What Sonepat Took Off India's Peak

"Show me the real value AI will deliver and the time frame within which I can realise it." That is the sentence an MSME leader put on the record of PwC India's March 2026 report on AI for Indian MSMEs. It is also, pretty closely, the sentence a soap-line engineer at Hindustan Unilever's Sonepat site stopped having to ask about her own plant some time last year. On 22 June 2026, HUL announced that its Sonepat factory had joined the World Economic Forum's Global Lighthouse Network in the Sustainability category, with a 29 percent improvement in energy efficiency and a 99 percent reduction in Scope 1 and 2 emissions. The announcement sits inside a longer sequence that began at Davos in January and has not stopped putting names on the map since.

The piece worth sitting with is not the Sonepat number on its own. It is the Sonepat number read against the grid above it and the floor beneath it, both of which have moved in the same window.

What did Sonepat ship, at the meter?

Sonepat is a detergent and personal-care plant in Haryana. The 29 percent energy-efficiency figure is reported against the site's own prior consumption baseline, not against an industry average. Alongside it HUL names three distinct interventions: an AI-led energy management layer, an IoT-enabled cleaning-in-place system, and a shift to renewable supply. The press release does not disaggregate their contributions. That is the first honest limit on the headline figure. We know the plant draws fewer kilowatt-hours per unit of output than it did. We do not know how much of the 29 percent sits on top of the renewable procurement swap versus the AI layer underneath it. The CII Green Business Centre's energy benchmarking methodology for the Indian consumer-goods sector is the right reference against which an independent auditor could later check the claim; HUL has not yet published a split the methodology could run against.

The companion recognition at Haridwar is where the agentic layer gets explicit. The site runs what HUL calls a Supply Chain Nerve Centre, built on agentic AI that tracks demand signals, adjusts production, and anticipates material risk. The reported outputs are a 72 percent cut in response times and a 24 percent reduction in inventory days on hand, with an on-time-in-full delivery rate of 99 percent. Those are months of released working capital, measured on a soap-line calendar. They do not show up directly on the national grid. The energy case lives at Sonepat and Gandhidham. The capital case lives at Haridwar. The two sit under the same operations team inside HUL, which is why the press release treats them as one programme. For an external reader they are two different problems with two different test regimes, and conflating them is where the vendor pitch deck does its real damage.

What did the January 2026 round signal about the Indian subset?

On 15 January 2026, ahead of the WEF Annual Meeting in Davos-Klosters, the Forum announced 23 new Global Lighthouse sites and launched Lumina, an AI-powered industrial-intelligence platform built on eight years of Lighthouse data and more than one thousand transformations. The Indian names in that round are worth separating from the general announcement, because they are what a sector-level reader can anchor the claim against. ACG Packaging Materials at Shirwal in Maharashtra reduced energy consumption by 31 percent. Unilever Gandhidham cut Scope 1 and Scope 2 emissions by 90 percent through renewable procurement and reduced water use by 17 percent. Unilever Pondicherry raised production volume by 25 percent and cut defects by 23 percent. None of those are vendor benchmarks. Each is a site-level self-reported figure the GLN's expert-panel review process has signed off on, under the Forum's own methodology.

Shortly after, on 28 June 2026, DCM Shriram's Jhagadia specialty-chemicals site in Gujarat was admitted to the Network with a Productivity distinction. Jubilant Ingrevia had crossed the same threshold earlier, when its Bharuch specialty-chemicals plant reported a 60 percent reduction in process variability and roughly doubled production volume after thirty integrated digital use cases. Added to the HUL roster, the Indian Lighthouse count has climbed into the mid-teens. That is a corpus the specific-energy-consumption modellers at the Bureau of Energy Efficiency could, in principle, build a sector curve against. In practice, nobody has.

What is the grid above all this reading?

