The industrial motor-control category looks stable from the outside. Hyperscale AI data centres break it. Soft starters are nearly absent in AI data centres. The three-stage drive separates: rectification moves upstream to a shared facility converter, inversion moves downstream to a card next to each motor, and over time the inverter moves inside the motor.
This companion essay reads the mechanical-plant drive stack through the lens of the DC-DC Transition Series. The underlying thesis is the one the series has been building: as today's 54 VDC rack distribution moves toward 800 VDC architectures and potentially toward facility-wide DC, and as direct-to-chip liquid cooling moves from niche to default on frontier training clusters, the industrial drive catalogue of the next decade starts to look very different from the one the primer catalogued. The primer is the right place to start if you want the baseline. This essay is about what the baseline becomes.
1. The two architecture shifts and where they point
The architecture changes are easy to name. One is the migration of rack-level DC distribution from today's 54 VDC architecture toward 800 VDC. GB200 and GB300 racks still use 54 VDC on the busbar; NVIDIA's Vera Rubin and the Kyber-class rack architectures move the distribution layer toward 800 VDC. The hyperscaler motivation is efficiency and copper. Every additional AC/DC or DC/DC conversion on the IT power path introduces another efficiency loss and another thermal burden, and 800 VDC drops the busbar current to a point where the copper and distribution cost of a 300 MW training campus becomes workable at all.
The other is liquid cooling. Direct-to-chip and immersion architectures are shipping at hyperscale with coolant distribution units of 500 kW to 2 MW serving racks via secondary-loop manifolds. The driver here is heat flux at the die. Blackwell-class accelerators dissipate 1200+ watts per package and modern CPU dies cross 400 watts. Air cooling at those thermal densities requires rack airflow volumes and CRAH footprints that collapse the economics of the data hall. Liquid cooling pulls the heat out at the source and sends it to the facility loop at a higher temperature than air can.
Both start as IT-side architecture changes. Each bleeds into the mechanical plant through a different mechanism, and together they push the drive content in the same architectural direction: integration of the inverter into the load, consolidation of the rectifier at the facility level, and disappearance of the standalone AC starter cabinet.
2. What the IT-DC transition does to the mechanical plant
A rack-level DC bus, today's 54 VDC or the emerging 800 VDC, leaves the mechanical plant alone. Chillers, chilled-water pumps, cooling-tower fans, CRAH fans and fire pumps all stay AC. The interesting question is whether the DC bus crosses the IT boundary. A facility-wide DC architecture is still a scenario, pushed forward by OCP and hyperscaler reference architectures.
If it does cross, the three-stage drive separates. The AC/DC rectifier externalises to a shared facility active-front-end (AFE); each drive on the DC bus becomes a lightweight inverter card. Harmonic compliance consolidates at the utility PCC under the facility AFE. Value in the drive supply chain migrates from the motor-side drive box toward the facility power-electronics layer.
Drive-box revenue per motor falls because the box is simpler. Facility AFE revenue per site rises because the AFE is new infrastructure. Vendors who sell both ends win; drive-only vendors lose.
Local DC-link capacitance stays. Bulk energy store migrates upstream to the facility AFE; high-frequency decoupling stays at the switching bridge for di/dt during each IGBT or SiC pulse.
DC fault interruption is harder than AC: no natural zero crossing to extinguish the arc. The figure below shows the options and their trade-offs. Standards: IEC 60947-10 (2026) covers semiconductor and hybrid breakers to 1,500 V DC; IEC 60947-2 covers mechanical LV DC breakers across the same range.
Grounding topology is a new facility decision. Each option has different fault-detection, bearing-current and personnel-safety consequences:
- Floating IT
- High-resistance grounded
- Solidly-grounded bipolar (plus/minus 400 VDC poles making an 800 VDC pole-to-pole bus)
Long DC runs with PWM switching noise from distributed inverters degrade PM motor bearings without isolation transformers or common-mode chokes upstream.
