On 21 October 2026, cloud company Vultr announced a 20 MW AI supercluster in Chennai, Tamil Nadu, built on AMD Instinct MI355X GPUs.
Vultr calls it the largest AMD Instinct deployment in India; its other large AMD AI site is in Springfield, Ohio.
It is meant for training, fine-tuning and high-throughput inference of AI models, with MI455X GPUs and AMD's Helios racks to follow in 2027.
The pitch is local compute: data residency, sovereignty and low latency for Indian users.
Training sets a model's parameters from large data. Fine-tuning adapts a pre-trained model to a task with extra data. Inference is the trained model producing outputs for new inputs. GPUs suit all three because they run thousands of calculations in parallel.
Simple Analogy: Training is schooling, fine-tuning is job training, inference is doing the job.
Build India's AI ecosystem, starting with public compute capacity
Key: Approved by the Union Cabinet on 7 March 2024 with ₹10,371.92 crore over five years; run by MeitY; its compute pillar targets 10,000+ GPUs offered to start-ups and researchers at subsidised rates
US chipmaker; makes the Instinct GPU line and the ROCm software stack
Runs the IndiaAI Mission
GS Paper 3 > Science and Technology > AI and computing
The principle that data is subject to the laws of the jurisdiction where it sits
Stacked memory placed next to the processor for very fast data transfer; HBM3E is a current generation