Multi-gigawatt AI training capacity at the rack-density, power architecture, and cooling design the Vera Rubin Ultra generation requires. 600 kW per rack. 800VDC distribution. Direct-to-chip liquid cooling. Designed for sustained training runs, not for the workload assumptions of the previous decade.
Rack-scale AI platforms have changed the operating point of data center design. NVIDIA Vera Rubin Ultra (NVL576) is specified at 600 kW per rack at production. AMD Helios MI450 follows close behind. Most existing Indian capacity was designed for a 5-20 kW rack envelope.
The gap is structural, not incremental. A facility designed for 10 kW per rack cannot host a 600 kW rack by adding capacity. The power distribution architecture is wrong. The cooling architecture is wrong. The mechanical and electrical infrastructure is wrong. Retrofitting is, in practice, demolition and rebuild.
For an AI lab planning a multi-billion-parameter training run in 2027 or 2028, the choice of training infrastructure is the choice of architecture: 800VDC or 415VAC; direct-to-chip liquid or evaporative air; AI-native or legacy retrofit; large unified campus or distributed micro-facilities.
HyperNext made these architectural choices at design time. The Kakinada AI Factory is 1.2 GW IT load, AI-native from foundation slab. The Hyderabad facility opens in December 2026 with the same architecture at smaller initial scale.
Four design choices, made at the architectural level, that distinguish this solution from a re-packaged commodity hosting offer.
Rack envelopes specified for sustained 600 kW operation, matching the Vera Rubin Ultra NVL576 production specification. Mechanical and electrical infrastructure sized to support the rack density at the campus level. Design basis documented in HN-RP-002.
Higher-voltage DC distribution reduces conversion losses, lowers PUE, and improves resilience at AI rack density. Sourced from 700 MW of captive solar at Khavda plus partnership wind. The arithmetic is in HN-RP-002.
Coolant-distribution architecture designed for the heat reject loads of NVL576-class systems. Closed-loop glycol coolant primary, sealed and recirculated. Heat rejected by dry coolers only, with no evaporative cooling. Cooling architecture in HN-RP-006.
1.2 GW IT load at Kakinada AI Factory (Q1 2028 first phase); 250 MW Hyderabad (initial 64 MW Dec 2026); 100 MW Nava Raipur DRaaS site. Anchor tenant capacity available now for 2027-2028 commissioning windows.
A worked example for a frontier AI lab training a multi-billion-parameter model in 2027.
The architectural choices on this page are documented in the HyperNext Research series. Methodology is published openly so that customers can verify the engineering claims and so that other operators can run the same analysis on their own facilities.
A 30-minute conversation with our business development team, oriented to your specific workload, regulatory requirements, and deployment timeline. No pricing reveals, no over-promised SLAs. Just a working conversation about whether HyperNext is the right fit.