Demand-side record · Enriched · Verified 22 Oct 2026
TCS–OpenAI 100 MW AI Capacity CommitmentOpenAI as first customer of HyperVault · host campus not named
Record PRJ-014National — customer commitmentTier A — demand signal11 public sources cross-checked
Signed 19 Feb 2026Host site unnamedTenant recordNon-additive
100MW
Signed capacity
1GW
Option ceiling
10%
Of campus design
1GW
First customer of a 1 GW build
19 Feb 2026
Announced
0
Sites named
On 19 February 2026, at the India AI Impact Summit, OpenAI became the first customer of TCS's HyperVault data-centre business, committing to 100 MW of AI-ready capacity with an option to scale to 1 GW under its global Stargate programme. It is the single most important demand fact in Indian AI infrastructure. It is also, deliberately, a record with no location: neither OpenAI nor TCS has named the campus. This page records what is signed, what is not, and why the number must never be counted twice.
1At a glance
Record ID
PRJ-014 — Gridvik India Data Centre Registry
Record
OpenAI customer capacity commitment within TCS HyperVault — a tenant record, not a facility
Parties
OpenAI and Tata Group, delivered through TCS's HyperVault AI data-centre business
Committed capacity
100 MW of AI-ready capacity in the initial phase
Option
To 1 GW over time, described by both parties as a stated path rather than a fixed schedule
Announced
19 February 2026, at the India AI Impact Summit in New Delhi
Programme
OpenAI's Stargate initiative — the build-out of AI infrastructure outside the United States
Host location
Not named. Neither OpenAI nor TCS identifies which HyperVault campus or which phase carries the 100 MW
Related records
PRJ-013 — the HyperVault programme. PRJ-063 — the 264-acre Hyderabad campus. Both contain this capacity. Neither is additive to it
Record type
tenant_deployment — carries customer capacity, not facility capacity
Gridvik tier
A — the demand signal that underwrites the largest AI build in India. But it is a commitment, not built capacity
2One number, two framings — the reconciliation you need
Both parties described the same 100 MW, and the wording differs in a way that matters commercially.
How each side framed the same 100 MW
Party
Wording
What it implies
OpenAI
"OpenAI will become the first customer of TCS's HyperVault data center business, beginning with 100 megawatts of capacity and with potential to scale to 1 gigawatt over time"
A customer buying capacity that someone else is building
TCS
"In the initial phase, TCS will develop AI infrastructure with 100MW capacity, with an option to scale to 1GW"
A builder committing to construct 100 MW of capacity
Gridvik read — one commitment, two descriptions, count it once
These are not two deals. OpenAI saying 'we will buy 100 MW' and TCS saying 'we will build 100 MW' describe the two halves of one arrangement. A registry that counted both would report 200 MW of Indian AI capacity that does not exist.
The registry therefore carries this as a single tenant record of 100 MW, flagged non-additive against PRJ-013 (the programme) and PRJ-063 (the Hyderabad campus). If a competitor quotes 1 GW of campus plus 1 GW of programme plus 100 MW of OpenAI, they have triple-counted one build.
The commercial point of the second framing: TCS committed to develop it, not merely to host it. That is a stronger statement than a lease and it is why the equity was raised before the anchor tenant was named rather than after.
3Where 100 MW sits — the scale check
Figure 1OpenAI's 100 MW against the HyperVault Hyderabad campus design of up to 1 GW, and against a single 150 MW build tranche at TCS's own stated capex rate. The signed commitment is 10% of campus design.
