The man who built Tesla on solar hype now thinks sunlight is better used 340 miles up.
The Summary
- Elon Musk's xAI is building data centers powered entirely by natural gas, abandoning the solar-electric vision he sold for two decades
- SpaceX is simultaneously developing orbital data centers that would run on space-based solar, suggesting Musk sees Earth's energy grid as unfixable
- The shift reveals what AI companies won't say publicly: training frontier models requires energy density and reliability that renewables can't yet deliver at scale
The Signal
xAI's new Memphis facility is a 100-megawatt natural gas facility with no solar panels in sight. Not even the token rooftop array for PR. The company filed permits for direct gas hookups and is building what amounts to a captive power plant adjacent to the GPU clusters. This from the CEO who spent 15 years promising that Tesla Energy would power the world on sunshine and batteries.
The numbers tell the story. Training a frontier model like Grok-3 requires sustained power delivery with near-zero downtime. Intermittency kills training runs. A cloud passing over your solar array can cascade into millions in wasted compute. Natural gas gives you 95%+ capacity factor. Solar in Tennessee gives you maybe 20% after accounting for weather and nighttime.
"AI training doesn't wait for sunny days. The economics brutally favor always-on energy."
Meanwhile, SpaceX is quietly advancing orbital data center prototypes that would run on space-based solar arrays. The pitch: 24/7 sunlight, no atmosphere, no permitting battles, and you can radiate waste heat directly into the void. Early tests suggest latency penalties around 40-80ms for consumer applications, but for batch training jobs, that's irrelevant. You send the job up, the models train in orbit, you download the weights.
The contradiction isn't hypocrisy. It's Musk reading the same data everyone else has and concluding Earth's energy infrastructure won't scale fast enough for the AI buildout. Solar works great for cars that charge overnight and houses that can buffer with batteries. It breaks down when you need a gigawatt of continuous power for a training cluster that can't ever shut off.
Key facts:
- xAI's Memphis facility: 100MW natural gas, zero solar capacity
- SpaceX orbital data centers targeting 2027 launch for first commercial prototype
- Industry-wide shift: 73% of new AI data center capacity announced in 2025-2026 is gas-powered
What's telling is the silence from the climate tech world that championed Musk for years. They know the math too. If you're serious about AGI timelines, you optimize for training speed, not carbon accounting. Tesla can still sell Powerwalls to homeowners. But xAI needs electrons, now, all the time. Natural gas delivers that. Solar doesn't. Not yet.
The orbital play is more interesting long-term. Space-based solar has been vaporware for 50 years, but combining it with data centers changes the economics. You're not beaming power to Earth, you're using it in-place for the one application that can tolerate the latency and can't tolerate downtime. If SpaceX proves the model, every hyperscaler will follow.
The Implication
Watch the divergence between what AI companies say about sustainability and what they actually build. xAI just told you that frontier model training is incompatible with intermittent renewables at current tech levels. OpenAI, Anthropic, and Google will face the same choice. Some will greenwash with carbon credits. Others will follow Musk to gas and eventually to orbit.
For anyone building in the agent economy, this matters. Inference is cheap and can run anywhere. Training is expensive and demands always-on power. The companies that secure reliable energy at scale will train bigger models faster. Geography will matter less than physics.