When Capital Outpaces the Land: The Infrastructure AI Actually Runs On
Pyse Capital · 1 July 2026
The global conversation about artificial intelligence is, at this time, a conversation about chips and software. The frame through which most sophisticated observers track the sector sits around Nvidia's market capitalisation, OpenAI's latest parameter count, the race between foundation model architectures. A legitimate frame, by all measures. And yet, the physical infrastructure beneath it receives almost no analytical attention, despite being the constraint that determines whether any of it scales.
Beneath every model or inference request, every training run, there is a layer made of electricity, land, cooling systems and grid connections. A layer that is finite in a way that software is not, and the competition to secure it is already underway.
Why the scale of power demand creates a land problem
A single hyperscale data center at full operational load draws at least 100 megawatts of power continuously, according to the International Energy Agency. One building, running permanently, consumes what it takes to keep half a million Indian households lit, cooled and connected. The largest campuses currently under construction are expected to draw 20 times that amount.
Imagine what a load of that magnitude requires to function. Power at this scale cannot be drawn from the general grid the way an office building draws it. It requires a dedicated substation with confirmed evacuation capacity, a transmission corridor that can handle continuous heavy load and a physical connection that has been engineered, approved and built specifically for that site. A hyperscale campus cannot be plugged into a standard industrial estate and expect the grid to absorb it. The infrastructure must be purpose-built around the load, and that process begins with the land.
Global data center electricity consumption stood at roughly 415 terawatt-hours in 2024, per the IEA. By 2030, the IEA's base case puts that figure at approximately 945 TWh, a near-doubling within six years. India's utilities generated approximately 1,840 TWh across all of FY 2025-26, per the Central Electricity Authority. The incremental power demand that data centers alone will add to the global grid between now and 2030 is, in absolute terms, roughly equivalent to half of everything India currently generates in a year. Each terawatt-hour of that demand has to find a specific substation, a specific corridor, a specific parcel. There is no generalized solution. There is only site-by-site origination work, repeated at a scale and pace the industry has never attempted before.
The division of labour that no one is talking about
The infrastructure that supports AI is not built by a single actor.
Hyperscalers, companies like Microsoft, Google and Amazon, are the anchor tenants and in many cases the developers of their own campuses. Utilities build and maintain the transmission infrastructure that delivers power to those campuses. Construction firms handle the build. On the power sourcing side, the picture is more complicated than the public sustainability commitments suggest: despite being the world's largest corporate buyers of clean energy, accounting for 43% of all global clean energy power purchase agreements signed in 2024 per Brookings, hyperscalers still drew 56% of their operational electricity from fossil fuel sources as of 2024, per the Environmental and Energy Study Institute. The renewable energy transition in data centers is a live, contested infrastructure problem which adds another layer of physical complexity to a chain that is already strained.
And beneath all these actors, at the earliest stage of the process, sits the work that none of them is structured to do. Identifying the land. Verifying title. Navigating regulatory approvals. Confirming grid evacuation capacity, the local grid's ability to actually transmit a massive continuous load outward, before a single dollar of construction capital is committed. Origination work — unglamorous, operationally intensive and almost invisible to the capital markets that follow it.
Budgets are not the bottleneck
Microsoft committed $80 billion to data center infrastructure in fiscal 2025 alone, per public disclosures. The Stargate consortium, OpenAI, Oracle and SoftBank, is targeting $500 billion in U.S. data center investment over four years. These are not small programmes managed by cautious committees. They are among the largest infrastructure capital deployments in the history of the technology industry.
And yet they are running into delays that no amount of additional spending can resolve.
The delay, in specific geographies, comes from the unavailability of parcels that meet a narrow set of physical requirements: proximity to a substation with confirmed evacuation headroom, land with clear title and contiguous acreage, access to power under a bankable long-term agreement, and the regulatory approvals that allow construction to begin. That combination is rare. Making it available requires ground-level origination work that the hyperscaler, the utility and the construction firm are each, for different structural reasons, not positioned to lead.
The result is a division of labour in which origination expertise, the capacity to identify viable sites before the demand reaches them, commands a structurally different position in the infrastructure chain than any of the parties that follow. It is not the loudest part of the AI story. It is, increasingly, the one that determines the pace of all the others.
The semiconductor conversation is sophisticated. The model architecture conversation is sophisticated. The conversation about the land, the substation and the approvals that make any of it operational is not happening at the same level of rigour.
The physical layer is a specific problem, not an abstract one
The physical layer beneath artificial intelligence could be misidentified as a global problem in the abstract. Correctly identified, it is a specific problem, in specific corridors, in specific geographies.