The macro frame is tight, and the plants are drawing their electricity from inside it. The IEA's Electricity Mid-Year Update 2026 forecasts Indian electricity demand to rebound by seven percent in 2026 after a sub-two-percent 2025, with a peak load of 270.8 gigawatts recorded on 21 May 2026 and shortages across the January-to-May window more than double the same period a year earlier. The IEA attributes a share of the rebound to stronger industrial activity. Industrial activity is what Sonepat and Shirwal and Jhagadia and Bharuch are, in the ledger.

Walk the arithmetic. If a Lighthouse site is running 29 percent lower on specific energy consumption and India's industrial draw is still rising seven percent year on year, the per-site improvement is being spent as fast as it is banked. The Lighthouse plant is not building a smaller grid for India. It is giving the central-sector dispatcher a few more megawatts of elbow room on top of a demand curve that is tightening every month regardless. That is a different claim from the one the press-release reader instinctively hears, and the honest version is the one the plant's own engineering team already knows: the badge is a floor the next five years' output growth will erode, unless the next cohort of plants runs the same programme.

Where does the disconfirming reading sit?

The strongest reading in the other direction does not come from an AI critic. It comes from the McKinsey Operations practice that helped build the Lighthouse methodology, which documented in its public report that the first Lighthouse cohort was selected from more than one thousand manufacturers scanned, picking sixteen. The selection ratio has gentled since, but the shape has held. Lighthouses are the plants that already had the data pipes, the leadership continuity, and the capital runway to pull a multi-year transformation to the finish line. They are the top of the pyramid, not a cross-section of it. Any read that treats the Lighthouse number as the ceiling of what Indian manufacturing can achieve is right. Any read that treats it as the floor the sector is operating at is wrong, and that is the read vendor marketing teams are quietly encouraging.

The PwC March 2026 report sharpens the point from the opposite end. India has roughly 1.59 crore registered manufacturing MSMEs — about 21 percent of all MSMEs, producing 35.4 percent of manufacturing value added and 48.58 percent of exports. The same report finds that 87 percent of surveyed MSMEs expect productivity gains from AI, 91 percent say the tool must be affordable first, and the top-25 listed SME manufacturers used AI-related terms twenty-nine times across three fiscal years, against 963 times for the top-25 listed large-caps. The gap sits on an invoice line, not on a vocabulary difference. The AI assistant the Lighthouse plant pays for is not the AI assistant the tool-and-die shop in Rajkot or the dye-house in Tirupur can price into its order book this quarter.

A morning on a dye-house floor

Two monsoons ago I spent a morning on a dye-house floor in Tirupur. The owner-operator was running a batch, a diesel genset was running outside against a scheduled grid cut, and his chief engineer was telling me, hands held two inches apart, the physical size of the kaizen manual that had carried the plant through the previous five years. The owner had seen the vendor pitch for an AI boiler-optimisation module two weeks before. He asked me three questions, in order. What is the payback period if the grid stabilises on its own. What is the payback period if it does not. And who runs the training on the module after my engineer retires. The vendor's slide pack answered the first in nineteen months. It did not answer the second. It did not answer the third. The owner had read the deck more carefully than the vendor had expected him to. The order was not placed.

I bring that morning up because Sonepat and Shirwal and Jhagadia are the plants where the Tirupur engineer's three questions have already been answered, by capital, by programme management, and by a corporate leadership team that could keep its promises on timeline. Those three answers are what the Lighthouse badge certifies. The engineer in Tirupur is still running his boiler on the kaizen manual.

The number that reframes this

A Lighthouse plant in India is one of roughly 220 sites in the global network, and perhaps fifteen in the country. India's registered manufacturing MSMEs sit at 1.59 crore. That is approximately one Lighthouse per one million registered manufacturing MSMEs.

Tarry Singh is the founder and CEO of Real AI, an enterprise AI advisory and deployment firm working with global enterprises on production agent systems, model risk, and AI sovereignty strategy. He also leads Earthscan for Energy AI startup, and is a founding contributor to the EU-funded HCAIM and PANORAIMA programmes for responsible AI education across European universities. He writes at tarrysingh.com.