For an investor or developer evaluating an AI data center, the model usually rests on two assumptions: that you can secure enough power, and that you can secure it soon. Both are getting harder to make.
The two inputs AI compute depends on most — grid electricity and water — have become the two inputs most likely to delay your project, cap its returns, or strand your capital. That isn't a sustainability footnote. It's the central risk in the asset class. And it's why we design AI data centers that depend on neither.
The Grid Has Become the Bottleneck — and a Balance-Sheet Risk
The demand picture is no longer speculative. The International Energy Agency reports that data centres accounted for roughly 1.5% of the world's electricity in 2024 — about 415 TWh — and projects that to more than double to around 945 TWh by 2030, naming AI as "the most important driver of this growth" (IEA, Energy and AI, 2025). In the United States, Lawrence Berkeley National Laboratory found data centers consumed 4.4% of U.S. electricity in 2023 and could reach between 6.7% and 12% by 2028, with demand having "more than doubled" between 2017 and 2023 "largely due to the growth in AI servers" (Berkeley Lab, 2024 U.S. Data Center Energy Usage Report).
The investor-relevant part isn't the demand — it's the grid's response to it. The IEA estimates that "around 20% of planned data centre projects could be at risk of delays" from grid strain, and notes that "nearly half of data centre capacity in the United States is in five regional clusters," each drawing power "as much as power-intensive factories such as aluminium smelters."
For a developer, "grid delay" is not abstract. It's an interconnection queue measured in years, a substation upgrade you don't control, and a capital structure where revenue can't begin until a utility says yes. Every month of delay is carrying cost on a nine- or ten-figure asset that isn't yet earning.
Water Is the Second Hidden Liability
The less-discussed dependency is water — and the peer-reviewed evidence is sobering. In Making AI Less "Thirsty," published in Communications of the ACM, researchers Li, Yang, Islam, and Ren found that training the GPT-3 model in Microsoft's U.S. data centers can directly evaporate 700,000 liters of clean freshwater, and projected that global AI demand could account for 4.2–6.6 billion cubic meters of water withdrawal by 2027 — more than the total annual withdrawal of four-to-six Denmarks, or roughly half the United Kingdom (Li et al., Communications of the ACM).
Most hyperscale AI facilities reject heat using evaporative water cooling. That makes water a gating factor for siting — and a growing source of regulatory and community friction in precisely the regions where land and power might otherwise be available. For an investor, water scarcity is a permitting risk, a reputational risk, and a future-moratorium risk stacked on top of the power problem.
So the question we started with is the wrong one. It isn't "can we secure enough grid power and water?" It's "what if the asset needed neither?"
Evaluating an AI data center development? Let's pressure-test the economics of an off-grid, water-free design for your site.
Schedule a Feasibility Consultation →A Different Foundation: an Industrial Property That Makes Its Own Power
The model we design begins from a different premise: an industrial commercial property that produces its own clean fuel on-site.
Hydrogen is generated at the facility from raw feedstock. That generation step yields a marketable industrial byproduct — sold to local industry as a standalone revenue line, turning part of the energy system into a contributor rather than a cost. The hydrogen then feeds on-site power generation, converting fuel to electricity at roughly 40–60% efficiency, and that electricity powers the data center.
The result is an asset with a fundamentally different risk profile:
- Energy sovereignty. The facility isn't waiting in an interconnection queue or exposed to grid congestion, curtailment, or transmission delay. Power availability is a function of the site's own generation — not a utility's capital plan.
- Continuity at data-center standards. AI infrastructure is held to a 99.999% ("five nines") uptime expectation — roughly five minutes of downtime per year. Dispatchable on-site generation is designed to meet that standard without depending on grid reliability.
- A built-in second revenue stream. The byproduct sale means the energy system isn't purely an operating cost; it's complementary income that strengthens project economics over the asset's life.
- Genuinely clean fuel. Power is produced from hydrogen rather than drawn from a grid mix the operator doesn't control — bringing clean energy directly onto the cloud-infrastructure layer.
It's the same principle we apply across our data center work: infrastructure should generate, not just consume. Hydrogen makes that literal — the property produces its own power and a saleable product from the same process.
Cooling Without Water — and Why Two-Phase Changes the Math
The second dependency, water, we remove at the source. Instead of evaporative water cooling, our facilities use two-phase, direct-to-chip cooling — a closed loop that uses no water at all.