Insulation coordination follows from the grounding choice. A floating (IT) or high-resistance-grounded 800 VDC bus pushes the full pole-to-pole voltage onto any single device during a fault, driving the inverter stage to 1,200 V-class SiC or IGBT. A solidly-grounded centre-tapped bus holds each pole to plus/minus 400 V against ground, which keeps the inverter devices in the 650 V to 750 V class with better switching efficiency and lower cost. The grounding scheme also sets the common-mode voltage profile that drives bearing currents, so motor insulation and shaft-grounding provisions downstream depend on this choice.
Rectifier savings per drive are concrete; DC protection is new capex and (on SSCBs) new conduction loss. Net balance-of-plant outcome depends on bus voltage, site size, breaker topology and conduction-loss budget.
Capex per MW redistributes across the drive stack.
Where the BOM lands depends on breaker topology:
- Fuse-based protection : DC-native BOM lands below AC baseline
- SSCB-based protection : DC-native BOM lands above AC baseline
Strategic read is the same either way:
- Full-stack vendor (AFE + DC switchgear + distributed inverter cards) captures every pool that emerges
- Drive-only vendor loses the rectifier pool with no backfill
Facility DC breakeven: high drive counts, short copper runs, fuse-based protection. The liquid-cooled AI hall hits breakeven first; the chiller yard later.
Where each piece of the VFD goes:
- Rectification moves upstream to a shared AFE
- Protection becomes a new infrastructure layer
- Inversion moves to a card next to each motor
- Inverter moves inside the motor over the medium term
3. What happens to the motor itself
The piece of the stack that gets the least attention in the DC-DC transition conversation is the motor. The architectural shift the AI data centre is driving goes beyond rectifier-moves-upstream and inverter-moves-downstream. The motor is migrating from induction to permanent-magnet synchronous, and then from PM-synchronous-with-separable-inverter to integrated motor-plus-inverter modules where the two cannot be sourced or serviced apart.
CDU secondary pumps already ship this way. Grundfos MGE, Wilo Stratos, Xylem Lowara, Armstrong Design Envelope: motor and inverter in one unit, no separable drive. EC fans from EBM-Papst and Ziehl-Abegg do the same on the fan hub. Chiller compressors are slower, but Danfoss Turbocor and other magnetic-bearing centrifugals ship with the inverter in the compressor skid.
Integration kills the drive-brand-sold-separately business. Inverter margin moves to whoever sells the motor. Pump OEMs figured this out a decade ago; fan OEMs did it on EC fans. Does the pattern reach mid-size industrial sockets where drive vendors still own the inverter? Open.
Integration scales awkwardly. At a few kW (CDU secondary pumps, EC fans), the motor's thermal mass and airflow handle the inverter heat. At 50 to 200 kW (chiller and condenser pumps), the heat load exceeds what the motor can dissipate; SiC or IGBT electronics in a 40 to 45 degC mechanical room with pump vibration degrade MTBF (mean time between failures) on capacitors and gate drivers quickly. Swapping a $20k to $100k integrated assembly in the field is a different operations problem from replacing a motor. That is why primary CDU pumps and chiller compressors have resisted integration longer. The threshold moves up with SiC thermal class and housing design, but does not yet cover the full industrial catalogue.
4. Soft starters are nearly absent as a growth socket in AI data centres
The soft starter is an AC motor starting device based on controlled thyristor conduction. Thyristors fire later in each AC half-cycle at the start of a ramp, then advance toward full conduction, limiting inrush current and starting torque. The mechanism depends on there being an AC waveform to control. On a DC bus the mechanism does not exist, and a different starting path has to be designed.
Every AC socket inside a DC-native facility has one of three possible outcomes. It stays on a residual AC subdistribution local to the socket (fire pumps, life-safety smoke fans, specific code-mandated loads). It migrates to a DC-fed inverter card that handles both the starting ramp and the running duty from the same box (make-up air handlers, sump pumps, mid-size facility-loop pumps). Or it vanishes from the design entirely because the architecture does not need a discrete starter at that location any more (dedicated chiller soft-start cabinets, bypass contactor assemblies). The soft starter was already losing ground inside catalogue LV drives with soft-start firmware for a decade. The DC transition finishes that migration cleanly inside AI data centres.
