Artificial intelligence is becoming cheaper to operate at the individual-task level, but the infrastructure required to support its rapid adoption is moving in the opposite direction. As AI becomes affordable enough for companies to embed into customer service, coding, search, analytics and automated workflows, overall usage is increasing. Gartner expects global spending on AI-optimised infrastructure-as-a-service to reach $42.3 billion in 2026, up 96.4% from $21.5 billion in 2025.
The biggest shift is occurring in inference, the computing required when AI models actually generate responses or perform tasks. Gartner estimates global inference spending will reach $23.3 billion in 2026, exceeding the $19 billion expected for AI training. This is important because the economics of AI are moving from a relatively small number of enormous training runs toward continuous, large-scale usage. Every chatbot query, AI-agent action and automated business process creates additional inference demand.
That creates a major electricity and infrastructure challenge. In India, the Ministry of Power has projected an additional 26.3 GW of electricity demand from data centres by 2031–32, with applications for around 17 GW already submitted to state transmission utilities. India's data-centre capacity has meanwhile grown from roughly 375 MW in 2020 to about 1,500 MW in 2025. Data centres also require substantial cooling, networking, backup power and transmission infrastructure, meaning the physical cost of AI extends well beyond GPUs.
The article identifies an important efficiency paradox: making AI cheaper does not necessarily reduce its total resource consumption. If the cost of an AI task falls, businesses may simply perform far more tasks with AI, causing aggregate compute and electricity demand to rise. For India, this means the AI race is increasingly also a power-grid and infrastructure race.
