Meta signs 168MW AI data center deal with Reliance in India
This 168MW deployment signals a strategic shift in Meta's infrastructure topology, distributing high-density compute closer to its largest user base while leveraging Reliance's local power grid capabilities. For engineers, this means lower latency for regional inference and a crucial diversification of training clusters away from constrained US power markets.
The announcement of Meta's first AI data center in India, a 168-megawatt facility built in partnership with Reliance, represents a critical pivot in global AI infrastructure distribution.
What Happened Meta has secured a massive 168MW power envelope and facility agreement with Indian conglomerate Reliance. The data center will specifically target Meta's global AI computing requirements, rather than just standard regional cloud services, and includes provisions for future expansion.
Technical Implications A 168MW facility is a massive deployment of high-density compute. To put this in perspective, modern AI clusters utilizing NVIDIA H100 or Blackwell GPUs require significant power and cooling overhead. Assuming a standard PUE (Power Usage Effectiveness) and ~1-1.2kW per accelerator (including networking, storage, and cooling overhead), this facility has the theoretical capacity to house roughly 100,000 flagship GPUs. Delivering this at a single site requires advanced power delivery networks and direct-to-chip liquid cooling infrastructure, which Reliance and Meta will need to co-develop to handle the local climate.
Why It Matters From an engineering and infrastructure standpoint, this move solves two major bottlenecks: power availability and inference latency. The US grid is currently severely constrained, delaying new large-scale data center interconnects. By partnering with Reliance—a company with deep roots in Indian energy and telecommunications—Meta bypasses North American power bottlenecks. Furthermore, India is Meta's largest market by user volume. Localizing a massive inference footprint drastically reduces round-trip latency for AI features embedded in WhatsApp, Instagram, and Facebook, while the cluster can still contribute to asynchronous global training workloads for future Llama models.
What to Watch Next Monitor the facility's cooling architecture—specifically whether Meta deploys its Grand Teton AI server architecture and how they handle the thermal challenges of the region. Additionally, track the timeline for grid interconnection and whether this prompts competitors like Google and Microsoft to accelerate their own high-density compute deployments in the APAC region to circumvent Western power grid limitations.