Electricity transmission towers across a rural landscape

Full Article

InfrastructureApril 3, 2026

Tech companies are building a shadow grid - and 30% of data centre power may soon be off-grid

Chevron is building a dedicated gas plant for a Microsoft data centre in Texas. Amazon secured 1.5 GW of dedicated solar. Roughly 30% of all planned data centre capacity is now expected to be on-site. The regulated grid is being bypassed at scale.

The regulated electricity grid was designed around a simple topology: centralised generation, transmission across long distances, and distribution to a dispersed population of end users. What it was not designed for is a class of industrial users large enough to require the equivalent of a small city's power supply in a single location, growing fast enough to outpace any utility planning cycle. The response from those users has been to stop waiting for the grid and start building their own. This is not a workaround or a temporary measure - it is an architectural shift in how the most capital-intensive technology infrastructure in history is being powered, and it is happening across every major hyperscaler simultaneously.

Chevron is working on a deal to build a dedicated natural gas plant for a Microsoft data centre in Texas. Amazon secured 1.5 gigawatts of dedicated solar capacity in the same state. According to a February 2026 report by Cleanview, a market intelligence firm, roughly 30% of all planned data centre power capacity is now expected to be on-site - up from almost nothing a year earlier. Forty-six data centre projects with a combined planned capacity of 56 GW are pursuing dedicated generation infrastructure outright. Microsoft's agreement with Constellation Energy to restart the Three Mile Island Unit 1 nuclear reactor is the most prominent example of a broader pattern: when grid power is too expensive, too slow to contract, or too unreliable in capacity-constrained regions, hyperscalers are moving directly to the supply side.

Nuclear power purchase agreements have emerged as the premium end of the dedicated generation market. The economics are straightforward: nuclear plants provide 24/7 baseload power with no fuel-cost exposure and near-zero carbon emissions, making them ideal for data centres with sustainability commitments and reliability requirements. Google's agreement with Kairos Power for six to seven small modular reactors, Amazon's investment in X-energy, and several other hyperscaler nuclear PPA announcements in 2025-2026 collectively represent a revival of commercial nuclear power demand that the existing fleet of large reactors - most of them built before 1990 - cannot fully satisfy. The result is a private-sector nuclear construction market developing in parallel to, and largely independent of, the public policy debates about nuclear subsidy and regulation.

This divergence - between AI infrastructure that is increasingly self-powered and everything else that depends on the regulated grid - has material consequences for both electricity markets and for the ratepayers who remain on it. Dedicated generation removes high-volume, technically predictable load from the grid's demand base, which would normally reduce capacity market costs. The complication is that it does not reduce the fixed infrastructure costs of the grid itself - transmission lines, substations, distribution networks - which were built to serve a certain total load and must still be maintained regardless of how much of that load migrates off-grid. Those fixed costs are then allocated over a smaller remaining customer base, implying rising per-unit costs for residential customers and small businesses who have no alternative.

The energy island model also creates a new category of infrastructure investment. Developers who can originate, finance, and build dedicated generation assets at data centre scale - whether gas, nuclear, or large-scale solar with storage - are operating in a market that did not meaningfully exist three years ago. The project economics are structurally attractive: long-dated offtake at contracted prices from creditworthy counterparties, with demand visibility that is orders of magnitude better than merchant generation. The critical bottleneck is not capital - there is significant institutional interest in infrastructure assets with technology-company counterparties - but execution: the engineering, permitting, and interconnection work required to deliver firm power to a specific site at a specific date is resource-constrained in ways that capital alone cannot solve.

The permitting and regulatory dimension is underappreciated. Dedicated generation that operates behind the meter - physically connected to a data centre without flowing through the public grid - faces a different regulatory framework than grid-connected generation in most U.S. jurisdictions. State utility commissions, FERC, and in some cases the Nuclear Regulatory Commission all have jurisdictional interests depending on the technology and configuration. Some states are moving to streamline permitting for data centre power agreements; others are resisting, viewing the shadow grid as a threat to their integrated resource planning authority. The regulatory patchwork is one reason that Texas - with its deregulated ERCOT market and lighter state oversight of merchant generation - has attracted a disproportionate share of early dedicated generation deals.

For energy modellers and power market practitioners, the shadow grid is already a significant modelling variable. The traditional assumption that data centre demand flows through the regulated grid is becoming incorrect at scale. Understanding the fraction of AI load that is off-grid, and how that changes marginal pricing, capacity market clearing, and transmission utilisation, is now a first-order input into any serious power market analysis. A model that assumes full grid dependency for AI load will overestimate forward demand on regulated networks and underestimate the rate at which fixed cost socialisation pressures accumulate for remaining grid customers - both errors with significant consequence for utility valuation, capacity investment decisions, and regulatory rate cases.

Model View

Grid demand D_grid(t) = total AI load L_AI(t) x (1 - shadow grid fraction f(t)). As f(t) approaches 30%, the capacity market clearing and transmission utilisation models that assume full grid dependency produce systematically biased forecasts.

Bottom Line

The one thing to remember — the strategic implication in its most compressed form.

The shadow grid is not a future scenario - it is already changing the economics of the regulated grid for everyone who remains connected to it.

Related briefs

These are the adjacent reads for this market arc. Each one adds a different layer: mechanism, allocation, infrastructure, or commodity backdrop.