Here's the mechanism, and why it matters. A low-pressure dielectric refrigerant circulates into a cold plate mounted directly on each processor. On contact with the chip, the fluid boils — and that phase change from liquid to vapor is what does the cooling. The vapor travels to a heat-rejection unit, condenses back to liquid, and returns to the chip in a sealed cycle. No municipal water enters the loop at any point.
The physics is the advantage. Conventional single-phase liquid cooling (water or glycol) carries heat away by raising the fluid's temperature — so its capacity is capped by flow rate and the fluid's heat capacity. Two-phase cooling instead exploits the latent heat of vaporization: the energy absorbed during a phase change exceeds simple temperature-rise heat transfer by orders of magnitude. The practical consequences are exactly what a high-density AI hall needs:
- It eliminates the water dependency the ACM research identifies as a scarcity and siting liability — removing a permitting constraint and a major source of community and regulatory friction, and widening the map of buildable sites.
- It removes water risk from the rack. The coolant is dielectric and non-conductive — chemically inert to electronics and safe to the IT equipment even in the event of a leak. Running at low pressure with a non-conductive fluid means a connection failure can't create an electrical hazard or corrode the most expensive silicon in the building.
- It keeps pace with rising chip power. As GPU thermal design power climbs generation over generation, two-phase moves far more heat per unit of fluid at lower flow rates — meaning smaller pumps and lighter cooling infrastructure while still cooling the densest accelerator racks.
In other words: the cooling architecture that protects your highest-value assets is also the one that erases your water liability. You don't trade performance for sustainability — you get both from the same decision.
What This Changes for Investors and Developers
Put the two design decisions together and the asset's risk profile is transformed at the foundation, before a single rack is installed:
- No interconnection-queue dependency. Revenue timing is governed by construction and on-site generation — not a utility's multi-year upgrade schedule. That directly addresses the ~20% project-delay risk the IEA flags.
- No water-scarcity siting constraint. The facility sidesteps the freshwater liability documented in the peer-reviewed literature, expanding viable locations and reducing moratorium and reputational exposure.
- A complementary revenue stream from the hydrogen byproduct, improving lifetime economics rather than treating energy purely as cost.
- A clean, sovereign energy story — the kind ESG-aligned capital and host communities increasingly require, turning a contentious development into a supported one.
None of this trades against performance or uptime. It's a different starting architecture — one designed so the two factors most likely to delay, constrain, or politicize an AI data center are simply not in the dependency chain. For an investor, that is what "secure infrastructure" actually means: returns that don't hinge on a utility's timeline or a region's water table.
The next generation of AI data centers won't compete only on rack density and PUE. They'll compete on independence — and the assets built for sovereignty will be the ones still bankable when the grid and the water supply can't keep up.
Frequently Asked Questions
A great deal. Peer-reviewed research in Communications of the ACM found that training GPT-3 in Microsoft's U.S. data centers can directly evaporate about 700,000 liters of clean freshwater, and projects global AI water withdrawal of 4.2–6.6 billion cubic meters by 2027 (Li et al.). Most hyperscale facilities use evaporative water cooling, making water a siting and regulatory constraint.
Demand is rising faster than grids can expand. The IEA projects global data center electricity use will more than double to ~945 TWh by 2030 and estimates ~20% of planned projects risk delay from grid strain (IEA, 2025). Berkeley Lab found U.S. data centers could reach 6.7–12% of national electricity by 2028 (LBNL).
In this model, yes. Hydrogen is produced on-site from raw feedstock and converted to electricity by on-site generation (~40–60% efficiency), powering the facility independently of the grid. Dispatchable on-site generation is designed to meet the data-center standard of 99.999% uptime (~5 minutes per year).
A low-pressure dielectric (non-conductive) refrigerant flows into a cold plate on each chip and boils on contact; that liquid-to-vapor phase change absorbs large amounts of heat. The vapor condenses back to liquid in a sealed loop and returns — no water involved. Because it uses the latent heat of vaporization rather than simply warming a fluid, it moves far more heat at lower flow rates, cools the densest GPU racks, and is safe to the hardware even in a leak.
On-site hydrogen generation yields a marketable industrial byproduct that can be sold to local industry, creating a complementary income line that improves the project's lifetime economics rather than treating energy purely as an operating cost.
It removes the two dependencies most likely to delay or constrain an AI data center — grid interconnection and water — while adding a complementary revenue stream. The result is faster, more controllable revenue timing, broader siting options, stronger ESG and community alignment, and an asset whose returns don't hinge on a utility's schedule or a region's water supply.