a data centre environment, as a certification baseline rather than a suggestion. Where a crossing is unavoidable it should be at right angles, and parallel runs of power and data should be kept short. Where AI deployments compress that separation because of space constraints, intermittent network errors are the usual result, and they are difficult to attribute after the fact.

05.1 Separating the overhead plane by function

Dividing the overhead cable plane into three functional zones removes most of the coupling problems structurally rather than by remedial measure. The value of the arrangement is that it prevents three distinct failure modes at once: heat rising from power cables into data cables, magnetic fields from power inducing noise in data lines, and maintenance work on power disturbing fibre runs.

One layout constraint comes directly from NEC 2026 rather than from good practice: Article 392 now requires vertical clearance between stacked cable trays, which prevents the lower tray from heat-soaking the cables above it. A four-tray stack that previously fitted inside a given vertical space now needs roughly twice that, which forces wider tray spreads or more horizontal runs — a real cost in a densely packed hall, and the reason tray layout needs deciding early. In tray stacks serving AI racks, limiting fill ratio to about 40 per cent rather than the code maximum buys a measurable reduction in heat buildup around the conductors, and it is one of those provisions above the minimum that repays itself in cable life. Cable tray sizing and fill ratio calculations are set out in our tray sizing guide.

06Standards in transition: NEC 2026 and the 800 V DC question

Two changes are running in parallel in 2026, and they are at very different stages of maturity. One is a code revision with published text; the other is a voltage architecture with published intent and no system standard. Treating them as the same kind of thing is how specifications end up citing a roadmap as if it were a normative document.

06.1 NEC 2026 gave medium voltage its own article suite

NEC 2026, published by NFPA in autumn 2025, is the first edition to give medium-voltage power distribution a standalone rulebook. Six articles now govern systems above 1,000 V AC or 1,500 V DC, and the content was relocated from articles that previously held it in scattered form.

The load-calculation article moved as well: Article 220 is now Article 120, and the new 120.7 allows power control systems to be factored into load calculations, which gives designers a code-sanctioned route to right-size services rather than defaulting to worst-case headroom on every feeder. Chapter 8 communications content was absorbed into Chapter 7. For anyone holding a 2023-era specification template, the renumbering is the practical problem: a document that cites Article 220 will read as out of date during plan review, and the fix is administrative rather than technical.

The provision with the largest commercial consequence is that medium-voltage equipment must now be listed or field-evaluated. For a buyer importing cable assemblies, terminations or prefabricated bus into a US project, that raises the documentation bar, because a recognised evaluation mark or a documented field evaluation is now the compliance evidence. The question worth settling during the RFQ rather than at inspection is which of those two paths a given product line uses.

06.2 Where 800 V DC actually stands

The second change is architectural. An 800 V DC distribution scheme replaces today's ladder of AC-to-AC and AC-to-DC conversions with a single conversion, and the argument for it is arithmetic: at higher voltage the current falls for the same power, and conductor losses fall with the square of current. The same reasoning applies inside a data centre at rack scale, where moving distribution voltage up reduces both current and cable bulk for a given rack load.

What matters for a specification is separating what exists from what is expected. The table below states the current position on each item, and the right-hand column is the one that keeps a submittal out of trouble.

The single most useful takeaway from that table is the separation of fire from voltage. Reaction to fire is a property of the cable as installed in the building, not of the circuit it carries, so EN 50575 classes, IEC 60332 flame propagation and IEC 61034 smoke measurement apply identically whether the bus is 415 V AC or 800 V DC. Keeping fire classification and voltage class in two separate columns of a specification is the way to avoid a cable being approved on the wrong basis. The full register, with the source for each line, is maintained in our 800 VDC data center cable specification guide.

