The forecasts disagree by 3.0 times,
and they are counting the wrong thing.

Every one of these forecasts is stated in built capacity, which is the same word the one filed estate shows does not mean earning. It is the finding rather than a caveat on it. On the one Indian estate where a filing prints both numbers, 188.04 MW of built capacity carried 111.37 MW sold to a customer. Applying that share is not a rival forecast. It is the same published numbers restated in the unit revenue would have to come from, and it moves the top of the range from 13.5 GW to 8.0.

One number is worth stopping on. Restated, the very top of the bull range lands at 8.0 GW, which is the same figure as the floor of that same bull range as published, 8.0 GW. The most optimistic case, counted in capacity that earns, arrives at the least optimistic case counted the way the industry counts. That is a coincidence of arithmetic rather than a result, and it is the clearest measure of how much work the unit is doing.

One estate is not a national conversion rate. It is the only rate anybody has published a filing for, which is the more useful complaint: an entire sector is being sized in a unit whose conversion to revenue has been measured exactly once. That estate is read here.

2025 was the best year India has had. It added 387 MW, 2.0 times the year before. It is still short of every forecast on this page. The least demanding of them, Colliers, needs 600 MW every year from here, which is 1.6 times a record year, sustained for 5.

The straight line is deliberate. A ramp would move megawatts between years without changing the total, and the total is the claim. Nothing here says the forecasts are wrong. It says what would have to happen for them to be right, in the one unit the sector reports annually, and that the doubling already achieved is not the hard part.

All 5 Indian operators in the table fell between FY2023 and FY2024. The average went from 13.47 per cent to 8.45, a drop of 37 per cent in a single year, while the largest global operator moved the other way. Capital is going in faster than it is coming back out, across the whole Indian set rather than at one company.

The issuer's own summary of this table, at printed page 261, reads: “Sify achieved ROCE of 9.88%, 6.75% and 7.92% in Fiscal 2023, Fiscal 2024 and Fiscal 2025, respectively, which is significantly higher than its global peers.” The comparison is to global peers. Against the Indian operators printed directly above them it ranks 3 of 5 in FY2023 and 3 of 5 in FY2024. In FY2024 it did not beat every global peer either: 6.75 against 6.92.

The source gives a reason on the same page: “Sify's depreciation rose from INR 2,083.00 million in Fiscal 2023 to INR 3,268.16 million in Fiscal 2025, resulting in a depreciation rate of 10.33% in Fiscal 2025 the highest among its peers. This partly weighed on its ROCE.” Heavy depreciation in a build phase does suppress the ratio, and it is the honest explanation for a low reading. It is also the explanation every operator in the Indian half of this table could give.

This is the clearest place in the sector to watch an announcement separate from a delivery, because every one of these providers accepted the same terms on the same date. 3 are reported as deploying and 2 as behind, and the two behind are among the largest names on the list. Size did not predict delivery here, which is the result the capacity pages reach from the other direction.

The largest group is the one nobody has reported on. For 5 of the 10 there is no public statement either way, which is a fact about the scheme's reporting rather than about those providers, so they are grouped by what was reported and not ranked against each other.

Microsoft, Google and AWS have between them pledged 45.2 billion dollars to Indian data centres. The market they are pledging into turned over 10 billion in 2025. Those pledges sit inside 94 billion committed by every investor since 2019, so three foreign firms account for 48 per cent of everything the sector has been promised in six years.

Only one of the three attached a capacity to the money. Google put 1,000 MW on the first phase of a single site at Visakhapatnam. India operates 1,500 MW in total. One announced phase of one campus is 67 per cent of a national estate that took two decades to build, and it is announced rather than built. The other 2 pledges name no capacity at all, which means the largest numbers in this sector are denominated in a unit that cannot be checked against anything that exists.

Per gigawatt actually running, the pledges come to 30 billion dollars, and the whole sector commitment to 63. Capital is not what the sector is short of. The listed names sit here, and none of them is the counterparty to any of these three announcements.

The Ministry of Power puts data centre electricity demand at 13.56 GW by 2031-32, against about 1 GW today. That is 13.6 times in six years, and it is a different measurement from the capacity forecasts in the first exhibit: this is what the buildings draw from the grid, those are what the servers inside them are rated for. Both are rising by an order of magnitude and neither is delivered by an announcement.

Priced at the published build cost of 60 to 70 crore rupees a megawatt, the most conservative forecast on this page needs 3,000 MW of new capacity costing 1.8 to 2.1 lakh crore. The IndiaAI Mission, the country's flagship programme for this technology and the scheme in the exhibit above, has a total outlay of 10,371 crore. The base case alone costs 17 times the whole mission, and it is the red rule at the foot of the chart.

Electricity is 65 per cent of what these operators spend to run, on a stabilised margin near 43 per cent. The single largest recurring cost is the one input that cannot be commissioned by writing a cheque, and the transmission that carries it has a measured delay record of its own, drawn against one company's deployment schedule here.