Every few weeks, another Indian state government stands next to a hyperscaler or a domestic conglomerate and announces a new number in gigawatts. Andhra Pradesh has a gigawatt-scale project, Gujarat has announced a 7.5 GW policy target, while Telangana, Maharashtra and Uttar Pradesh have announced or identified projects that add to their respective data center pipelines. The figures climb fast, the announcements often emphasize headline capacity, and the distinction between planned capacity and usable AI compute can easily get lost: a gigawatt announced today does not equal a gigawatt of commissioned AI compute tomorrow. It does not, and the gap between those two things is where the real story of India’s AI buildout actually sits.

The Gigawatt Has Become a Shorthand, Not a Measurement

Capacity figures often function as headline measures of project scale, even though they do not by themselves show how much computing capacity operators have commissioned or use. A company names a number, media outlets repeat it across coverage, and the headline figure can overshadow the project’s development stage. Google and its partners describe the Visakhapatnam hub as “gigawatt-scale,” a phrase that signals the planned buildout’s scale, while the public announcement does not specify a workload-utilization target for the facility.

Reliance’s Jamnagar campus carries similarly large figures within a seven-year investment horizon. None of this makes the projects illegitimate. It does mean that an announced power figure does not by itself show how much IT load operators have commissioned, energized or actually used for computing on any given day. Engineers who design these facilities distinguish between nameplate power, contracted power and delivered IT load, and the gap between the first and the third can become substantial once a facility starts operating

Announcement, Financing and Commissioning Are Three Separate Milestones

India’s current data center pipeline spans widely different stages, from feasibility studies and memorandums of understanding to active construction and projects nearing commissioning. Adani has outlined a $100 billion plan running through 2035. Reliance has pointed to a $110 billion, seven-year commitment alongside more than 120 megawatts expected to come online later this year. L&T Vyoma has signed a state agreement in Dholera targeting a 2028 start date. Each represents a different level of commitment, but a signed memorandum, a groundbreaking ceremony and an operational commissioning mark very different points in a process that can span years and depend on land acquisition, transmission approvals, water access and equipment lead times.

Power purchase agreements can undergo renegotiation. Access to advanced chips can also become a constraint as demand, supply and allocation decisions change across customers and hardware generations. A number announced in February does not guarantee the same capacity will reach commissioning three years later, and industry analysts covering India’s build-out have already identified execution as an important risk as developers move from large announcements to actual deployment.

Rack Density Is Rewriting the Assumptions Under Every Contract

The deeper complication is that facilities receiving financing now must account for accelerator generations whose power density and cooling requirements are changing rapidly. Rack densities that once appeared aggressive can look increasingly conservative as newer accelerator platforms raise power density and introduce more demanding cooling requirements. A data centre engineered around lower-density, air-cooled racks may need upgrades to power distribution and thermal systems before it can accommodate newer, higher-density liquid-cooled hardware. Such retrofits can add capital costs and operational disruption and, depending on the facility’s original design, may require changes to electrical distribution as well as cooling infrastructure. Operators who lock in a facility’s core specifications early face the risk that newer accelerator generations could make parts of the original design less economically attractive even while the infrastructure remains physically usable.

Training Bursts, Inference Persistence and the Utilization Question

There is a second shift happening alongside hardware turnover, and it concerns how these facilities actually get used once they are live. Training workloads can arrive in concentrated bursts tied to model-development cycles, creating periods in which the number of active jobs and the facility’s power demand can vary substantially. Inference, by contrast, is becoming an increasingly important data center workload as AI systems move from model development into production applications and real-time services.

A facility optimized primarily around burst-oriented training does not automatically provide the same operational profile required for sustained, latency-sensitive inference traffic, and the industry’s research on data center architecture is already pointing out that installed power tells an incomplete story about what a site can deliver reliably across its operating life. Utilization is one of the critical variables in determining whether a facility generates an adequate return on its capital, alongside power costs, financing, pricing, efficiency and the revenue generated by its computing capacity.

What India’s AI Ambitions Actually Need to Measure

None of this argues against India’s current wave of investment, which is substantial and backed by government policy that treats AI infrastructure as a strategic priority. It argues for a different set of public metrics. Commissioned megawatts, contracted power arrangements, actual utilization rates and workload mix would provide a more complete picture than another gigawatt figure attached to a groundbreaking photo. Government agencies tracking the India AI mission, along with the state bodies signing these deals, could do the country a service by publishing delivery timelines against announced capacity rather than letting the two numbers blur into one. Until that distinction becomes standard practice, every new gigawatt headline deserves a follow-up question: how much of that number is steel and power lines today, and how much of it will still be relevant compute once it is finally switched on.