HyperVault, the AI data centre subsidiary of Tata Consultancy Services, said on 5 September 2026 that it has secured 264 acres in Hyderabad and that it and its partners expect to invest up to Rs 70,000 crore — roughly $7.4 billion — in a campus of up to one gigawatt. It is the largest single number any Indian IT services company has ever attached to physical infrastructure, and the coverage has understandably led with it.
The figure deserves a closer reading than the headline gives it. “Up to Rs 70,000 crore” is a build-out ceiling contingent on customer demand, not a cheque that has been written. The capital actually committed to this business in public is Rs 18,000 crore of equity from TCS and TPG, announced in November 2025. The rest arrives only if tenants do.
That is not a criticism of the announcement. Phased, demand-linked capital is the correct way to build a data centre, and TCS has been unusually explicit about the arithmetic. What follows sets the announced ceiling beside the committed capital, the chief executive’s own cost-per-megawatt figure and the one anchor tenant that has been named, so you can see which parts of this are contracted and which are intent.
We have written before about how data center operations and AI strategy have stopped being separate conversations. This is the clearest example yet: a services company buying land and grid capacity because the AI work it wants will not fit anywhere else.
Table of contents
- What HyperVault Actually Announced on 5 September
- The HyperVault Capital Stack: Rs 18,000 Crore Committed, Rs 70,000 Crore Announced
- HyperVault by the Megawatt: TCS Priced It at $1bn per 150MW
- OpenAI Is HyperVault’s Anchor Tenant, at 100 Megawatts
- The 264 Acres: Why Telangana Won the TCS AI Campus
- Inside the HyperVault Design: 170kW Racks and Direct-to-Chip Cooling
- The Water-Neutral Claim Nobody Has to Verify
- How HyperVault Compares With India’s Other Gigawatt Campuses
- What Analysts Dislike About the HyperVault Bet
- What HyperVault Means for India’s Data Centre Market
- What This Changes for Businesses Outside India
- How to Read the Next HyperVault Announcement
- Frequently Asked Questions About HyperVault
- References
What HyperVault Actually Announced on 5 September
The press release, datelined Hyderabad and Mumbai, is short and unusually specific about physics and unusually vague about money. Reading the two halves separately is the fastest way to understand the story.
The figures that are firm
HyperVault has secured 264 acres. The campus is designed for up to one gigawatt. The design targets rack densities above 170kW, direct-to-chip liquid cooling, large power blocks and resilient network connectivity. The stated customers are frontier AI companies and hyperscalers running training, inference and advanced computing workloads. Those are engineering commitments, and they are checkable later against what gets built.
The figures that are conditional
The Rs 70,000 crore is attributed to “HyperVault and its partners”, is prefixed “up to”, and is tied to development “in a phased manner, in line with customer demand and technology requirements”. Every one of those qualifiers moves risk off the balance sheet and onto the tenant pipeline. The jobs figure — around 7,000 direct and indirect, per local reporting — carries the same conditionality.
What was not announced
No phase-one megawatt figure. No energisation date. No power purchase agreement. No named tenant beyond the one already public from February. No capital structure for the Rs 70,000 crore. The absence is not evasion; those things are usually settled per phase. But it means the headline number is currently the least informative figure in the release.
| Element | Stated figure | Status |
|---|---|---|
| Land | 264 acres, Bharat Future City, Hyderabad | Secured |
| Capacity | Up to 1 GW | Design target, phased |
| Investment | Up to Rs 70,000 crore (~$7.4bn) | Ceiling, HyperVault “and partners” |
| Rack density | Above 170kW per rack | Design target |
| Jobs | ~7,000 direct and indirect | Projection at full build |
| Phase 1 size, date, tenant | None given | Not in the record |
The HyperVault Capital Stack: Rs 18,000 Crore Committed, Rs 70,000 Crore Announced
The gap between what has been committed and what has been announced is the single most useful thing to hold in your head about this project, and it is entirely public.
What TPG actually agreed
On 20 November 2025, TCS said TPG would invest up to Rs 8,820 crore — about $1 billion — for a stake of between 27.5% and 49% in HyperVault, with TCS retaining 51%. Together the two committed up to Rs 18,000 crore in equity. TPG is investing through TPG Rise Climate and its Global South Initiative, alongside its Asia Real Estate business. N. Chandrasekaran, TCS chairman, framed it as building “large GW-scale AI data centers in India”.