07Selection toolset for data centre electrical designers

07.1 Routing each zone to a construction

07.2 What a one-megawatt pod actually bills out as

A 1 MW AI training pod built as eight racks at approximately 125 kW each produces a bill of materials with six distinct construction families in it, and the pattern is worth seeing because it shows how little the stages share. The main feeder from UPS to pod PDU runs 4-core 185 mm² CU/XLPE/LSZH with PH120 circuit integrity over roughly 50 m per cable, one cable serving two racks. PDU to rack-level distribution uses 4-core 70 mm² CU/XLPE/LSZH/SWA/LSZH at around 15 m per run. The rack PDU whips are 3-core 16 mm² CU/XLPE flexible with an LSZH jacket. Data is Cat6A S/FTP for management and 10G, with Cat8.1 S/FTP for the short GPU interconnects. Liquid cooling pump power uses 4-core 6 mm² flexible LSZH. Six families, three voltage classes, two fire categories.

One detail in that bill of materials is worth isolating because it changes the tray count. Using a 4-core 70 mm² armoured cable rather than four separate single-core 70 mm² runs occupies roughly a third of the tray space, because one cable at an outer diameter around 35 to 40 mm replaces four individual cables. Fitting 1 MW of cabling into a 600 mm wide overhead tray is decided by choices of that kind, not by the conductor material. Where a single PDU feeds several racks and the feeder current climbs past roughly 300 A, busway becomes more space-efficient than cable, and many current hyperscale designs use a hybrid — busway for the main floor distribution and flexible cable for the final connection to the rack PDU. The crossover is generally in the 300 to 400 A region and varies by manufacturer.

07.3 Verification checks before the cable is installed

1. Read the PH rating on the sheath. For BS 8519 Category 3 the cable should be marked with the fire survival time and the test standard. A cable marked only as fire resistant, without a PH designation, is not a circuit-integrity claim and should not be accepted for the critical path.
2. Check the LSZH marking against the test standards. The sheath should reference IEC 60754 for halogen-free performance and IEC 61034 for smoke. A burn test on a sample is a crude but effective field check, because PVC produces dense black smoke and acidic fumes while a genuine LSZH compound produces thin white smoke with no halogen.
3. Measure the outer diameter. Tray fill calculations depend on it, and a cable dimensionally over specification by more than a small margin changes whether the installation fits. This matters more in a data centre than in any other facility type because the trays are packed.
4. Request alien crosstalk data for copper data cable. In high-density racks, near-end alien crosstalk is the limiting performance factor for Cat6A and Cat8. Ask for the manufacturer's test results before accepting delivery rather than after commissioning failures appear.
5. Verify the meter marks. On a long UPS-to-PDU run a small length error compounds, and meter marks should be checked against the actual length on a sample drum.

08Engineering questions we are asked most often

These are the questions that recur in design reviews, answered from the same basis used above. Where a figure is a general engineering threshold rather than a company measurement, it is described as such.

Can I use PVC-sheathed cable in a data centre?

Not in the IT space. PVC produces dense smoke and releases hydrogen chloride when burned, both hazardous to people and corrosive to equipment, which is why LSZH is required in plenum and air-handling spaces under NFPA 75, the EU CPR framework and ANSI/BICSI 002. In plant rooms and on exterior runs — generator feeds and the medium-voltage incoming — PVC remains acceptable, though many operators standardise on LSZH across the site for consistency in procurement and spares.

Why is aluminium effectively absent from data centre power distribution?

Because the binding constraint is tray space rather than metal cost. Copper carries roughly 1.3 times the current of aluminium at the same cross-section, so an aluminium conductor needs around 1.6 times the cross-section to match, which is bulkier in an already congested tray. Voltage drop over the 30 to 100 m runs typical from UPS to rack row compounds it, aluminium terminations loosen under the daily thermal cycling of a reconfigurable facility, and the continuous low-frequency vibration from liquid cooling pumps and fans adds a fatigue consideration. Several design standards recommend copper explicitly for power distribution inside the IT space.

Does higher rack density mean I need busway instead of cable?