The counter-argument is that the AI data centre is a tiny fraction of global soft-starter demand. True. The soft-starter category runs on upstream oil-and-gas, mining, water-and-wastewater, cement and marine, where induction motors and flat-duty loads dominate and will not go away in a decade. The claim here is narrower. Inside hyperscale AI buildouts, the soft starter is a residual-AC footprint; growth comes from elsewhere; vendors who positioned data-centre expansion as a soft-starter growth story need a different pitch.
5. Liquid cooling has already pre-collapsed the pump drive on the secondary side
Inside a coolant distribution unit, two pump architectures coexist. The primary side circulates facility water through the CDU's heat exchanger, commonly with induction pumps sized in the tens of kW on external LV VFD cabinets. The secondary side circulates the rack-side coolant between the CDU and the rack manifold, commonly with permanent-magnet synchronous pumps of a few kW each whose inverter electronics are built into the pump housing. Published vendor sizing varies by rack density and loop topology; the two sides look nothing alike on the drive side.
The secondary-side architecture is already the thing that a facility DC bus would generalise. The external drive cabinet is absent. The inverter is integrated into the mechanical load. The communication path can be a direct Modbus or BACnet interface from the pump controller to the CDU control layer, subsuming the separate drive panel wired into a BMS. The vendor shortlist on secondary-side pumps is small (Grundfos MGE, Wilo Stratos, Xylem Lowara ecocirc, Armstrong Design Envelope) and the drive is bundled into the pump listing.
Primary-side pumps remain on the catalogue-LV architecture. External VFDs from ABB, Danfoss, Yaskawa and Siemens pair with the pump skid. The primary pump is bigger, the induction motor dominates on cost grounds, and the facility-loop ΔP control problem is well-served by an external VFD with real-time pressure tracking. The architecture has not collapsed on the primary side yet. A broader shift becomes more plausible if the facility bus moves DC, which externalises the rectifier stage of that same VFD, or if IE4 and IE5 efficiency requirements and pump economics accelerate the migration toward PM-synchronous motors with integrated inverters.
The important observation for the category is that liquid cooling is pre-collapsing the drive form factor one socket at a time, inside the pump, before the facility bus does the same thing wholesale. The secondary-side CDU pump is a working proof of concept for the DC-native inverter-card architecture at a smaller scale.
6. The permanent-magnet motor arithmetic reframes the soft-starter question
The question of whether a given socket uses a soft starter, a VFD or direct-on-line depends on the motor type before anything about the duty profile matters. The soft starter is an AC motor starting device based on controlled thyristor conduction, and the thyristor action limits the magnetising inrush that an induction motor exhibits at startup. A permanent-magnet synchronous motor requires controlled commutation from zero speed, which an inverter produces. The induction motor's magnetising-current problem is specific to that motor type. The motor-type decision has already made an inverter-based starting architecture necessary, before any duty-profile optimisation is considered.
That changes how to think about affinity-law savings in the data-centre cooling plant. If the pump is PM-synchronous, the inverter ships regardless of whether the duty profile is flat or variable. The affinity-law argument matters only on induction-motor sockets where the operator can genuinely choose between a soft starter (flat duty, lower capex) and a VFD (variable duty, lower opex). A growing share of electronically controlled data-centre pump architectures uses PM-ECM or permanent-magnet synchronous motors on the secondary side (Grundfos MGE, Wilo Stratos, Xylem Lowara ecocirc, Armstrong Design Envelope are the usual references), which forces the inverter on the drive-selection decision.
Soft starters survive in the induction-motor + flat-duty + no-BMS + capex-sensitive corner. Hyperscale AI facilities place high weight on efficiency, uptime and BMS integration alongside first cost, and increasingly favour PM pumps where footprint and efficiency justify the premium. All three forces push the socket out of the soft-starter corner into the inverter corner. The structural outcome is the same as the DC-bus argument: the discrete soft starter inside the hyperscale AI DC is a residual footprint; growth categories are elsewhere.
