Powering India’s AI Compute: Why 300kW Racks Need a New Design

India’s AI story is moving past standard cloud applications. We are now building high-performance computing capacity made for AI, and policy is backing it. The IndiaAI Mission, with a ₹10,372 crore allocation, is setting up shared computers so startups, researchers and enterprises can reach tens of thousands of GPUs. Data centers now have Infrastructure Status, which has eased access to credit and brought in long-term investment. Operational capacity stood at 1.6 GW in mid-2026 and is projected to reach 6 GW by 2029.

So, what does that mean inside the data center? We are no longer designing for rows of CPU servers. We are designing power and cooling together for dense GPU clusters.

How fast rack power is rising

Not long ago, 10kW to 30kW per rack counted as high density. AI has moved that line. This is the range we see today:

• 20kW: traditional enterprise IT
• 60kW: high-density cloud and analytics
• 150kW: AI training and GPU deployments
• 300kW: large AI clusters
• 600kW+: next-generation AI factories and supercomputing centers

At 300kW and above, conventional AC power and air cooling start to struggle. Here is where the old design breaks.

Four problems

First, current. Delivering 300kW to one cabinet at 400V AC takes hundreds of amperes, which means heavy copper busbars and thick cables. They crowd the rack, block airflow, add floor load and push up cost.

Second, losses before the server. Energy lost in a conductor rises with the square of the current. Push high current through low-voltage paths and power turns into heat before it even reaches the server. If you are working to lower PUE, that hurts.

Third, too many conversions. The usual chain steps down utility voltage, converts AC to DC for batteries, converts back to AC for distribution, then converts to low-voltage DC at the chip. Every step loses energy and needs more equipment, and that equipment takes space you could use for revenue-generating IT hardware.

Fourth, air has a limit. Air cooling tops out at roughly 30kW to 50kW per rack. Beyond that, fan power, noise and hot spots go up, and the processors throttle. You get less computing for the power you pay for.

What works at 300kW and above

So what do we do? We look at power and cooling as one system.

Start with power. Moving rack distribution to 800VDC lowers the current needed for the same power. That reduces conductor losses by up to 40%, removes redundant AC-DC conversion steps, frees up rack space and makes better use of copper.

Then cooling. Direct-to-chip liquid cooling puts cold plates right on the GPUs and CPUs, removing up to 80% of the heat at the source. Coolant distribution units and heat exchangers manage fluid temperature and flow in the secondary loop, so a facility can run at peak load continuously with PUE close to 1.1 to 1.2.

Then speed. Operators are building on tight timelines, so we use prefabricated Power PODs. Each one brings together UPS systems, medium-voltage switchgear and intelligent power distribution as a pre-tested block. There is less integration risk on site, build time comes down from years to months, and you can add capacity in steps as you grow.

Finally, the grid. AI training can cause sharp jumps in power demand, which stresses the facility and the local grid. Battery energy storage, active power factor correction and microsecond-level telemetry help smooth peak load, keep power quality clean and bring site-level renewables into the main power system.

Working with operators

If you are building or running a data center in India, you need a partner who understands both high-voltage power electronics and thermal control. Delta Electronics India brings global engineering experience in high-density power, along with local manufacturing, R&D and application support. We work with operators from facility design and thermal modelling through execution and lifecycle management, so the move to AI-ready architecture is safe, efficient and commercially scalable.

Going from 20kW enterprise racks to 300kW+ AI factories is a structural change in how digital infrastructure is built. With high-voltage DC, direct liquid cooling and modular power, we can give India’s compute ecosystem a foundation that delivers hyperscale performance while keeping energy use in check.