That AI Datacenter Site Isn't the Capacity You Think It Is

We ran three real datacenter sites in Maharashtra, India through our verification platform. Here is what we found.

• Capacity gaps of 17% to 33% when you model actual workload zones instead of assuming uniform rack density
• Behind-the-meter coverage ranging from 51% to 0.8% across three sites in the same state
• Hidden buildability premiums of +10% to +40% from factors that never reach the investment committee as a single number
• Contested grid infrastructure at every site, with queued demand far exceeding current spare capacity
• Hardware transition decisions required within months of project start, on campuses that will not reach full power for years

This post explains why these gaps exist, how they change at different scales, and what questions you should be asking your development team before your next site goes to IC.

AISIGINT is the independent verification layer for AI datacenter investments. We provide a managed audit and governance service that verifies the ground truth of every site. Give us a site coordinate and a target capacity, and we return an evidence-linked decision signal covering power, grid, permits, buildability, chip roadmap alignment, and more. Typically within days. No software to install. We integrate directly into your deal flow.

I built it because the AI infrastructure buildout is the largest construction wave of our generation, and it needs better intelligence. After two decades in the AI and computing industry, from chipset design at Qualcomm to leading global AI infrastructure research at Omdia, I had a clear view of how fast compute demand was growing and how much infrastructure needed to be built to keep up.

When we pointed the platform at real sites in India, the findings surprised us. Not because good operators don't eventually uncover these issues, but because they typically surface months into the process, across different teams, after consultants have been engaged and money committed. The platform surfaced them from a site coordinate and a target capacity, before any of that. We ran a 150 MW site on 11 acres, a 280 MW site on 20 acres, and an 1800 MW campus on 153 acres. Three very different scales. The same fundamental problems showed up at every one.

The capacity gap

A datacenter in 2026 is not one homogeneous building. It is a campus with fundamentally different zones. Enterprise colo at 5 kW per rack goes in multi-storey air-cooled halls, four floors high. Hyperscale cloud at 15 kW per rack is similar. Three-storey, air-cooled, reasonably compact. But GPU-dense AI halls at 60 to 250 kW per rack, depending on workload and design, are a different building entirely. Most operators build these as single-storey facilities due to the weight of liquid cooling infrastructure, floor loading requirements of 20 to 25 kN per square metre, and the complexity of routing coolant distribution vertically. Multi-storey is technically possible, but expensive and rarely the default choice.

The moment you add GPU-dense workloads to your campus mix, your shared infrastructure requirements increase significantly -- centralized liquid cooling, water treatment, larger substations -- which reduces the proportion of site area available for IT buildings. Experienced operators understand this, and their engineering teams will model it during detailed design. But at the site selection and pre-IC stage, when you are deciding which sites to pursue and what price to negotiate, the analysis often still relies on simpler density assumptions because the full engineering team has not yet been engaged. That is the gap. The information arrives too late to influence the decisions it should be shaping.

At the 150 to 300 MW range, which covers the majority of Indian datacenter developments today, the capacity gap is roughly a third. At the 1800 MW campus scale the gap narrows to 17%, but the site still exceeds available land by over 30 acres.

Assumptions behind these numbers. Each site was modelled as a zoned campus with four workload tiers. Enterprise colo (5 kW/rack, 4-storey air-cooled), hyperscale cloud (15 kW/rack, 3-storey air-cooled), AI inference (60-250 kW/rack, single-storey liquid-cooled), and AI training (up to 600 kW/rack, single-storey liquid-cooled). Land take includes shared infrastructure (substations, cooling plant, logistics, roads) at approximately 40% of gross site area. Rack densities and building form factors are based on current industry reference designs and NVIDIA hardware specifications at projected power-on dates. If your workload mix or density assumptions differ, the gap will shift, but the direction holds for any campus that includes meaningful GPU-dense capacity.

A note on the workload assumption. This analysis assumes a mixed workload campus. Enterprise colo, hyperscale cloud, and AI inference on the same site. If you are building a pure enterprise colo facility with no GPU-dense workloads at any point in the asset's life, the capacity gap largely disappears (because pure colo campuses need far less shared infrastructure -- no centralized liquid cooling plant, no water treatment, a smaller electrical switchyard -- leaving more site area for stackable IT buildings).

