Delhi’s distribution network was once among the leakiest in the world. A quarter-century of reform has made it comparable with western Europe. A parallel shift in computing, from training AI models to running them at scale, explains why that achievement matters more than ever.
In 2002 more than half the electricity fed into Delhi’s distribution network was never paid for. Some of it vanished as heat in overloaded wires and failing transformers; much of it was stolen through illegal hooks or lost to broken meters and unpaid bills. Daily power cuts lasting hours were routine. Today the capital’s combined technical and commercial losses stand at five to six per cent, in the same range as France and Belgium, and its supply reliability exceeds 99.9 per cent. The story of how that happened is told in the October issue of IEEE Spectrum by Mini Shaji Thomas, a power systems professor at Jamia Millia Islamia who lived through it.
A grid on the brink
The numbers from the starting line are stark. When the Delhi Vidyut Board was broken up in 2002, the private distributor now known as Tata Power-DDL inherited losses of about 53 per cent in north Delhi. BSES, now part of Reliance Infrastructure, faced 51.5 per cent in South Delhi and 63.1 per cent in East Delhi, and some pockets reached 83 per cent. Utilities were collecting payment for less than half the power they supplied. About 11 per cent of transformers failed each year. Tata Power-DDL’s chief executive told the magazine that the company took over more than 100,000 unresolved billing complaints.
The Electricity Act of 2003 supplied the legal framework: it unbundled generation, transmission and distribution, opened the sector to private operators and created mechanisms to prosecute theft.
How the losses came down
No single technology did the job. Supervisory control and data acquisition systems, known as SCADA, gave operators a live view of the network and the ability to switch equipment remotely. Hundreds of capacitor banks, some of them mobile, were installed to supply reactive power and steady the voltage, cutting the current that lines had to carry. Bare overhead conductors were replaced with insulated bundled cables that are difficult to tap illegally.
Metering advanced in steps: electronic meters in 2003, automated reading in 2004, data analytics on meter readings from 2006 and smart meters from 2017. The transformer failure rate fell below one per cent.
The social measures were as important as the hardware. Tata Power-DDL recruited women from 223 slum clusters as local agents to collect payments, alongside literacy and financial-literacy programmes. Payment rates in those neighbourhoods now match the rest of the city. Round-the-clock kiosks, online payment and police support against organised theft completed the picture. As the chief executive of BSES Rajdhani Power told the magazine, technology enables, but lasting gains depend on execution, accountability and engagement with consumers.
The result was achieved while demand soared. Delhi’s peak load reached 8,748 megawatts this year, roughly three times the 2002 level.
Unfinished business
The turnaround has limits. Illegal charging of e-rickshaws still drives high losses in some areas. Smart meters are not yet universal. And Delhi remains an exception: the article notes that losses in several large Indian states still run at roughly 17 to 23 per cent. Delhi’s dense, relatively affluent consumer base and sustained regulatory attention are not easily copied in a state with long rural feeders and subsidised farm supply. Even so, the model has drawn interest from countries with similar problems, from Bangladesh to Kenya.
The other revolution: AI moves from training to inference
The same issue of IEEE Spectrum carries a feature on a change in computing that will bear on every grid operator. For several years the heaviest demand for AI hardware came from training, the process of building a model. In 2026, the magazine argues, the centre of gravity has moved to inference, the act of using a trained model to answer queries, write code or run autonomous software agents around the clock.
Inference stresses hardware differently. Processing a user’s prompt can be done in parallel, which suits graphics processors. Generating the reply happens one word-fragment at a time, and each step requires reading the model’s full set of numbers from memory. The bottleneck is therefore memory speed, not raw calculation. Researchers cited in the article found that high-end graphics processors running open language models sat idle for 50 to 80 per cent of the time during this phase.
That inefficiency has opened the field. Nvidia has reportedly paid about 20 billion dollars for the technology and staff of the start-up Groq, and now pairs its graphics processors with specialised chips that hold their memory on the same piece of silicon. Amazon Web Services combines its own processors with wafer-sized chips from Cerebras. Smaller firms are stacking processors directly on memory, or designing chips that can run only one family of AI model but do so at extraordinary speed. Some of the performance claims come from the companies themselves and await independent verification.
Why the two stories belong together
Faster, cheaper inference means more AI use, and more use means more electricity drawn continuously by data centres, a load quite unlike the evening peak of a residential city. India is courting that investment vigorously, with large facilities planned around Mumbai, Chennai, Hyderabad and the National Capital Region.
A data centre needs what Delhi spent 25 years building: a distribution network that delivers nearly all the power it receives, with stable voltage and almost no interruptions. A grid that loses a fifth of its energy cannot sensibly host loads of that scale, and the cost of the waste falls on other consumers.
The lesson of the Delhi experience is that such reliability is not purchased with a single technology. It comes from metering every unit, fixing accountability on the operator, investing steadily in unglamorous equipment and persuading customers that paying for power is worthwhile. As India’s states compete for the computing industry of the coming decade, the quality of their distribution grids may prove as decisive as land or tax incentives.
