SpaceX plans on-site gas plants and batteries to run $16.8 billion Terafab off the grid
Behind-the-meter power moves the reliability problem rather than removing it, and that trade-off is worth understanding before you design a site.
SpaceX intends to run its planned Terafab semiconductor and advanced-computing campus in Grimes County, Texas, on electricity generated on site rather than drawn from the grid, according to a tax-abatement agreement with the county reported by Data Center Knowledge on 11 August 2026. The fully executed agreement states that the facilities are expected to be powered by on-site plants and are not expected to take grid electricity once operational.
Riley Trettel, who leads energy and data centre development for SpaceX, told county commissioners the company would bring its own power in the form of natural gas-fired plants together with very large battery arrays for storage. The agreement also references a battery storage project and describes plans for on-site wastewater treatment and for minimising groundwater use.
The project was announced on 6 August, when SpaceX and Tesla committed an initial $16.8 billion to build a campus of roughly 100 million square feet. Data Center Knowledge described it as a four-phase development with a potential total investment of $119 billion, intended to make chips for Tesla vehicles, the Optimus robot and SpaceX systems.
Not everyone reads self-supply as a complete answer. Neil Osnato, founder of Persistence Analytics Group, said behind-the-meter generation relocates the reliability problem rather than eliminating it, since the operator now carries the responsibility a utility would otherwise hold. The report also noted that the agreement's wording, which says the site is not expected to use grid power, stops short of ruling out a connection to the Texas grid operated by ERCOT, leaving the final electrical architecture unclear.
Why it matters
Large AI and semiconductor loads are increasingly choosing to generate their own power because waiting for a utility interconnection can take years. Gas turbines paired with batteries have become the default recipe for doing that quickly, which ties the compute build-out to gas supply and to local air and water permitting. Taking the grid out of the picture does not remove the need for redundancy; it moves it onto the operator's balance sheet and engineering team. For learners, the lesson is that power architecture is now a first-order design decision for any large computing site.
Large AI and semiconductor loads are increasingly choosing to generate their own power because waiting for a utility interconnection can take years.
IT and fab load
Chip manufacturing and computing equipment; demand can swing sharply as jobs start and stop.
UPS and distribution
Conditions power and bridges the milliseconds between a disturbance and the battery response.
Battery arrays
Very large battery storage covers seconds to minutes of ride-through and smooths spikes.
On-site natural gas plants
Steady generation for the bulk of demand; the operator, not a utility, owns fuel and maintenance.
Grid tie: not expected to be used
The county agreement says the site is not expected to draw grid power, without ruling out a connection.
What you can learn from this
- Behind-the-meter generation. Behind-the-meter means electricity is produced on the customer's side of the utility meter and consumed directly, so it never crosses the public grid. The main attraction is time: a new large load normally waits in an interconnection queue while the utility studies how it would affect the network, and self-supply skips that study. The cost is that every function a utility performs, from fuel procurement to maintenance to reserve capacity, becomes the operator's job. That is what the comment about relocating rather than eliminating reliability risk refers to.
- Why gas turbines and batteries are paired. Gas turbines and reciprocating engines provide steady output for hours or years, but they respond slowly to sudden changes in demand and take time to start. Batteries respond in milliseconds, so they cover the seconds and minutes between a load spike or a generator trip and the moment other generation catches up. Computing loads, and AI training in particular, can swing sharply as jobs start and stop, which makes that buffering essential. The two technologies together approximate the stability a large grid provides through sheer scale.
- Interconnection queues and time-to-power. Before a large customer connects, the grid operator must confirm that lines, transformers and generation can carry the extra load without harming other users, and any upgrades must be built. Those studies and upgrades commonly take several years for loads measured in hundreds of megawatts. Operators speak of time-to-power as the binding constraint on new capacity, which is why sites are chosen for available electricity as much as for land or fibre. Self-generation is one of several responses, alongside locating next to existing plants.
- Redundancy: N, N+1 and 2N. Engineers describe redundancy by how many spare units exist beyond what the load needs. N means exactly enough, N+1 adds one spare so a single failure does not interrupt service, and 2N duplicates the whole system. A grid connection quietly provides enormous redundancy because thousands of generators back each other up; an isolated site must buy that resilience explicitly with extra turbines and battery capacity. The number of spares chosen is a direct trade between capital cost and tolerable outage risk.
- Water as a constraint. Semiconductor fabrication and high-density computing both consume large volumes of water for process steps and cooling. Rural sites often depend on groundwater, so commitments to treat wastewater on site and limit aquifer use appear in agreements alongside power. Water and power interact: evaporative cooling saves electricity but uses more water, while dry cooling does the opposite. That trade-off explains many siting decisions that look puzzling from the outside.
We teach this
How to use this in practice
- Draw a single-line diagram of an off-grid campus. Sketch gas-fired generators feeding switchgear, a battery system on the same bus, uninterruptible power supplies, and finally the IT or fab load, with a dashed optional grid tie. Mark each point where a single failure would drop the load and note what spare would prevent it. Done looks like a one-page diagram with at least three failure points labelled and a spare or bypass drawn for each.
- Work a redundancy calculation by hand. Assume a 100 megawatt load. Calculate how much battery energy is needed to ride through fifteen minutes, then how many 50 megawatt generating units are required for N+1 and for 2N. Extend the exercise by asking how the numbers change if the load can shed a quarter of its demand during an outage. Done looks like a half-page worksheet with the arithmetic shown and a sentence on which configuration you would choose and why.
- Read the grid operator's large-load process. Find ERCOT's public documentation for connecting large loads and summarise the steps a new site must complete, from initial request through study to energisation. Note which steps involve the transmission provider and which involve the market operator. Done looks like a numbered list of the steps in your own words and an estimate, taken from the documents themselves, of how long the sequence typically takes.
- Test the smallest power system you own. If you have an uninterruptible power supply at your desk, check its reported runtime at current load, then unplug it from the wall and time how long your equipment actually stays up. Compare the two numbers and check the battery's age; sealed lead-acid batteries lose capacity as they age, which is why measured runtime often falls short of the estimate. Done looks like a recorded runtime, a replacement date if the battery is old, and a sentence on why the estimate and the measurement differ.
Sources
- SpaceX Plans to Power $16.8B Terafab Without the Grid — Data Center Knowledge
Our reporting is an original summary; full coverage is at the links above.
Don't just read about it — build it.
Square 1 teaches the skills behind the headlines, with every line of your work graded by AI. Find your starting point in 3 minutes.
Get your free skill report