Thirty megawatts of AI workload now runs on direct-to-chip liquid cooling in Navi Mumbai. Vertiv and CapitaLand India Trust completed the installation at CLDC Mumbai 01 in Airoli on 16 October. Vertiv calls it India's first high-density liquid-cooling deployment at this scale.
The number is small against India's buildout. It is large against the country's experience. Air-based systems still took 65.85% of Indian data-centre cooling spending in 2026, according to Mordor Intelligence. Liquid cooling adds roughly $1.3 million per megawatt, a premium Sify's group CFO put on the record this year.
Adoption is no longer the question. Margin is. Global integrators bring reference designs, commissioning crews and balance sheets. Indian manufacturers bring factories, lower logistics costs and a domestic service model.
Both models now have a reference project. The next 24 months decide which one travels.
The integrators draw first blood
Vertiv supplied the secondary fluid network, the skids that carry coolant between the facility loop and the server racks. Its teams balanced the technology loop with SFN-Tab kits, then ran Level 4 testing before the hall went live. The end customer is a hyperscale operator that Vertiv did not name.
Commissioning is the hard part. A direct-to-chip system fails on dust, pressure imbalance and coolant chemistry long before it fails on cooling capacity. Vertiv built installation procedures with CapitaLand Investment specifically to keep dust out of the technology loop while the skids went in.
The CapitaLand Investment project, particularly the collaboration, helped us accelerate the development of our manufacturing and engineering capabilities.— Subhasis Majumdar, managing director, Vertiv India
The secondary fluid network is the part buyers underrate. It sits between the facility water and the rack manifold, so it owns the interface where most leaks and pressure faults appear. A vendor that controls that layer also controls the service contract that follows.
Vertiv is not alone. Schneider Electric, LG and FläktGroup have all announced high-density cooling products this autumn. FläktGroup produced its first 2.5 MW DX coolant distribution unit on 28 October for a US customer, at its Arkansas plant.
The incumbent advantage is procedural. Listed vendors carry audited service networks, spare-parts depots and insurance that a 30 MW operator can underwrite. That matters more than a lower invoice when the workload is a training cluster.
India wants the factory, not the invoice
KühlTherm is the counter-case. The Ahmedabad company launched locally made rear-door heat exchangers and coolant distribution units on 7 August, after a $1.1 million seed round led by Arkam Ventures. Its RDHx scales from 20 kW to 100 kW per door. Its in-row CDU scales from 250 kW to 1.5 MW.
The pitch is vertical integration. KühlTherm designs and builds the thermal loop from chip to chiller inside India. It runs what it describes as the country's only integrated liquid-cooling validation lab, and it plans to lift manufacturing capacity from 300 MW to 2 GW by 2027.
KühlTherm's cold plates cover 300 W to 2,500 W per plate, the range where current accelerators sit. A 2 GW manufacturing target is a supply-chain claim. The gap between that claim and a shipped, certified product is the investment risk.
The founders came out of Ola Electric, which is a useful signal. EV thermal management is a volume manufacturing problem, and Ola built battery packs and home energy systems at scale. Whether that experience transfers to data-centre reliability is the open question.
KühlTherm also lists immersion modules alongside direct-to-chip hardware. Immersion handles higher densities but complicates maintenance and coolant supply. A supplier that offers both can bid on retrofits and new builds with the same sales team.
KühlTherm's own count puts India's installed data-centre capacity at 1.6 GW in October. Treat that as a company estimate. The direction is not in dispute. Tata Power has outlined a data-centre power push measured in gigawatts, and Mordor counts Google, Sify, Equinix and STT GDC among the operators adding liquid-ready capacity.
The number that decides the split
India's data-centre cooling market was worth $2.99 billion in 2026. Mordor Intelligence projects $3.75 billion in 2027 and $11.38 billion by 2032, a 24.86% compound annual growth rate. Liquid cooling grows faster than the market it sits in: 26.7% a year through 2032.
Base-year cooling spend
Air and liquid systems combined, before the liquid retrofit cycle. · Mordor Intelligence, 2026
The per-megawatt math explains the urgency. The colocation operator's group CFO put the liquid-cooling premium at about $1.3 million per megawatt. On a 30 MW hall, that is roughly $39 million of thermal equipment. On a 2 GW pipeline, the same arithmetic becomes a multi-billion-dollar procurement decision.
Cost of the thermal upgrade
Sify's group CFO estimate for adding liquid cooling per megawatt of capacity. · Sify, 2026
Air cooling is not disappearing. It held 65.85% of 2026 spending and still cools most of the installed base. The split moves at the top of the density curve, where accelerators push racks past 100 kW and air stops being an option.
| Decision | Global integrator | India-built |
|---|---|---|
| Design and commissioning | ✔ Reference designs, L4 testing | ◐ In-house lab, fewer references |
| Local content | ✗ Imported skids and CDUs | ✔ Chip-to-chiller in India |
| Cost per MW | ◐ Integration fee premium | ✔ Lower labour and logistics |
| Bankability | ✔ Listed, audited, insured | ✗ Early-stage balance sheet |
Qualitative comparison from company disclosures and Mordor Intelligence data, October 2026.
Read the table as a risk map. Incumbents win on bankability. Local suppliers win on cost and response time. The hyperscale buyer optimises for uptime. The domestic colocation buyer optimises for capex per megawatt.
What would move the split
Three variables decide whether India's thermal bill stays imported. Certification comes first: OCP-style qualification and insurer acceptance for locally built CDUs. A reference project above 10 MW comes second. Localisation speed comes third, and it cuts both ways.
Probability: 60% — the 30 MW reference removes the "no local precedent" objection, and a $1.3M per MW premium gives operators a reason to shop locally.
As we wrote in our piece on 800-volt DC power, the binding constraint in AI infrastructure keeps migrating. It moved from chips to power delivery. It is moving now to heat rejection. The vendor that owns the interface between power and cooling owns the next procurement cycle.
Three ways the next 24 months land
🟢 Local manufacturing breaks through (55%)
Consequences: Local content passes 25% of thermal spend and the $1.3M per MW premium compresses for domestic buyers.
🟡 Two-tier market settles (30%)
Consequences: Imports stay dominant by value while local unit volumes grow, which protects incumbent margins.
🔴 Certification slips (15%)
Consequences: Imports extend their lead, and the 2 GW manufacturing target slips beyond 2027.
For investors, the cleanest exposure is the interface, not the box. Service contracts, coolant supply and commissioning labour carry better margins than sheet metal. Vertiv already sells that bundle. KühlTherm is trying to build it.
The 30 MW site does not settle the contest. It gives both sides a number to argue over. Vertiv can point to a live hyperscale hall. KühlTherm can point to a 2 GW manufacturing plan and a domestic cost base. Watch the next award above 10 MW. That order will say more than any launch event.