7. DC standards, protection and operational reality
The AC-fed drive has a developed standards stack. The DC-fed equivalent has product-level pieces ready and system-level pieces open. The map below reads from the motor side up to the facility interface, flagging which layers an existing standard covers and which have open application-level specification work.
The commercial opportunity is in writing the OPEN layers before the hyperscaler reference design is frozen. The AC-era precedent is instructive: once SCCR became embedded in industrial control-panel and drive-package qualification, every subsequent vendor had to meet the same bar or lose the specification. The DC-native equivalent is being written now; participation sets test methodology, qualification requirements and architectural assumptions for the follow-on sites.
The commercial boundary that gets less attention than the standards work is the chiller OEM pushback. Trane, York (Johnson Controls), Carrier and Daikin package their own semi-hermetic compressor drives inside the chiller skid, with drive, motor, compressor and controls factory-tested and warranted as a unit. Each will resist a raw 800 VDC bus input from a third-party facility AFE because it:
- Invalidates UL and CE factory certification
- Breaks the AHRI part-load performance rating
- Shifts warranty exposure from the chiller OEM to the facility stack integrator
A second operational gap gets less attention than the product-level standards. IEEE 1584 does not translate cleanly to DC arc-flash incident-energy calculations, and NFPA 70E Annex D.5 for DC is both conservative and under-adopted in industrial facilities. The field-technician base for 800 VDC busbar maintenance is small. Plant technicians certified on 400V and 480V AC do not transfer directly to high-energy DC without additional training, procedure development and arc-flash PPE specification. The first hyperscalers to adopt facility DC will spend disproportionately on training and commissioning before the architecture becomes routine.
8. Vendor positioning: both ends of the DC stack win
The competitive map of the DC-native data centre is organised around a single question. Can the vendor sell both the facility AFE and the drive catalogue, as one commercial relationship into the same customer. The vendors that can do this are a narrow set. Vendors whose exposure remains concentrated in standalone motor-side drives face a different structural risk.
ABB, Siemens, Schneider Electric and Delta each have different combinations of industrial drives, UPS and power-conversion portfolios, and data-centre electrical-infrastructure businesses. The capability matrix below separates the layers; one label would conflate them. The DC-native ambition is that one commercial relationship covers all six rows across all six capabilities; the current reality is that each vendor has partial coverage, and the race is to close the gaps. ABB's current 800 VDC portfolio explicitly bundles power conversion, DC distribution, DC protection and cooling optimisation into one offer. The architectural convergence this essay is pointing at is already in a shipping vendor catalogue aimed at hyperscale AI today.
The drive-focused vendors (Danfoss, Rockwell, Yaskawa, Nord, KEB, Lenze) have a different exposure story. These names have strong catalogue positions and differentiated capabilities in particular drive applications. The architectural question is what they do when the hyperscale buyer standardises on an AFE-plus-inverter-card specification. The options are partnering into someone else's facility-power stack or being designed out of the DC-native data-centre socket. Rockwell presents a different exposure profile: a strong automation and control position, with less breadth across the facility-power stack than the large European electrical majors.
Inovance is the interesting Chinese position. Its EV charging and ESS businesses have built DC-fed inverter experience at scale, which gives it direct architectural read-across to DC-native data-centre drives. The challenge outside China is less about power-electronics capability than about hyperscale qualification, customer trust, installed-base weight, service infrastructure and specification acceptance. Those obstacles can take years to clear and are the right place to watch for signals.
Grundfos, Wilo, Xylem and Armstrong occupy a different kind of top-right position. These are integrated-inverter pump vendors. The secondary-side CDU socket is theirs already. In a DC-native architecture, they continue to own the secondary pump drive, with the question of whether primary-side induction pumps also migrate to their IE4/IE5 PM pump catalogues.