How the rest gets funded
Equity of that size does not build a gigawatt. Business Standard reported the partners also aim to raise $4.5–5 billion in debt. That is the layer that turns Rs 18,000 crore of equity into a Rs 70,000 crore programme, and it is the layer that has to be raised phase by phase against signed contracts.
The proportion that is actually committed
Rs 18,000 crore against a Rs 70,000 crore ceiling is 25.7%. Put the other way, just under three quarters of the headline figure is capital that has not yet been raised, from lenders who have not yet been named, against demand that has not yet been contracted beyond one customer.
HyperVault by the Megawatt: TCS Priced It at $1bn per 150MW
The most valuable number in this whole story was not in the September release. It came from K Krithivasan on an analyst call, and it makes the headline figure legible.
The chief executive’s own arithmetic
Krithivasan told analysts TCS had set a target of one gigawatt, would build it in phases, and expected to get there over five to seven years. Then the useful part: “Roughly about every 150 megawatt would be about a billion dollars.” That is $6.67 million per megawatt, and it is a number you can hold the announcement against.
Where the Rs 70,000 crore ceiling comes from
Run it out. Six and two-thirds tranches of 150MW reach a gigawatt at roughly $6.67 billion. The announced ceiling is $7.4 billion, about 11% higher — which is exactly the sort of margin you would expect once land, grid connection and a contingency are folded in. The HyperVault headline is not an inflated round number; it is the CEO’s own unit cost with a sensible buffer on top.
The clock that matters more than the number
Five to seven years for a gigawatt means the last tranche lands somewhere between 2031 and 2033. Anyone reading $7.4 billion as near-term capex is reading it wrong by roughly a factor of six. The annual run-rate implied is closer to $1.1–1.5 billion, spread across phases that each need their own tenant.
| Tranche | Cumulative capacity | Cumulative cost at $1bn/150MW | Share of 1 GW |
|---|---|---|---|
| 1 | 150 MW | $1.0bn | 15% |
| 2 | 300 MW | $2.0bn | 30% |
| 4 | 600 MW | $4.0bn | 60% |
| 6 | 900 MW | $6.0bn | 90% |
| 6.67 | 1,000 MW | $6.67bn | 100% |
| Announced ceiling | 1,000 MW | $7.4bn (+11%) | 100% |
OpenAI Is HyperVault's Anchor Tenant, at 100 Megawatts
A gigawatt campus is a leasing business, and leasing businesses live or die on the first signature. HyperVault already has one, and it is a good one.
What was agreed in February
At the India AI Impact Summit on 19 February 2026, TCS and OpenAI announced a strategic partnership under which OpenAI becomes the first customer of the HyperVault data centre business, starting at 100 megawatts with the potential to scale to a gigawatt over time, as part of OpenAI’s Stargate programme. That is the anchor the September land announcement is built on.
What 100 megawatts represents
One hundred megawatts is 10% of the campus design capacity, and roughly two thirds of a single 150MW tranche at the CEO’s own unit costing. It is enough to justify breaking ground and financing phase one. It is not enough to underwrite a gigawatt, and nobody has claimed it is.
Why the anchor decides the phasing
“In line with customer demand” is not boilerplate. Each 150MW block needs a tenant before the debt behind it can be raised economically. If OpenAI’s option to scale converts, HyperVault’s phases follow quickly. If it does not, the campus stops at whatever the pipeline supports, and the 264 acres simply sit there — which is why the land was secured up front and the capital was not.
The 264 Acres: Why Telangana Won the TCS AI Campus
Location is the part of this story that was genuinely competitive, and the state has been open about that.
The site
The campus sits in Bharat Future City, the greenfield development south of Hyderabad that Telangana has been assembling as its next industrial corridor. At 264 acres for up to a gigawatt, the plan implies about 3.8 megawatts per acre — dense by Indian standards and only achievable with the liquid-cooled, high-rack-density design the release describes.