But almost nobody is building that site anymore. CtrlS has committed $2 billion for AI-ready campus expansion. Yotta has pivoted from traditional colocation to an AI-cloud platform. NTT is deploying direct-to-chip liquid cooling across new Indian facilities. Industry estimates suggest 50% of datacenter capacity could be AI-related within a decade. Sites evaluated today will operate for 15 to 20 years. Within that horizon, mixed workloads are not speculative. They are the baseline planning assumption.

The capacity gap exists because GPU-dense AI halls at 60 to 250 kW per rack, the workload profile increasingly relevant to India, require substantially more shared infrastructure -- centralized liquid cooling plant, water treatment systems, larger electrical switchyards -- than air-cooled colo. That shared infrastructure footprint, not the IT buildings themselves, is what drives the land shortfall. You do not need frontier training racks at 600 kW to hit this constraint.

The risk changes with scale

At 150 to 280 MW, the dominant issue is physical. Can the campus fit? At both smaller sites, the answer is no. The workload-zoned campus requires 144 to 149% of available land. These sites need to either reduce their marketed capacity, change their workload mix to favour stackable colo over GPU-dense halls, or find adjacent land. These are solvable problems, but only if you know about them before going to your investment committee.

At 1800 MW, the land gap still exists (121% utilisation, 30+ acres short) but the dominant risk shifts to something far harder to resolve. The time dimension.

The 1800 MW campus has a 5-phase energization schedule spanning 72 months. Phase 1 delivers 150 MW on existing 230kV infrastructure by mid-2027. Phase 5 delivers the final 150 MW in early 2032, six years later.

That six-year buildout crosses multiple hardware generations. Phase 1 deploys on current-generation hardware running comfortably on standard 480V AC. Phase 3, reaching power-on in 2028 to 2030, coincides with next-generation inference hardware expected to operate at 250 kW per rack and above, where liquid cooling becomes standard and the electrical backbone should be designed for future 800VDC. Phase 4, reaching power-on in 2028 to 2031, enters the territory where training-class hardware at 600 kW per rack is likely to require 800V DC power distribution.

The platform flagged specific procurement warnings. If 800VDC is needed for Phase 4, equipment must be ordered by approximately month 6 of the project. For Phase 5, the order window is approximately month 12. Equipment lead times for this class of power distribution are 18 to 24 months. These are procurement decisions that need to be evaluated within the first year of a project that does not reach full power for six years.

The economics are clear. Provisioning the electrical backbone for higher-density future hardware during initial construction adds 5 to 8% to cost. Retrofitting the same facility after it is built and tenanted adds 25 to 40%. For Site C, that decision covers over 1000 MW of capacity across Phases 3, 4, and 5. The difference between provisioning now and retrofitting later is a capital allocation decision worth potentially hundreds of crores, and it needs to be made during Phase 1 design.

Most operators I speak to are aware of this transition in general terms. What is harder to do without automation is map exactly which phases, which zones, and which procurement deadlines are affected for a specific site. Without that specificity, the decision gets deferred. And deferral is itself a decision. It is a decision to pay 25 to 40% more later for something that costs 5 to 8% today.

Same state. Completely different power and grid stories.

All three sites are in Maharashtra, within the broader Navi Mumbai corridor. All three face intense demand pressure from competing datacenter projects. But the power story and grid position at each site are completely different.

Site A (150 MW) has a behind-the-meter pathway covering 51% of target capacity (76 MW), deliverable approximately 8 months faster than grid. That is a genuine fast-track option with direct revenue implications. On the grid side, the corridor has over 20 announced or filed datacenter projects competing for limited transmission capacity. The primary substation is contested, and queue position will determine connection timeline.

Site B (280 MW) has effectively no BTM fallback (2 MW, or 0.8% of target). Entirely grid-dependent. And the full facility draw requires 400kV infrastructure that does not exist yet. Phase 1 can deliver just 38 MW on existing infrastructure. The remaining 242 MW requires a multi-year grid programme. The 132kV backup substation is heavily oversubscribed and not a viable alternative.

Site C (1800 MW) has 93 MW of BTM (5.2% of target), mostly biomass, enough to accelerate partial capacity by roughly 2 months. But at this scale the grid story dominates. 4 to 6 injection points required, dedicated 400kV transmission, and a phased programme spanning 3 to 5 years. The nearest viable substation is already contested by multiple large datacenter and industrial projects. A site at this scale is not connecting without significant grid augmentation regardless of current spare capacity.

