A megawatt does not mean much in the abstract. Operators evaluating off-grid options need to know what one megawatt of continuous gas-fired power, on a real campus, actually supports in 2026 workload terms.
The short version: roughly 800 to 1,000 modern AI accelerators at full training load, or 3,000 to 5,000 mid-tier inference GPUs running hot. Or about 1,500 traditional CPU servers in a non-accelerated rack mix. The number depends on which accelerators and which workload profile, but the band is real.
That is what one megawatt buys. The question for an operator is whether one megawatt is the right unit to be thinking about at all.
The continuous part is the part that matters
"Continuous" is the operating word in 1 MW continuous. Not peak. Not name-plate. Not theoretical.
Continuous means the load can pull a megawatt 8,760 hours a year without the source flickering. That is what AI training requires and what intermittent renewable sources alone cannot provide without significant storage layered on top. A solar-plus-battery installation rated at 1 MW peak does not deliver 1 MW continuous unless the battery is sized large enough to bridge the entire daily and seasonal envelope, which usually it is not.
Natural gas generation, fed from an operating field with dedicated allocation, does deliver continuous. Combined-cycle units are designed for base load. The fuel arrives at the generator at the rated pressure, the generator runs at the rated output, and the campus pulls what it pulls.
For an AI training run that takes 60 days, this distinction is everything. A 60-day training run that gets interrupted on day 41 because the solar contribution dropped below threshold for six hours does not produce a model that is 70 percent trained. It produces a checkpoint that has to restart from the last clean save, with whatever throughput penalty that imposes.
What 1 MW supports in current accelerators
Working from the public specs for the accelerators most operators are deploying in early 2026:
A modern training-grade accelerator pulls roughly 700 watts at sustained load. Add cooling and the supporting infrastructure at a PUE around 1.3 and a single accelerator's real draw is closer to 900 to 1,000 watts on the campus side of the meter. One megawatt supports about 1,000 of these accelerators, give or take, before the math starts pulling capacity from cooling.
Inference accelerators draw less, around 300 to 400 watts each at sustained load. One megawatt supports roughly 3,000 to 5,000 of these, depending on the cooling architecture.
A traditional CPU server with no accelerators pulls 400 to 700 watts, and a megawatt of campus power supports about 1,500 of those in a standard rack mix.
These are floor numbers, not optimization targets. A campus designed for a specific workload profile, with the cooling and the network designed around that profile, can extract more useful work from the same megawatt. But the band gives an operator something to compare against the deal they are actually trying to support.
What scaling looks like
The reason the 1 MW number is worth discussing is not that 1 MW is the deal size. It is that 1 MW is what is operating today and the same field has room to scale.
The field we have been pointing operators toward in Eastern Kentucky has 140 active wells and 23 miles of pipeline. The 1 MW continuous is what is flowing now. The math on the field's reserves and well capacity points to 1 gigawatt plus of headroom on the same physical infrastructure.
For an operator, that scaling potential matters more than the current production number. A campus designed for 50 MW does not get built on a 1 MW source. It gets built on a source that can demonstrably get to 50 MW within the campus build timeline. The field's existing pipeline and well count are the proof points.
The phasing usually looks like this:
Phase 1, the first 1 to 5 MW, comes off existing well capacity with light additional surface equipment. This is the proof phase, where the operator demonstrates the campus pulls what it is supposed to pull at the prices the deal was structured around.
Phase 2, scaling to 10 to 30 MW, requires additional gen-sets on the campus and additional pipeline tie-ins. The well capacity already exists. Surface buildout is the gating factor.
Phase 3, beyond 30 MW, gets into the full 1 GW scaling math. New wells, additional pipeline capacity, and substantially expanded surface equipment. This is the longest-lead phase but also the one that opens up genuinely large campuses.
What this is not
This is not a solar-replacement story. Operators with a working renewable strategy and grid interconnect already in place do not need off-grid gas. The math on continuous natural gas as primary power makes sense specifically for operators who are blocked on grid timing and want a path that does not depend on storage scaling.
This is not a cheap-power story either. Natural gas in 2026 is competitive with grid-delivered power on a fully loaded basis, especially after factoring in the cost of grid queue delay. It is not 2014-style cheap natural gas. The case is reliability and timing, not unit cost.
And this is not a turnkey story. The operator still has to build the generation plant, contract the EPC, manage the permitting, and run the operations. The gas allocation solves the upstream fuel problem. Everything downstream of the meter is the operator's lane.
The honest framing
Stone Path facilitates the introduction between operators and KYTX, the company that owns the field. Gas is allocated project by project at the field owner's set price. Land adjacent to the field is available for buildout through a separate landowner, also reachable through the same intro.
If a campus is sized for a workload that 1 MW continuous can support, and grid timing is the constraint, the math is worth running. If the campus needs 100 MW from day one and cannot phase, this is the wrong tool. Most real campus plans phase, which is why the conversation is worth having.
Walk through what 1 MW continuous would look like on your campus and we will scope phasing against your build timeline.