E2E Networks proposes to expand capacity by around 5.4 MW for NVIDIA HGX B300 systems and other GPUs - CNBC TV18 - LinkedIn

5.4 MW is not a capacity figure—it is a territorial claim in India's AI compute landscape. Technically, B300's per-node power draw jumps roughly 40% over H200. At 5.4 MW that translates to 80–100 node clusters, forcing a migration from cold-plate to immersion cooling and a shift toward 800V DC power delivery. Upstream, TSMC's 4nm allocation and SK Hynix's HBM3E supply cadence will dictate actual delivery timelines, not E2E's internal roadmap. On risk, E2E's NVIDIA bet is a CUDA lock-in play, but single-vendor dependency is a single point of failure under export-control escalation. The 2023 A100 restriction proved a 72-hour supply shock can paralyze an entire pipeline. India's summer grid deficit adds a hidden operational cost most analysts overlook. Competitively, Reliance, Adani, and Yotta are all racing to build AI data centers. 5.4 MW is modest in absolute terms, yet in India's power-approval regime, first-to-secure-the-allocation carries quasi-exclusive value. NVIDIA leverages this to cement its reference architecture and squeeze AMD's MI350X window. 12–24 month outlook: inference workloads will overtake training as the dominant GPU consumer. Power costs will breach 45% of TCO. Power before silicon replaces chips before everything—grid capacity, not GPU count, becomes the primary KPI for data center competitiveness.