Data Centres and the Cloud
AI and the Demand for Compute
How training and running AI models require vast numbers of specialised chips, why this is driving a data centre boom, and India's IndiaAI Mission to provide compute.
Artificial intelligence needs enormous computing power, called compute.
GPUs
- AI uses graphics processing units (GPUs), mostly from Nvidia.
- AI servers use far more power per rack than normal servers.
Training vs inference
- Training a large AI model can take thousands of GPUs for months.
- Inference: answering users’ questions uses compute every time.
Global boom
Tech giants plan to spend hundreds of billions of dollars a year on AI data centres.
IndiaAI Mission
- Approved in March 2024 with about 10,372 crore rupees.
- Provides subsidised GPU access to startups and researchers, with tens of thousands of GPUs contracted from private providers.
- Supports Indian foundation models.
Opportunity for India
- Lower costs and a big talent pool.
- Global firms like Microsoft, Google and Amazon announced large AI data centre investments in India in 2024-25.
Challenges
- GPU supply and export controls.
- Power needs.
The startup's GPU access
An Indian AI startup rents subsidised GPU time through the IndiaAI compute portal to train a model for Indian languages.
Thinking AI runs on ordinary computers
It needs specialised chips and huge data centres.
Key takeaways
- AI needs vast compute using GPUs.
- Training and inference both use data centres.
- The IndiaAI Mission (2024) subsidises GPU access.
- Global firms are investing in Indian AI data centres.
No recording for this one yet - EconReader can read it aloud for you.