AICassindra
Research review / india

India's AI Compute Pipeline in One Week: Vultr Plans a 20 MW AMD MI355X Cluster in Chennai and TCG's Cortex DC Targets 2 GW -- Both Are Announcements, Not Capacity Yet

Two private announcements, on October 21 and 23, add to India's sovereign-compute story without dates, sites or prices. The useful questions for builders are what a 20 MW AMD cluster can actually serve, how ROCm fits an existing CUDA stack, and why India still has no official count to measure any of this against.

By Sandeep Belgavi · Published 2026-10-24

Two Indian data-centre announcements landed this week. Neither is operating capacity yet, and both should be read that way.

Vultr and AMD in Chennai

On October 21, 2026, Vultr said it "will deploy AMD Instinct MI355X GPUs in a new 20 MW supercluster in Chennai, India, creating the country's largest AMD Instinct deployment". Each MI355X carries "288GB of HBM3E memory and 8TB/s of memory bandwidth". The cluster will run AMD's ROCm software stack, and Vultr plans "significant additional capacity expansion in 2027, including AMD Instinct MI455X GPUs and the AMD Helios rack-scale AI architecture." CEO J.J. Kardwell framed it as sovereignty: the cluster "gives organizations in India the capacity to train, fine-tune, and run inference on in-country infrastructure." The release gives no date for availability, no pricing and no GPU count, and the "largest" claim is Vultr's own.

Cortex DC

On October 23, Indian Infrastructure reported that Cortex DC, the "artificial intelligence (AI) first subsidiary of The Chatterjee Group (TCG)", plans "2 gigawatts (GW) of capacity across India" for hyperscalers, cloud providers and AI workloads. It targets "operational readiness for its first data centre within 18 months of first shovel" and aims to "source 51 per cent of its energy from green sources". No sites, investment figure, partners or first-shovel date were given. TCG's chairman, Purnendu Chatterjee, offered the slogan: "The industrial age rested on steel and energy; the intelligence age will rest on compute."

What 20 MW of MI355X means for a model team

The following is our own arithmetic, not a vendor figure. Eight MI355X GPUs in one node hold about 2.3 TB of HBM. A 3-trillion-parameter model, the class Moonshot says it will release as open weights by November 4, needs about 3 TB for weights alone at 8-bit precision, so it would span at least two such nodes before any KV cache for long contexts. At 4-bit, the weights (about 1.5 TB) fit in one node with room for cache. That is the practical case for high-memory accelerators: frontier-scale open models served from fewer nodes, on infrastructure that stays in India. The GPU count behind 20 MW depends on whether that figure is IT load or total facility power, which the release does not say.

The software question matters as much as the hardware. Teams whose kernels, quantisation tools and serving stacks assume CUDA should budget time to validate them on ROCm, test the exact model and precision they plan to run, and measure tokens per second per rupee rather than relying on peak FLOPS.

The missing baseline

These announcements arrive as the IndiaAI Mission's compute agency, C-DAC, seeks empanelment for up to 10,000 more AI compute units. Earlier this month, a parliamentary committee report recorded MeitY saying it keeps no official figures on national data-centre capacity. This digest covered both on October 15. Without a public count of what is energised, contracted and available, it is hard to tell whether 20 MW or 2 GW changes what Indian startups can actually rent.

What remains uncertain. Commissioning dates for Chennai, pricing in rupees, whether capacity will be offered through IndiaAI's subsidised compute portal, and every site-level detail of Cortex DC's plan.

Analysis: India's compute story is shifting from GPU counts to megawatts and gigawatts, and that is the right unit, because power and cooling are the binding constraints. But a megawatt announced is not a megawatt energised. Founders should plan on capacity they can contract now, ask any provider for a delivery date and a committed price, and keep their stack portable between NVIDIA and AMD so a new regional cluster is an option, not a rewrite. Policymakers could help most by publishing the capacity register that does not yet exist.

Sources and evidence

Vultr said on October 21 it will deploy AMD Instinct MI355X GPUs (288GB HBM3E each) in a 20 MW Chennai supercluster with MI455X and Helios capacity planned for 2027, and TCG's Cortex DC said on October 23 it plans 2 GW of AI data-centre capacity with the first site ready 18 months after groundbreaking; neither gave dates, prices or sites, and India still lacks an official data-centre capacity count to measure them against.