India Just Published the Number Everyone Else Leaves Out

Of the 633 AI deals in our tracker, 26 say how much electricity they will consume. That is 4.1 percent. This week's addition came from Hyderabad.

Tata Consultancy Services' subsidiary HyperVault announced a data center campus of up to 1GW on 264 acres near Hyderabad, with partners expecting to invest up to 700 billion rupees, roughly $7.4 billion, across multiple phases. The dollar figure is large and, in a year of multi-billion compute leases, unremarkable. The 1GW is the part almost nobody else prints.

That matters because power is the constraint that actually gates these projects. A dollar commitment can be restructured, delayed, or quietly reduced without anyone noticing. A grid interconnection cannot. When a company states a capacity figure it has handed you something falsifiable: either the substation gets built or it does not. When a company states only a dollar figure, you are reading a press release. Our running tally puts announced capacity across those 26 disclosing deals at 59.2 GW.

## Three kinds of capital, one week

The infrastructure money is the TCS campus. The venture money arrived the same week: Krafton, the Korean gaming company behind PUBG, earmarked $250 million over three to four years for Indian AI, robotics and deep tech startups outside its own gaming portfolio. And the acquisition money showed up as Adobe buying Rilo, an Indian AI workflow startup, in a team-and-technology deal for its agentic marketing products.

Those three move on completely different clocks. A data center campus is a decade. A venture pool is three to four years. An acquihire closes in weeks. They rarely point at the same market in the same week by coincidence.

## The direction is the story

The reflexive read on Indian AI is delivery capacity: cheap engineering hours staffing somebody else's roadmap. Each of these three points the other way.

TCS is building the asset rather than staffing one. Krafton is treating India as a source of deep tech rather than as a user base, which is close to the opposite of how most gaming companies have looked at the market. Adobe bought a team to put inside a product line it sells globally.

Hold that loosely. One week is not a trend, none of these has broken ground or closed, and our Asia deal map still shows the region's capital heavily concentrated in China, Korea and the Gulf rather than India.

## What our own number does not tell you

The 4.1 percent disclosure rate deserves its caveats stated out loud. Seven of the 26 figures are hedged with "up to" or "roughly", and the TCS number is one of them, so the 59.2 GW total is partly an upper bound rather than a commitment. It counts announced capacity, not energised capacity. And it is a floor: it captures deals whose descriptions state a figure, so anything disclosed only in a filing we have not scanned is missing.

The rate itself has barely moved. Eleven new deals entered the tracker since our last count and exactly one of them, the Hyderabad campus, named a capacity. Ten did not.

## What to do with this

**Read announcements for megawatts, not dollars.** If an AI infrastructure announcement gives you a number of billions and no capacity figure, treat its timeline as unconfirmed. The companies that know their power position tend to say so, because it is the part that is hard to fake.

**Watch for the Hyderabad interconnection.** The falsifiable event here is not the 700 billion rupees, it is a published power agreement for the 264 acres. If that lands in the next two quarters, the 1GW is real capacity. If it does not surface by mid-2027, this was a land announcement with a GPU headline on it.