As a share of campus design
100 MW is 10% of the 1 GW Hyderabad campus
As a share of one tranche
100 MW is roughly two thirds of a single 150 MW tranche at TCS's stated unit cost of about US$1 bn per 150 MW
What that justifies
Breaking ground and financing phase one. It does not underwrite a gigawatt, and neither party has claimed it does
Capital implication
At about US$6.7 m of capex per MW, 100 MW implies of the order of US$670 m of build cost — about a third of the ₹18,000 crore of committed equity
The option, valued
If the full 1 GW option were exercised, it would commit the entire Hyderabad campus and most of the programme. That is the strategic prize, not the current contract
4What is not named — and why that is the interesting part
This is the rare data-centre record whose most important fact is an absence. Every published account of the February announcement — OpenAI's own release, TCS's statement, Reuters, TechCrunch, the Tata newsroom, and the trade press — declines to say where the capacity goes.
Host campus
Unnamed. Neither party identifies the site
Host phase
Unnamed. Whether the 100 MW is phase one, an early tranche, or drawn across multiple buildings is not stated
Term
Not disclosed
Pricing
Not disclosed
Exclusivity
Not disclosed. Whether OpenAI is the sole tenant of the phase it occupies is unknown
Is the 1 GW option at fixed terms?
Not disclosed
Which entity signs the lease
Not disclosed — OpenAI, a group entity, or a Tata entity contracting on OpenAI's behalf
Why this matters
The Hyderabad campus was announced in September 2026, seven months after the OpenAI deal. That sequence — tenant first, site second — is unusual and suggests the campus was chosen to suit an already-negotiated requirement, which in turn implies the specification is being driven by the tenant's AI architecture rather than by a developer's standard design
The inference Gridvik will make, and the one it will not
Will make: the 100 MW almost certainly lands on the Hyderabad campus at Bharat Future City. It is the only HyperVault campus with secured land, and it was described from the start as the anchor for frontier AI companies. Gridvik treats Hyderabad as the probable host.
Will not make: Gridvik will not label Hyderabad an 'OpenAI data centre', and it will not split the campus capacity into 'OpenAI's part' and 'the rest'. Both would be invention. The registry's own review note is explicit on this: keep the host location and phase unconfirmed.
5The enterprise side of the same deal
The infrastructure commitment was announced alongside a commercial agreement that is often glossed over. It matters because it explains why OpenAI wanted Indian capacity at all.
ChatGPT Enterprise
Rolling out across the Tata workforce over the coming years, beginning with hundreds of thousands of TCS employees — described as one of the largest enterprise AI deployments globally
Codex
TCS intends to standardise AI-native software development on OpenAI's tools across its engineering teams
The strategic logic
Tata gets compute plus tooling; OpenAI gets a captive enterprise deployment, a sovereign foothold in India, and a builder with a listed-company balance sheet
Sovereignty rationale
OpenAI said the infrastructure lets its models run within India, cutting response times and supporting data residency and compliance needs, including government workloads
Programme framing
Positioned under 'OpenAI for India' and the global Stargate initiative
Not capacity
None of the software agreement is data-centre capacity. If a market map shows Tata as a large OpenAI compute site plus a large software deal, only the first part belongs in a capacity register
6Timeline
TCS announces TPG's investment in HyperVault — up to ₹8,820 crore for 27.5–49%, with combined commitments up to ₹18,000 crore. No anchor tenant named.
OpenAI becomes HyperVault's first customer: 100 MW initial, option to 1 GW, under Stargate. Announced at the India AI Impact Summit alongside the ChatGPT Enterprise and Codex agreement.
Restated in TCS's Q4 FY26 results — 100 MW initial phase, option to 1 GW.
HyperVault announces 264 acres at Bharat Future City, Hyderabad, for a campus of up to 1 GW. The host site is named for the first time — seven months after the tenant commitment.
Land also acquired in Visakhapatnam and Pune, widening the possible host set beyond Hyderabad.
Any statement identifying where the 100 MW physically lands, when it is delivered, or what it costs
7Why this record matters to vendors
Read this before the table
A tenant record has no procurement of its own — but it is the reason the procurement exists. This is the demand fact that converts a 1 GW design into a funded phase. Everything below is a Gridvik estimate.