The architecture divides vendors into distinct camps. The full-stack integrators (ABB, Siemens, Schneider Electric, Delta) own both the medium and low-voltage power-electronics layer and the drive catalogue, and they capture the expanded margin pool as rectification, DC protection and inverter content redistribute. The pure-play drive and motor vendors (Danfoss, Yaskawa, Rockwell on the drive side; Nord, KEB, Lenze on the motion side) have portfolios concentrated in standalone LV AC drives and face a share-of-wallet contraction as rectification moves upstream, as the secondary loop integrates on the pump OEM side, and as the AFE service contract becomes the anchor for the broader electrical service relationship. Where each name lands by the end of the decade depends on which pools they build, buy or partner into between now and the first hyperscale DC-native greenfield.
9. The service pool concentrates and moves upstream
The installed-base service pool in the current drive category is a distributed field-work business. Each VFD cabinet is a service touchpoint. Fleet contracts cover the chiller VFDs, the cooling-tower fan drives, the pump VFDs, and the handful of soft starters. The vendor's regional service network dispatches technicians, stocks spare parts across hundreds of SKUs, and maintains the installed base through its 15 to 25 year life.
In a DC-native facility, the service pool restructures. The inverter card next to the motor is simpler, smaller and (for integrated pumps) replaced as a unit; field service goes away. Field-service hours per MW on the drive side fall. The facility AFE is a different asset class:
- Fewer units per site
- Higher per-unit value
- Longer service contracts, measured in SLA credits + uptime guarantees instead of hourly billing
The AFE service pool is the kind of contract UPS and switchgear businesses have long sold.
The commercial implication for the drive vendors is that the AFE service contract becomes a natural anchor for the broader electrical service relationship. Where the AFE vendor can offer a bundled contract covering the AFE, the DC switchgear and the downstream inverter cards, a hyperscale facility manager has little reason to run a parallel field-service relationship purely for the drive side. Vendors who built their installed-base economics on local drive-service relationships face a share-of-wallet risk that is bigger than any one socket decision.
10. What to watch
The signals compound:
- Hyperscaler publishing a facility-wide DC reference architecture accelerates the standards work
- A Western Inovance win forces the catalogue vendors to respond
- The first greenfield vendor selection sets the de facto reference stack for the follow-on cycle
Watching all four in parallel is cheaper than acting on any one in isolation.
11. The through-line
The drive catalogue of the next decade starts to look like a facility AFE plus a fleet of integrated inverter cards and PM-ECM pumps inside AI data centres. The three-stage AC VFD cabinet survives on the retrofit side of the installed base and across the industrial economy outside AI data centres. The discrete soft starter is a residual code-driven socket class inside the hyperscale AI data centre. None of this rewrites the drive category. It just moves money around inside it and adds one new revenue pool, facility AFE service, previously absent from the drive catalogue.
Most of the mechanical capacity added in the next five years is brownfield expansion, with pure greenfield a smaller share. A hyperscaler adding a liquid-cooled AI hall alongside a legacy air-cooled plant runs two architectures in parallel: the legacy mechanical plant on AC with existing VFD cabinets, and the new hall on whatever DC-native or DC-adjacent architecture the vendor stack supports. The common transitional step is a local AFE serving the AI hall inverter cards and integrated units, with the facility backbone remaining AC. Drive-vendor exposure plays out slowly even after the architecture is proven at the first greenfield site, because the installed-base retrofit cycle runs 15 to 25 years. The greenfield wins are the leading indicator; the existing fleet is the lagging tail.
For investors: only a handful of vendors stand to win from this, and the real exposure is bigger than the revenue mix shows. For operators: the vendor stack you pick at your next new-build site stays with you for 15 to 25 years, and the bar is rising faster than any product brochure will tell you.
The companion motor-control TCO calculator now includes a DC-fed inverter (DCI) scenario alongside DOL, soft starter and VFD. Default DCI capex runs at roughly 60 to 70 percent of VFD $/kW with a facility AFE allocation adjustable from $0 to $25/kW. Toggle the inputs to see where DCI changes the lowest-cost pick on your motor and duty profile.