The constraint revolves around specific substations, transmission corridors and land parcels, and it compounds faster than supporting infrastructure is being built. Understanding where it concentrates, and why, is the analytical work that most of the capital tracking this sector has not yet done.
The geographies that matter to this conversation
India's data center capacity stood at approximately 1.4 GW as of mid-2025, per industry data, and is expected to double within two years. That growth rate is not being driven by domestic technology demand alone. It is being driven by global hyperscalers looking for geographies that combine grid access, land availability, a skilled technical workforce and a regulatory environment that has begun to orient itself toward receiving large-scale infrastructure investment.
The GCC is the other geography worth holding in mind. Sovereign wealth funds from Mubadala to Saudi Arabia's PIF are now among the largest data center investors in the world. Middle East capacity is projected to triple from 1 GW to 3.3 GW by 2030, per PwC analysis. The Middle East data center market, valued at $3.48 billion in 2024, is projected to reach $9.49 billion by 2030, per Research and Markets.
What is not visible in those growth projections is how much of that capacity is waiting on a viable parcel with confirmed grid access. The investment pipeline is considerably longer than the origination pipeline behind it.
What Mumbai is already telling us
Mumbai and Chennai together account for roughly 73% of India's national data center capacity, per CBRE. Both cities built that position on the back of specific grid infrastructure, specific submarine cable landing stations and specific regulatory relationships that took years to establish. And both are now running into the limits of what that infrastructure can absorb.
Mumbai is the clearest case. Grid congestion on the Mumbai-Kalwa 400 kV corridor is causing new feeder approvals to be delayed by up to 18 months, per Mordor Intelligence's India hyperscale market analysis. Operators who arrived in Mumbai expecting to move quickly have found that the primary corridor cannot receive them. Some have already pivoted to Navi Mumbai and further into Gujarat, not because of cost or connectivity preferences, but because the land they needed, in the corridor they needed, was simply not available on the timeline their capital required.
This is what an origination constraint looks like in practice. It does not announce itself in advance. It becomes visible only after the capital has committed, the site selection process has run, and the approvals queue has revealed how long it actually takes. By that point, the operators who identified viable positions in adjacent corridors earlier are already moving. The ones who did not are waiting.
The pressure that built in Mumbai is now redirecting capital toward states where the physical prerequisites are still available and where policy has been constructed to receive it. That redirection is not a story about Mumbai failing. It is a story about scarcity propagating outward from established corridors into newer ones, exactly as it has in every mature infrastructure market before this one.
Where the next positions are being taken
Karnataka has entered the picture with purpose. The state published a Data Centre Policy for 2022-27, offering incentives on land, power tariffs and approvals. Falling renewable tariffs under ISTS-exempt power-purchase agreements, arrangements that allow data centers to procure renewable power without paying interstate transmission charges, are making Karnataka economically competitive with markets that were established far earlier. Operators who cannot move quickly on viable parcels in congested corridors are losing ground to those who identified their positions earlier and elsewhere.
The capital flowing into India's data center geography is confirmed at scale and moving fast. The Google-Adani AI campus is targeting 1,000 MW backed by a $15 billion commitment. Reliance Industries announced a 1,000 MW facility in December 2025. OpenAI has announced plans to build a 1 GW data center in India, per industry reporting. Blackstone-backed Lumina CloudInfra is developing a reported 500 MW campus in Navi Mumbai, per Houlihan Lokey's December 2025 India real estate data center report.
These are live capital programs, with active land and power procurement pipelines behind them. The origination work that made their sites viable was completed before the announcements, not after. The sites that were not ready in time are not in the pipeline at all. That is a precise description of how infrastructure markets work: the positions that matter are taken quietly, well before the capital that validates them arrives.
The scarcity that compound demand cannot resolve
A grid-connected parcel of the right size, in a corridor with confirmed evacuation headroom, in a state with active data center policy and bankable power access, is a finite asset: physically and regulatorily finite. More spending does not create more substations. More ambition does not shorten the approvals process. The origination pipeline, parcels identified, titled, grid-connected and approved ahead of formal demand, is, globally, far shorter than the investment pipeline sitting behind it.
What the AI infrastructure moment has produced is a concentration of capital chasing a category of asset that was never built speculatively, because the demand to justify building it speculatively did not previously exist. The market is now in the process of discovering what that means in practice, corridor by corridor, state by state.
The semiconductor conversation is sophisticated. The model architecture conversation is sophisticated. The conversation about the land, the substation and the approvals that make any of it operational is not happening at the same level of rigour, or with the same analytical precision, as the layers above it.
That asymmetry is, in itself, a kind of signal.
Pyse develops, originates and structures income-generating real assets across India and the UAE, at the intersection of energy, electrification and data infrastructure.