Meta reserves capacity in two pre-commercial energy technologies for AI data centers

Two reservations, both attached to 2028 demonstrations

Meta announced deals with two energy companies on the same day. With Overview Energy, it reserved up to 1 GW of space solar — satellites in geosynchronous orbit roughly 22,000 miles up that collect sunlight and beam it down to existing solar farms as low-intensity, near-infrared light. With Noon Energy, it reserved up to 1 GW/100 GWh of ultra-long-duration storage built on reversible solid oxide fuel cells and carbon-based storage, rated for over 100 hours.

The important detail is timing. Overview's orbital demonstration — the first time it would beam energy from space to a solar farm — is planned for 2028, with commercial grid delivery starting 'as early as 2030.' Noon's demonstration also targets 2028, and the reservation begins with a modest 25 MW/2.5 GWh pilot. Neither headline gigawatt figure describes power flowing today. They describe options Meta has purchased on technologies that have not yet been proven at any scale.

The constraint Meta is naming out loud

The announcement is unusually candid about the limits of the clean energy Meta already buys. It states plainly that solar depends on sunlight, wind depends on weather, and the grid still lacks the storage to make full use of either. That framing matters because Meta has already contracted more than 30 GW of clean and renewable energy and backs 7.7 GW of nuclear. The problem it is solving isn't total capacity — it's availability across the full 24-hour cycle that AI training and inference run on.

Both new partnerships attack that specific gap. Overview aims to make solar farms produce during the hours they currently sit idle. Noon aims to carry clean power across days rather than the few hours lithium-ion delivers. Read together, they are two attempts to convert intermittent generation into the baseload profile a data center actually consumes.

Why the reuse of existing infrastructure is the real argument

Because the technology will build on solar infrastructure that's already in place rather than requiring new facilities, it can come online faster at scale than traditional buildouts.Montana Labs

This line is the load-bearing part of the Overview case. The pitch is not that space solar produces cheaper electricity in the abstract, but that it raises the output of solar farms that already exist — no new land, no new grid interconnection. Noon makes a parallel claim: its modular design lets storage capacity grow alongside Meta's data center footprint without building entirely new infrastructure.

For a company adding data centers on a compressed schedule, incremental capacity from installed assets is worth more than greenfield projects that take years to permit and connect. That is why Meta frames early-stage support as rational rather than speculative — the value is in unlocking existing capacity faster, not waiting for a fully mature technology.

What committing to unproven power supply signals about AI planning

The practical implication is that Meta is now treating energy the way it treats compute: something to be secured years ahead of need, through reservations on capacity that may or may not materialize. A capacity reservation for space solar and one of the industry's largest commitments to ultra-long-duration storage are hedges against a future in which AI demand outruns the grid's ability to deliver clean, round-the-clock power.

For applied AI teams, the lesson is not about satellites or fuel cells specifically. It is that the largest operators now consider the physical energy system a first-order dependency of AI roadmaps — enough to fund technologies whose first demonstrations are still two years out. The risk Meta is accepting is real: if the 2028 demos fail, these gigawatts stay theoretical. The bet is that the downside of not having reliable power lined up is worse than the cost of backing something early.