Procurement window and scope
Gridvik assessment · 22 Oct 2026
Order window
The 100 MW is the first tranche, so 2026–2028. Phase one is expected to complete around 24 months from breaking ground; the HyperVault campus itself was not yet under construction as of October 2026. The build that serves OpenAI is the first ordering event on this programme.
Procurement heat
High. A named frontier-AI tenant with a stated option to 1 GW changes the specification risk profile: the buyer is no longer guessing what the workload will be.
Who buys
HyperVault procures; OpenAI specifies. This distinction matters. In anchored AI builds the tenant's technical requirements — rack density, cooling, power quality — flow into the specification, and vendors who have no relationship with the tenant can be excluded before a tender is even drafted
DG / gensets (indicative)
At 100 MW of AI load with N+1 redundancy, roughly 80–95 gensets of 2,000–2,250 kVA class are implied, though frontier-AI campuses often specify larger units and parallel more of them for ride-through on training loads
UPS (indicative)
AI training loads are unusually sensitive to power-quality events. Expect higher-specification UPS and more battery autonomy than an equivalent enterprise colocation build
Cooling (indicative)
Liquid cooling is the entire basis of this deal. HyperVault is designed for rack densities above 170 kW, which is only achievable with direct-to-chip liquid cooling. A 100 MW AI hall implies CDU plant, cold plates, high-flow manifolds and leak detection at a scale no operating Indian facility currently runs
Categories in scope
HV substations · DG and gensets · distributed power and rack-level battery · CDUs · cold plates and manifolds · high-density racks · rear-door heat exchangers · chilled water plant and dry coolers · busway · high-speed optics and structured cabling · fire detection and suppression · leak detection · BMS/DCIM · water treatment · renewables
Decision role
Public decision roles — in GridFeed
Stage-change history
Dated change log — in GridFeed
All estimates are labelled as estimates. No tender dates claimed, no private contact data published.
Capacity commitment, option, announcement date and venue, the enterprise agreement, both parties' wording, the campus announcement date, and the rack density design target
What is derived
The capital implication of 100 MW, the procurement section, and the Hyderabad host inference which is labelled as an inference
Additivity rule
Non-additive. The 100 MW is customer capacity inside the HyperVault programme (PRJ-013) and, in all likelihood, inside the Hyderabad campus (PRJ-063). Summing this record with either would double-count the same megawatts
Capacity basis
Not disclosed as IT load versus facility power. Gridvik records 100 MW with the basis unstated, because neither party has specified it
Host assumption
Stated as probable Hyderabad and explicitly not asserted. The registry keeps location as undetermined until a party names it
Currency
Capex implications use TCS's stated US$1 bn per 150 MW, giving about US$6.7 m per MW. INR conversions use the fixed rate of ₹88.40 per USD
9Limits and open questions
Host site
Unnamed by both parties. The single largest gap on the page
Delivery date
Not stated. No ready-for-service date has been given for the 100 MW
Term and price
Not disclosed
Option status
The 1 GW option is a path, not a schedule. No tranche dates exist
Capacity basis
IT versus facility power is not specified
Tenant exclusivity
Unknown
Delivery risk
The campus was at land-and-permitting stage when announced in September 2026. A tenant commitment of this size against unbuilt capacity puts the schedule risk on the operator, not the customer
What we will not claim
No site, no date, no price, no equipment brand, and no assertion that the option will be exercised
10Sources
This record was compiled entirely from publicly available sources, cross-checked against one another.
OpenAI, 'Introducing OpenAI for India'·Tata Sons newsroom·TCS press release and Q4 FY26 results·Reuters·TechCrunch·Times of India·DatacenterDynamics·Outlook Business·Indian Infrastructure·National Herald India·datacenters.com provider record
Gridvik does not publish source URLs on record pages; citations are retained in the internal source register and are available on request.
What this page is, and what it is not
This is a public-source record of a signed customer commitment inside a larger programme. The commitment is fact; the location is not.
No tender insider information, no private contact data, no operator-paid placement. Corrections are free and published with a date.
