Microsoft Project Kilby: When Bypassing the US Grid Costs Less Than Using It

Microsoft and Chevron have signed what is now the largest disclosed behind-the-meter AI power deal in US history. Project Kilby: 2.67 GW of dedicated gas generation near Pecos, Texas. 20-year Microsoft PPA. $7 billion capex. FID later 2026. First power 2028. Engine No. 1 holds a 50% option with matching capex rights. Seven GE Vernova turbines committed despite a multi-year manufacturer backlog.

Zero grid draw. Zero utility involvement. Zero ERCOT exposure.

This is not a power purchasing agreement. It is a declaration that the US grid is not a viable path for large-scale AI data center power at the scale and timeline that hyperscalers require.

WHAT THIS ACTUALLY MEANS

Project Kilby is not an energy strategy. It is a grid arbitrage decision. Microsoft has concluded that the cost of building its own 2.67 GW gas plant — $7 billion in generation capex alone, before a single server is installed — is lower than the cost of waiting in the US grid interconnection queue and navigating PJM or ERCOT tariff uncertainty.

That conclusion has profound implications. It means the US grid queue is not merely slow — it is so slow and so uncertain that the world's most capital-efficient technology company prefers to vertically integrate into power generation rather than use the existing infrastructure. Microsoft is becoming a utility. Not because it wants to. Because the alternative is worse.

The Permian feedstock logic reinforces this: associated gas is frequently flared in the Permian Basin due to takeaway infrastructure constraints. That flared gas is essentially free — a waste product of oil production that creates structurally cheap input costs for anyone willing to build the capture and generation infrastructure. Microsoft is not just bypassing the grid. It is accessing energy that the market cannot currently monetise.

THE COST OF BYPASSING VS THE COST OF WAITING

THE COREWEAVE PARALLEL — CONTRACTED VS LIVE

CoreWeave's situation illustrates the same dynamic at a different scale. The company has 3.5 GW of contracted power capacity and 1 GW active. The 2.5 GW gap is not a financial failure — it is a grid execution problem. Between contracted and live sit interconnection studies, transmission upgrades, tariffs, local politics, and construction timing.

Project Kilby is Microsoft's solution to never ending up in CoreWeave's position at hyperscale. By owning the generation asset, Microsoft eliminates the gap between contracted and active power. The 2.67 GW will be operational when Microsoft needs it, not when the grid operator's queue allows it.

The problem is that this solution costs $7 billion and produces gas-fired power — not the zero-carbon, stable-priced electricity that Microsoft's public commitments require. Project Kilby is a workaround, not a strategy.

WHY FRANCE IS THE STRUCTURAL ANSWER TO THE SAME PROBLEM

Project Kilby solves the US grid problem with $7 billion of private capital and 2 years of development time. France's RTE solves the same problem with €2–20 million of connection costs and 18 months of deterministic process — for a developer who identifies the right brownfield HTB site.

The comparison is not fair to Project Kilby at 2.67 GW — France's fast-track sites top out at 1 GW each. But for the 50–500 MW deployments that represent the majority of AI data center development activity, the France alternative is structurally superior on every dimension: cost, timeline, carbon intensity, regulatory certainty, and absence of private generation dependency.

THE BROADER SIGNAL — HYPERSCALERS ARE BECOMING UTILITIES

Project Kilby is the clearest proof point yet of a structural shift: hyperscalers are vertically integrating into power generation because the utility model cannot serve their deployment speed and scale requirements. This is not a temporary workaround. It is a new industry structure emerging in real time.

The implications extend beyond Microsoft. XOM, Coterra, and other Permian-heavy operators are the logical next counterparties. Amazon, Google, and Meta will face the same calculus. The question is not whether hyperscalers will build their own generation — it is whether they will do it in the US, where the grid queue forces them to, or in markets like France where RTE's deterministic process makes private generation unnecessary.

A grid hookup decides whether an AI data center is a project or a site plan. Project Kilby proves that when the grid hookup takes too long, hyperscalers will spend $7 billion to avoid needing one. France's brownfield HTB sites offer the same outcome — operational power in 18 months — without the $7 billion and without the gas turbines.

→ Related: FERC June 18 — the power-queue ruling · $55M to save 4 months — the cost of delay · France site selection guide

Sources: Project Kilby deal disclosure June 2026 · Bloomberg · Insurance Journal · BNEF US data center capacity forecast · CoreWeave 2025 Form 10-K · FERC June 18, 2026 orders · GridReadiness field intelligence June 2026.