Who else was in the running
Local reporting says Telangana beat five other states for the investment. Chief Minister A. Revanth Reddy called the project historic “not just for Telangana but for India”, adding that “this decade is one of Artificial Intelligence” and that “access to AI models and compute is fast-becoming public infrastructure”. IT and Industries Minister D. Sridhar Babu made the pitch plainly: “Telangana Means Business.”
What a state actually contributes
Land assembly at this scale, and the grid connection behind it, are state functions. That is the real competition between Indian states for AI campuses — not tax breaks but the ability to deliver contiguous acreage and firm power on a schedule. Telangana delivered the first; the second is the part still to be proved for HyperVault and for every rival campus in the country.
Inside the HyperVault Design: 170kW Racks and Direct-to-Chip Cooling
The engineering claims are the most concrete part of the announcement, and they are also the part that makes the economics work.
What 170kW a rack actually changes
A conventional enterprise hall runs 5–10kW a rack. Modern colocation pushes 15–30kW. Above about 50kW, air stops being viable at any sensible cost. Designing HyperVault for more than 170kW puts it firmly in the GPU-training tier, where a single rack draws more power than a small office building and the entire mechanical design has to change to match.
Why liquid cooling is not optional at that density
Deepesh Kiran Nanda, chief executive and managing director of HyperVault AI Data Center Limited, put the design brief in one sentence: “We are building infrastructure for where AI is going: higher density, liquid cooling, larger power blocks and faster deployment.” Direct-to-chip cooling moves heat in liquid straight off the package rather than pushing it through the room, which is the only way to keep a 170kW rack thermally sane.
The density dividend
High density is why 264 acres can hold a gigawatt. It is also why the capital cost per megawatt can land near the CEO’s $6.67 million — fewer buildings, less air-handling plant, more compute per square foot of shell. The trade is that the mechanical design becomes far less forgiving, and retrofitting it later is close to impossible.
| Tier | Typical rack density | Cooling method | Typical workload |
|---|---|---|---|
| Enterprise server room | 5-10 kW | Room air | Business applications |
| Standard colocation | 15-30 kW | Contained hot aisle | Cloud and hosting |
| High-density colocation | 40-60 kW | Rear-door heat exchangers | Mixed AI inference |
| HyperVault design target | 170 kW and above | Direct-to-chip liquid | Frontier model training |
The Water-Neutral Claim Nobody Has to Verify
The release says the campus will be built on “green energy and water-neutral design principles”. Both phrases are welcome and neither is defined anywhere in the announcement.
Why the wording matters in Hyderabad
Water-neutral can mean a closed-loop design that consumes almost nothing, or it can mean consuming freely and offsetting elsewhere through recharge projects. Those are very different outcomes for a city that already manages summer supply carefully. Nothing in the HyperVault announcement says which one is meant, and no Indian rule compels an answer.
Europe requires the number; India does not
The contrast is instructive. Under EU Delegated Regulation 2024/1364, data centres with at least 500kW of installed IT power must report annually — by 15 May — into a European database: total water consumption, drinking water consumption, water usage effectiveness, energy demand, renewable energy share and waste heat recovery. India has no equivalent reporting obligation. As one write-up of the announcement put it, water-neutral design is a phrase in Hyderabad and a filed number in Frankfurt.
The disclosure that would settle it
A single published water usage effectiveness figure per phase, alongside a power usage effectiveness number and the renewable share of supply, would convert the claim into a commitment. HyperVault is under no obligation to publish any of it. Whether it does anyway is a reasonable proxy for how seriously the sustainability language was meant, and it costs the company nothing to answer.
How HyperVault Compares With India's Other Gigawatt Campuses
TCS is not early to this. It is entering a market where several Indian groups have already announced gigawatt-scale AI campuses, and the announced capital per gigawatt varies by a factor of three.
The per-gigawatt spread
Google and Adani are building a roughly one-gigawatt AI hub at Visakhapatnam on about $15 billion over five years. Reliance has a 1.5GW cluster in Andhra Pradesh reported at $17 billion, and a much larger Jamnagar programme reported at $20–30 billion for 3GW. Lodha has announced $11 billion for a 2.5GW park in Maharashtra. Against those, the HyperVault figure of $7.4 billion for a gigawatt sits at the cheap end.