Three sites, same corridor. One has a fast-track path covering half its load. One has zero fallback and a voltage gap. One needs a multi-year transmission programme with 4 to 6 injection points. That variation does not show up in a corridor-level assessment. It only surfaces when you map each site against its actual grid options, the competing demand in that corridor, and the available BTM alternatives.

The costs that arrive too late

Good development teams assess each of these factors individually. Flood risk, water availability, CRZ exposure, seismic compliance, cooling requirements. The problem is not that operators miss them. The problem is that they are typically assessed by different workstreams, at different stages, and reported in different sections. What rarely happens is aggregation into a single buildability cost premium that reaches the investment committee as one number.

Site B's +40% is driven by CRZ-III classification. The site falls within a Coastal Regulation Zone due to proximity to tidal waters and notified creek systems, triggering special industrial clearance from MoEFCC/SCZMA. That alone adds an estimated +12 months and +15% to project costs.

Site A's +35% comes from a different stack. Flood mitigation, over-exploited groundwater requiring alternative sourcing, and seismic Zone III compliance. Each seems manageable alone. Together they add over a third to baseline cost.

Site C is the lowest at +10%, a greenfield site with safe groundwater, no CRZ issues, and simpler terrain. But at 1800 MW, 10% is a far larger absolute number than 40% on a 280 MW site.

These factors are typically assessed in separate workstreams and rarely aggregated into a single cost impact. The result is that investment committees can approve budgets that are materially understated.

From hearsay to evidence

The first question we get is, "How do you collect this information? Can we trust it?"

It is the right question. In India, most infrastructure intelligence travels through back channels. A call to someone at MSETCL. A conversation with a local fixer. A consultant who knows someone in the district collector's office. This is how the industry works, and it works well enough for individual data points.

The problem is not that back channels are wrong. The problem is that they are impossible to cross-reference, impossible to audit, and impossible to scale. When you are committing hundreds of crores to a 15-year asset, your investment committee needs more than "someone told us." They need to know where each finding came from, whether it lines up with what other sources say, and what happens when sources contradict each other.

AISIGINT collects intelligence from public sources. Utility filings, MERC dockets, Parivesh environmental clearance records, tender portals, satellite imagery, MIDC allotment data, policy documents, and more. We do not treat everything we find as accurate. Our AI agents parse, cross-reference, and flag contradictions across multiple domains simultaneously. When MSETCL's tender portal says one thing and the MERC docket says another and satellite imagery shows something else entirely, the system surfaces that conflict and assigns confidence accordingly.

And we do not rely on AI alone. AISIGINT operates as a human-plus-AI system. Our team includes infrastructure professionals who cross-check every finding against on-the-ground reality, validate assumptions with direct source documentation, and feed corrections back into the system. Human judgement continuously improving machine output. The system gets smarter with every site it evaluates because human expertise is embedded in the loop, not bolted on afterwards.

Your back channels are still valuable. They give you information we cannot get from public sources. But the question is whether you want to go to your investment committee with back-channel intelligence alone, or with back-channel intelligence validated against an independently assembled, evidence-linked picture of the same site. The two together are far more powerful than either one alone.

Every finding in this blog was generated from a latitude, longitude, and a target MW. No site visits. No consultant briefings. No weeks of manual assembly across different teams.

Six questions to ask your development team this week

If you have sites in your pipeline right now, these are the questions that surface the gaps before they become budget overruns.

1. What mix of workloads are we assuming over the next 10 years, and how does that change our MW-per-acre math?
2. What is our realistic buildable capacity once we model separate zones for GPU-dense halls?
3. Do we have a fast-track power pathway (BTM, captive, open access), and how much of our target capacity does it cover?
4. What are the top 3 regulatory approvals that could add 6 to 12 months, and have we started them?
5. What are the "order-by" dates for long-lead equipment, and are we tracking them against our energization schedule?
6. What is the single biggest unknown that could kill this site, and do we have evidence to resolve it?

If your team can answer all six with evidence-linked documentation, your site is in good shape. If any of them produce silence or "we think so," that is the gap AISIGINT is built to close.

Run your pipeline through AISIGINT.

Send us a site coordinate, a target MW, and your intended workload mix. No software to install. No integration required. We return an IC-ready decision signal with evidence-linked findings, typically within days.

Get in touch →