The announcement that crossed my desk yesterday stopped me mid-scroll: National Grid Ventures is investing $1.75 billion for a 35% stake in Joulent, a US energy startup building dedicated power infrastructure for AI data centers. The lead project — codenamed Kilby — is a 2.67-gigawatt natural gas facility in Reeves County, Texas, co-developed with Chevron's Energy Forge division and GE Vernova, under a 20-year power purchase agreement with Microsoft. First power delivery is targeted for 2028, with a final investment decision expected before the end of 2026.
What's a Grid Company Doing Funding Off-Grid Power?
The headline almost writes itself: National Grid, the British infrastructure operator that literally runs the grid, just committed over a billion dollars to a company whose core value proposition is bypassing the grid. That's not a contradiction — it's a signal.
The "across-the-meter" model that Joulent pioneered places generation capacity directly adjacent to the data center campus. Electricity flows from turbine to server without ever touching the public transmission system. No interconnect queue. No waiting for utility upgrades. No dependence on transmission capacity that may or may not materialize on any timeline you can plan around.
If you've spent any time in infrastructure, the logic is immediately familiar. It's the same force that turned serious operators away from shared hosting toward dedicated servers and colocation in the early 2000s. When the shared resource becomes the critical path — and that resource's capacity is controlled by parties with no particular interest in your delivery timeline — you bring the resource in-house.
That's exactly what's happening here, just at the power layer.
The Interconnect Queue Is Real
Let me give some texture to what "interconnect queue" actually means in practice, because it's easy to read past the phrase.
In Texas's ERCOT market, a developer who wants to connect new generation capacity to the grid files an interconnection request with the system operator. Those requests are reviewed and processed in queue order. As of this year, that queue stretches years — not months, years. If you want to open a data center in West Texas today and power it from the public grid, you may be waiting until 2029 or 2030 before you have a reliable interconnection agreement. That's before construction, permitting, and equipment delivery lead times even begin.
AI data centers can't accept that schedule. The business case for a model training cluster dissolves if the cluster can't come online for four years. The capital sitting idle while waiting for grid access is a carry cost that no CFO will sign off on, and the competitive window for the hardware you bought doesn't stay open that long.
Joulent founder Chris James put it plainly: "American innovation is moving faster than the power infrastructure built to support it."
Joulent's answer is to stop waiting. Build your own generation. Place it next to the campus. Run a wire across the meter.
The Scale of What's Being Built
2.67 gigawatts is a number worth sitting with for a moment. For context:
- A typical large combined-cycle natural gas plant runs between 400 and 800 megawatts.
- Three Mile Island's Unit 1 reactor — the one Microsoft restarted in late 2024 to power its Pennsylvania data centers — ran at about 835 megawatts.
- The Kilby project, at 2.67 GW, is roughly three times that capacity, serving a single customer under a single contract.
This is not a backup generator. This is a dedicated utility, built from scratch, to serve one customer for two decades.
The industry numbers behind that commitment are stark. According to Data Center Knowledge, data centers consumed roughly 1.5% of global electricity in 2025. Demand from AI-specific workloads grew 17% that year, against 3% growth in overall global electricity use. By 2030, projected data center demand is 945 terawatt-hours annually — double current levels. AI-specific demand, currently about 20% of the data center total, is expected to reach 40% by 2030.
Those numbers explain why a 20-year PPA at 2.67 GW isn't considered aggressive. It may not be enough.
The Structural Shift: Power-First Site Selection
For most of the history of commercial hosting and data centers, site selection worked a certain way. You chose a location based on network proximity, fiber diversity, tax incentives, labor market, and regulatory environment. Then you worked with the local utility to arrange power delivery. Power was a utility abstraction — something you specified in your buildout requirements and assumed the system would provide.
That model is breaking.
As one analyst quoted this week noted: "power availability is becoming the primary determinant of data center development." Developers are now selecting sites based on where power can be delivered on a credible timeline — and then routing fiber to meet the data center wherever the power dictates. The energy layer is driving the topology of the internet's compute layer. The implications for how hyperscalers select regions, negotiate cloud pricing, and structure availability zones will take years to fully surface.
The Joulent model is also designed for eventual grid interconnection — the across-the-meter approach isn't intended as a permanent bypass of the public system, but as a way to get operational while transmission infrastructure catches up. Once grid interconnection becomes available, the plant can tie in and potentially sell surplus capacity into the market. That optionality is part of what makes the economics work for National Grid Ventures: the investment isn't just a bet on serving Microsoft for twenty years, it's a bet on a future in which large-scale private generation eventually integrates with public grids on terms that favor the generator.
What This Looks Like from the Hosting Side
I've been running hosting infrastructure long enough to remember when colocation was the edgy alternative to leasing rack space in a shared facility. The serious operators moved first. They wanted dedicated power circuits, redundant fiber paths, and infrastructure that wasn't shared with the tenant in rack 47 who was running a poorly-tuned database at midnight and taking the shared UPS with them when it fell over.
The hyperscalers are doing the same thing to the utility system that operators did to shared hosting twenty years ago. They are leaving the shared resource — the grid — for dedicated infrastructure they control. The difference in scale is staggering, but the logic is identical: when your operational requirements exceed what the shared resource can reliably deliver on your timeline, you build your own.
The open question is what this means for everyone who can't build their own. If large hyperscalers progressively exit the grid as grid customers — building dedicated generation that meets their needs — they also reduce the demand base that justifies the transmission investments needed to serve everyone else. The economics of the shared infrastructure shift when the biggest and most creditworthy customers leave it.
I don't think that's the dominant outcome in the near term. Joulent's model anticipates eventual grid integration, and 2.67 GW is a rounding error on national load. But it's a dynamic worth watching: infrastructure built to serve one dominant use case tends to reshape the landscape for everyone downstream of it.
The Operational Bottom Line
For anyone running serious infrastructure at any scale, the Kilby announcement is a useful reminder of a principle I keep coming back to: in a constrained system, whoever controls the bottleneck controls the timeline.
Right now, the bottleneck for AI compute expansion is power delivery. The companies that recognized this earliest — and moved to vertically integrate around it — are the ones building 2.67 GW campuses in West Texas with 20-year contracts in hand. The companies that assumed power was a utility abstraction they didn't need to think about are the ones waiting in an interconnect queue that stretches into the next decade.
The lesson isn't that every infrastructure operator needs to build a gas plant. The lesson is that identifying which layer of your stack is becoming the critical path — and building strategy around it before the queue gets long — is fundamental to staying operational and competitive. Whether it's power, fiber, silicon, or IP address space, the pattern repeats: the resource that seems abundant today becomes the constraint that determines who can scale tomorrow.
National Grid just put $1.75 billion behind the bet that power is that resource for the next decade of AI infrastructure. With Big Tech collectively planning to spend $700 billion on AI infrastructure in 2026 alone, the power constraint isn't going away. On the evidence, I think National Grid's bet is right — and I think the companies that don't have a power strategy in the next two years will find out the hard way why they needed one.