Why the numbers are not like for like
They cover different things. Google’s figure explicitly bundles subsea cable capacity and green energy infrastructure. Reliance’s Jamnagar programme includes generation. Announced totals also mix committed and aspirational capital in different proportions — which is precisely the ambiguity this article started with. Treat the column as a ranking of announcements, not of costs.
The comparable number
The one figure that travels across all of them is capital per megawatt of IT-adjacent capacity, and on that basis HyperVault at $7.4 million per megawatt is close to the CEO’s own build cost and roughly half Google’s announced figure. That is either commendable discipline or an incomplete scope. Which one it is becomes visible only when phase one is energised.
| Campus | Announced capital | Capacity | Implied $ per GW |
|---|---|---|---|
| Lodha, Maharashtra | $11bn | 2.5 GW | $4.4bn |
| Reliance, Jamnagar | $20-30bn | 3 GW | $6.7-10bn |
| TCS HyperVault, Hyderabad | $7.4bn | 1 GW | $7.4bn |
| Reliance, Visakhapatnam | $17bn | 1.5 GW | $11.3bn |
| Google + Adani, Visakhapatnam | $15bn | ~1 GW | ~$15bn |
What Analysts Dislike About the HyperVault Bet
The market reaction to TCS’s data centre ambitions was not enthusiastic, and the objection is coherent enough to state properly.
The return-ratio argument
TCS’s core business posted a return on equity of 51% and a return on invested capital above 80% in FY25. Those are the numbers of a company that sells people’s time and owns very little. BOBCaps estimated a data centre business would return equity in the teens. Blending a low-teens business into an eighty-per-cent one dilutes the ratios that have historically justified the multiple, and it does so for a decade.
The strategic-fit argument
PhillipCapital’s Karan Uppal made the structural version of the point: TCS is moving away from a capex-light model to a capex-intensive one. Analysts also flagged limited overlap with core IT services. Building and leasing halls is a real estate and power business with an infrastructure cost of capital, not a services business, and the skills only partially transfer. TCS shares fell as much as 1.5% on the day the wider plan landed.
The case for doing it anyway
The counter is straightforward. TCS booked $40.7 billion of total contract value in FY26 and reported annualised AI revenues above $2.3 billion, and the constraint on growing that is access to compute its clients can legally use. Data localisation rules mean some of that work cannot leave India. Owning the capacity converts a supply constraint into an asset, and HyperVault is the vehicle for it. Whether that is worth the ratio dilution is a genuine disagreement, not a settled question.
What HyperVault Means for India's Data Centre Market
Zoom out and the campus stops looking like an outlier and starts looking like a symptom.
The capacity curve
India’s installed data centre capacity was around 1.5GW in 2025 and is forecast to reach roughly 6.5GW by 2030. CBRE expects about 30% growth in 2026 alone on around 500MW of new supply. Against a 2030 figure of 6.5GW, a single one-gigawatt HyperVault campus is over 15% of the entire national fleet — from one subsidiary that did not exist two years ago.
AI has taken over leasing
The mix has shifted violently. AI workloads accounted for 78% of data centre leasing activity in 2025, against 23% the year before. That is the demand signal every one of these announcements is responding to, and it explains why the designs have jumped straight to liquid cooling rather than arriving there gradually.
The binding constraint is power
The sector’s announced pipeline in India runs to roughly $90 billion. Land is available and capital is available; firm, clean, gigawatt-scale power on a fixed date is the thing in short supply. That is why the HyperVault release leads on “large power blocks” and green energy integration, and why the power purchase agreements — none of which have been announced — are the documents that will actually determine the schedule.
What This Changes for Businesses Outside India
A campus in Hyderabad is not obviously a UK or European procurement question. In two ways it is.
Sovereignty claims need a metric
The pitch behind HyperVault is that compute located in India, funded and operated locally, satisfies data residency in a way rented hyperscaler capacity does not. That argument travels. When a supplier tells you their AI runs on sovereign infrastructure, the follow-up is which entity owns the building, which owns the silicon, and where the inference actually executes — three questions that frequently have three different answers.
