AI is getting cheaper to run, but powering its growth is getting pricier

AI models are becoming cheaper to run, but rising demand for data centres, computing and electricity is pushing up the infrastructure cost of AI, including in India

Artificial intelligence is becoming cheaper and easier to deploy, but the infrastructure needed to run it at scale is moving in the opposite direction. As companies use AI for more tasks, demand for computing capacity, data centres and electricity is rising. Gartner expects worldwide spending on AI-optimised infrastructure as a service (IaaS) to reach $42.3 billion in 2026, up 96.4 per cent from 2025. At the same time, countries are having to plan for the additional electricity demand that will come from the expansion of AI data centres.

 

For India, the challenge is crucial as the country expands its data-centre capacity and AI computing infrastructure. The Ministry of Power has projected an additional 26.3 GW of electricity demand from data centres by 2031-32, according to a recent clarification reported by ETEnergyWorld. Of this, applications for about 17 GW have already been submitted to state transmission utilities, while another 9.3 GW has yet to enter the formal grid-connection process.

 

Cheaper AI can mean more AI use

 

The falling cost of running individual AI tasks does not necessarily mean lower overall infrastructure demand. In fact, lower costs can encourage businesses to use AI more frequently and across a wider range of applications. If an AI query becomes cheaper to process, companies may run more queries. As AI tools become affordable enough to be integrated into customer service, coding, search, analytics and other workflows, the number of tasks handled by AI can increase rapidly.

 

This is reflected in the growth of AI infrastructure spending. Gartner expects total IaaS spending worldwide to grow 29.3 per cent in 2026, compared with a 96.4 per cent increase in AI-optimised IaaS spending. Gartner expects spending on AI-optimised IaaS to rise from $21.5 billion in 2025 to $42.3 billion in 2026, before reaching $66.1 billion in 2027.

 

A key driver of this growth is inference, the computing required when trained AI models generate responses or perform tasks. Gartner expects global spending on inference to reach $23.3 billion in 2026, exceeding the $19 billion projected for AI training. As companies move from developing models to using them continuously across business operations, every query, automated workflow and AI agent task adds to the demand for computing capacity.

 

The infrastructure supporting these systems includes high-performance computing equipment, networking and data-centre capacity. It also requires reliable electricity and cooling systems to keep servers operating. 

 

Data centres are becoming an electricity issue

 

The impact is already visible in markets where data-centre construction has expanded rapidly. Reuters reported last week that the growth of data centres in the US state of Virginia, the world's largest data-centre market, has increased pressure on Dominion Energy's power system. The utility's fuel costs are projected to reach $4.35 billion through June 2027, about 88 per cent higher than the $2.31 billion recorded in 2021, according to filings cited by Reuters.

 

The utility also expects to purchase 23 per cent of its energy supply from the wholesale electricity market by 2027, compared with 14 per cent in 2021. Reuters reported that regulators and consumer advocates have raised concerns that rapid data-centre-driven load growth is increasing exposure to volatile wholesale electricity prices.

 

The Virginia example illustrates the broader infrastructure problem. Data centres can bring investment and economic activity but their electricity requirements can also put pressure on grids and power markets when demand grows faster than generation and transmission capacity.

 

India is preparing for a much larger load

 

India's data-centre industry is also expanding rapidly. Data released by the Ministry of Electronics and Information Technology in March showed that the country's total data-centre capacity had increased from about 375 MW in 2020 to around 1,500 MW by 2025.

 

The government also said about 38,231 GPUs had been onboarded through 14 empanelled service providers and data centres under India's AI compute capacity framework. These GPUs are being made available to eligible startups, researchers and other users at a subsidised average rate of Rs 65 per hour, according to the Ministry of Electronics and Information Technology.

 

The expansion is not limited to established data-centre locations. According to ETEnergyWorld, major hubs include Mumbai, Navi Mumbai, Chennai, Hyderabad, Bengaluru, Delhi-NCR and Jamnagar, while newer investment destinations include Andhra Pradesh, Madhya Pradesh, Chhattisgarh and West Bengal.

 

This growth means electricity planning is becoming an increasingly important part of India's AI strategy. 

 

India's power requirement estimate has risen

 

The latest government clarification points to a significantly larger electricity requirement than an earlier estimate. In March, the Ministry of Electronics and Information Technology said electricity demand from data centres was estimated to reach 13.56 GW by 2031-32.

 

By August, however, the Ministry of Power had projected an additional 26.3 GW of electricity demand from data centres in 11 major states by the same 2031-32 period. ETEnergyWorld reported that the figure was nearly double the 13.56 GW estimate cited in an earlier Rajya Sabha reply.

 

The ministry clarified that around 17 GW of the 26.3 GW projected requirement already has connectivity applications filed with state transmission utilities. Applications for the remaining 9.3 GW, or roughly 35 per cent of the total projection, have yet to be submitted.

 

The government has said transmission infrastructure is being developed in phases to match growing electricity demand and planned generation capacity. ETEnergyWorld reported that 154 inter-state transmission projects were under construction, involving 39,792 circuit kilometres of transmission lines and 401,510 MVA of transformation capacity.

 

Power is only one part of the infrastructure challenge

 

Electricity is not the only resource required to support AI data centres. Cooling is another major consideration because high-performance computing systems generate significant heat.

 

The Ministry of Electronics and Information Technology said data-centre operators are adopting technologies such as direct-to-chip liquid cooling, adiabatic cooling and immersion cooling to reduce water use. Operators are also deploying high-density racks to support high-performance computing and AI workloads while seeking to reduce power and water consumption.

 

This means the cost of AI infrastructure extends beyond GPUs and servers. Operators also need cooling systems, transmission connections, backup power, networking equipment and suitable data-centre facilities.

 

The efficiency paradox

 

The central issue is that improvements in AI efficiency do not automatically translate into lower infrastructure demand. More efficient models can reduce the computing required for an individual task. But if those savings make AI affordable enough to be used across more products and business processes, total demand can continue to rise.

 

Gartner's numbers show this shift clearly. AI-optimised IaaS spending is expected to nearly double in 2026, while inference spending is projected to overtake training spending for the first time.

 

For companies, this means the cost of an AI service cannot be measured only by the price of an individual query or model API. The wider system behind that service, computing capacity, electricity, cooling and network infrastructure also has to expand as usage grows.

 

For India, the question is therefore not only how cheaply AI can be run, but whether the physical infrastructure can keep pace with demand. As AI becomes more deeply embedded in business and consumer applications, the industry's next constraint may be less about the cost of producing an AI response and more about the cost of powering the machines that produce it.