Above roughly 300 A per circuit, busway becomes more space-efficient than cable. Where a single PDU feeds four to six 100 kW racks the feeder current can exceed 600 A, which is comfortably inside busway territory. Many current designs use a hybrid approach: busway for the main floor distribution between switchboard and row, and flexible cable for the final connection to the rack PDU. Busway installs and reconfigures faster but carries a higher upfront cost, so the crossover point is a project decision, generally in the 300 to 400 A region.

What documentation should I require to verify an LSZH claim?

Three things for the exact model being shipped: the certificate reference with the issuing body named, the relevant test report for the tested construction, and batch traceability from the delivered drum back to production. Then verify the certificate number with the issuing body rather than accepting a scanned copy. For the LSZH property specifically, the test report should reference IEC 60754 for halogen acid gas and IEC 61034 for smoke density, because a manufacturer marking LSZH on a PVC sheath is a documented counterfeit pattern. Our cable supplier audit guide works through conductor purity to spark test for a structured assessment.

Do I need to change fire classification work because of 800 V DC?

No. Reaction to fire is a property of the cable as installed in the building, not of the circuit it carries, so EN 50575 classes, IEC 60332 flame propagation and IEC 61034 smoke measurement apply the same way whether the bus is 415 V AC or 800 V DC. Keep fire classification and voltage class as two separate lines in the specification, because treating them as one question is how a cable ends up approved on the wrong basis.

Should a normal enterprise data centre plan for 800 V DC now?

Probably not for the next build. The public roadmap points at 800 V DC power hardware in the second half of 2026 and full rack or row deployments around 2027, tied to specific processor generations, and independent analysis expects hybrid AC and DC deployments to dominate in the meantime. There is no announced date for when this becomes procurable outside hyperscale-scale builds. What a non-hyperscale buyer should take from it is procurement hygiene: name your standards explicitly, require DC test evidence rather than voltage claims, and put a standards-change clause in supply contracts so a future transition can be handled without renegotiating the whole agreement.

09Why Sorivo, and how to check our claims

09.1 Team experience, stated precisely

Sorivo's senior team brings more than 15 years of combined experience in cable design, production control and quality management. We describe it as team industry experience rather than company age, because that is what it is — the brand began exporting industrial cable in 2023 on the back of a long-standing team and a manufacturing partnership, and the figures are published that way on our about page. Projects and buyers currently span 300+ clients across 30+ countries.

09.2 Application engineering on the four calculation-heavy decisions

Data centre cable specifications turn on four calculations: conductor sizing against derated ampacity in a hot aisle with grouped circuits, voltage drop over the actual route from UPS to row, jacket chemistry matched to the enclosure and to coolant exposure, and phasing and separation practice around a dense tray installation. Support covers those before quotation rather than after a commissioning problem. The cable engineering tools section carries free utilities for repetitive work, including the ampacity calculator and the derating guide.

09.3 What comes with the cable

For any published model we can provide the certificate reference covering the exact construction, the relevant test report, and batch traceability from the delivered drum back to production. On the data centre side the published range covers CU/XLPE/LSZH/SWA/LSZH armoured power cable for the distribution stages, the BS 6387 CWZ fire-resistant armoured cable for circuit-integrity duty, Cat6A and Cat8 S/FTP shielded data cable, and the ESS energy storage cable for battery-side work. Two checks are worth doing regardless of supplier: verify the certificate number with the issuing body rather than accepting a scan, and confirm that the certificate covers the model code actually being shipped rather than a similar construction in the same family. The global certification guide and the supplier qualification checklist set out the process from the buyer's side.

Reviewed by Luo Qiang — Senior Cable Application Engineer, Sorivo
15+ years in industrial and renewable energy cable specification. Experienced in cable specification aligned with IEC standards.

How this was written: every standard number and product claim on this page is taken from the corresponding published Sorivo product or technical article, and the standard identifiers were checked against the issuing body's own catalogue before publication. Claims about the 800 V DC transition are marked with their status rather than presented as settled standards. Where a figure is a general design threshold rather than a company measurement, it is described as such.

Standards referenced on this page

10Frequently asked questions