Cybersecurity and control move with the workload
Moving model training onto a partner’s campus changes your control boundary. The cybersecurity questions that follow are the ordinary ones — tenancy isolation, key custody, incident notification, subprocessor lists — but they get asked of a landlord rather than a cloud provider, and the contractual language is far less standardised. This is the same discipline any serious cloud migration demands, applied to a newer kind of supplier.
Capacity pricing is about to move
If India adds several gigawatts of AI-grade capacity by 2030, the price of high-density compute in the region falls, and the arbitrage between training somewhere cheap and serving somewhere near the user gets wider. That is a planning input for anyone budgeting model work over a three-year horizon, whether or not they ever touch HyperVault.
How to Read the Next HyperVault Announcement
The next twelve months will produce more releases. Three habits make them easy to grade.
Count megawatts energised, not crores announced
Capital figures are ceilings and they get restated. Energised megawatts are physical, dated and hard to spin. A release that says “phase one, 150MW, live” is worth more than one that raises the headline number to Rs 90,000 crore.
Watch who signs after OpenAI
One anchor at 100MW justifies phase one. The second and third tenants are what convert HyperVault from a large project into a gigawatt business. If twelve months pass with no new named customer, the phasing language in the September release is doing exactly the work it was written to do.
Watch the power purchase agreements
Green energy integration is currently an adjective. A signed PPA — with a named supplier, a volume and a term — is the point at which it becomes a plan. The same applies to the water-neutral claim: a published water usage effectiveness figure would settle in one line what the press release leaves open.
Watch the phase-one groundbreaking date
Land secured is not construction started. Between them sit environmental clearance, grid connection agreements and a financing close. The interval between the September land announcement and the first pour is the cleanest available measure of how firm the HyperVault pipeline really is.
Frequently Asked Questions About HyperVault
What is HyperVault?
HyperVault AI Data Center Limited is the AI data centre subsidiary of Tata Consultancy Services, majority owned by TCS with TPG holding a minority stake of between 27.5% and 49%. It builds and operates high-density, liquid-cooled capacity for AI training and inference rather than selling services.
Has TCS committed the full $7.4 billion?
No. The announcement says HyperVault “and its partners” expect to invest “up to” Rs 70,000 crore, phased against customer demand. The publicly committed equity is Rs 18,000 crore from TCS and TPG, about 25.7% of the headline figure, with the balance expected from debt raised phase by phase.
When will the campus be operational?
No date has been published for the Hyderabad campus. TCS has said separately that it expects to reach one gigawatt of capacity over five to seven years, built in phases of roughly 150 megawatts each.
Who will use the HyperVault campus?
OpenAI was named in February 2026 as the first customer of the data centre business, starting at 100 megawatts with an option to scale, under its Stargate programme. The stated target market is frontier AI companies and hyperscalers. No other tenant has been named.
Is this the largest AI data centre in India?
Not by announced capacity. Reliance’s Jamnagar programme is reported at 3GW and its Visakhapatnam cluster at 1.5GW, and Lodha has announced 2.5GW in Maharashtra. At up to 1GW, the HyperVault campus is in the same tier as the Google and Adani hub at Visakhapatnam.
Why does rack density matter so much here?
Above roughly 50kW a rack, air cooling stops being economic. Designing for more than 170kW commits the campus to direct-to-chip liquid cooling, which is what allows a gigawatt to fit on 264 acres and keeps the cost per megawatt near the $6.67 million TCS has quoted.
References
TCS’ HyperVault to establish large-scale AI data center campus in Telangana
TCS Secures $1Bn Investment from TPG to Accelerate AI Data Center Business HyperVault
TPG to pump $1 billion into TCS data centre business HyperVault
TCS to build 1 GW AI data centre in India
TCS subsidiary HyperVault to invest up to Rs 70,000 crore in Hyderabad AI data centre
This decade is one of AI: TCS unit to build Rs 70,000-cr data centre in Hyderabad
Adani and Google Partner to Build India’s Largest Data Centre Campus in Visakhapatnam
India’s installed data centre capacity to rise 4-fold by 2030
Commission Delegated Regulation (EU) 2024/1364 on a common Union rating scheme for data centres
India’s Tata launches business entity for 1GW data center build-out
TCS closes FY26 with Improving Sequential Growth Momentum and Strong Deal Wins
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.