In December 2025, the Financial Times reported that Big Tech’s AI data-center plans would require an estimated 44 GW of additional capacity by 2028. In November 2026, Google agreed to a dedicated Blue Origin launch of TPU-powered satellites in Q1 2029. Google’s data-center energy head says the United States has enough electricity in aggregate.
Key takeaways
- Big Tech’s AI data-center plans were estimated to require 44 GW of additional capacity by 2028, according to a December 2025 Financial Times report.
- Google agreed on November 20, 2026, to a dedicated Blue Origin New Glenn launch of TPU-powered Project Suncatcher satellites in Q1 2029.
- Rune raised $40 million in September 2026 to deploy modular, off-grid compute at solar farms and other renewable-energy sites.
- Crusoe and Redwood Materials opened a Nevada data center powered by repurposed EV batteries and equipped with 2,000 GPUs in June 2025.
- Blue Origin has sought FCC approval for nearly 52,000 solar-powered satellites in its proposed Project Sunrise system.
Firm power, cooling equipment and permits must arrive together before an operator can run a large GPU cluster. Builders are moving compute toward controllable energy through on-site generation, batteries and modules beside renewable projects. Google’s orbital experiments take that siting question beyond the grid.
A megawatt only counts when a site can use it
The 44 GW estimate is a forecast of what Big Tech’s plans need, not electricity already committed to those sites. An operator cannot run GPUs on aggregate national generation if the local connection, cooling system or permitted site is missing.
The Wall Street Journal reported in April 2024 that the rush to build data centers had produced shortages of property, power and parts; cooling-system lead times had reached five times their earlier length. By 2026, AI labs were deploying on-site gas generators rather than waiting for grid service. A generator can reduce dependence on a grid connection, but it does not install cooling equipment or secure a permit.
Google’s data-center energy head pointed to slow permitting, sharpening the question behind data-center capacity underwriting: who has firm service at a specific location, and when can the facility use it?
Rune and Crusoe move GPUs toward the power
Crusoe and Redwood Materials launched a Nevada data center with 2,000 GPUs powered by repurposed EV batteries in June 2025. AI labs’ on-site gas generators take a different route to local supply. In September 2026, Rune raised $40 million to deploy modular, off-grid compute at solar farms and other renewable-energy sites. Each brings the cluster within reach of an energy system its operator can control.
Exowatt offers another version: the company raised $120 million around concentrated solar power for data centers, with a stated target of one cent per kilowatt-hour. That price is an ambition, not a demonstrated cost across data-center deployments. Rune’s modules and Exowatt’s power system answer different questions—where to put the computers, and how cheaply a nearby system might supply them.
Moving compute has a cost of its own. Chip Huyen notes in AI Engineering that remote electricity can be unattractive for inference workloads with strict network-latency requirements. A solar-adjacent site may suit one workload and fail another’s response-time budget. Operators that build more of their own energy-to-compute chain gain control over its timing, but take responsibility for more of the engineering. That trade-off is already visible in the split between shared-grid dependence and integrated infrastructure.
Orbit adds a spacecraft to the data-center design
Spaceborne computing began as a hardware test, not a capacity plan. In 2018, HPE’s Spaceborne Computer made a teraflop-class system available for scientific use in orbit; coverage reported that nearly half its hard disks failed after solar-radiation exposure. A satellite carrying AI chips must still keep those chips powered, cooled, connected and operational without a technician arriving at the rack.
Starcloud reported in December 2025 that it had run an Nvidia H100 in orbit and queried Google’s Gemma; it also reported training an LLM there. Google’s Project Suncatcher prototype took a deliberately small step: four TPUs on a mission designed to last a year. After its October 2026 launch, Google said the prototype was online and running Gemma 4. Those demonstrations establish that useful compute can run on the hardware. They do not establish how to service a large, sustained cluster.
At scale, satellites need solar collection and storage to feed their processors, thermal hardware to reject heat, and ground and inter-satellite links to carry work and results. Amazon validated optical links for its Kuiper satellites in 2023, showing that one networking component can work in orbit. A fast link between satellites does not, by itself, settle the latency of a request traveling to and from Earth. Orbital compute has to be designed as one distributed machine, rather than servers with a rocket ticket.
Blue Origin sells the launch and seeks a fleet
Google’s November 20 agreement with Blue Origin puts a dedicated New Glenn launch of TPU-powered Project Suncatcher satellites on the schedule for Q1 2029. It moves Google beyond a four-TPU feasibility mission to a planned heavy-lift deployment. The launch is still a plan, not evidence of commercial-scale orbital capacity.
Blue Origin has separately sought FCC approval for nearly 52,000 solar-powered satellites in its proposed Project Sunrise AI-data-center system. SpaceX has sought permission for as many as one million satellites it says could use solar power for AI data centers. Neither filing is a deployed constellation or an exemption from spectrum and collision rules. Launch providers see orbital compute as a potential customer and, in Blue Origin’s case, a system they might operate themselves.
Starcloud was reported to be valued at $1.1 billion after a $170 million Series A in March 2026. By August, its reported valuation was $2.3 billion after a $250 million extension that included a reported $25 million Nvidia investment. Financing can purchase hardware work and launch opportunities. It cannot turn a prototype’s operating history into a reliability record for a fleet.
Earth remains the cheaper repair shop
Google’s data-center energy head said orbital facilities could cost up to 100 times as much as Earth-based ones. A separate analysis estimated $51.10 per watt of orbital compute against $15.85 per watt for terrestrial capacity. Those estimates use different comparisons and should not be treated as interchangeable price quotes. Both put a substantial burden on claims that sunlight alone makes orbit economical.
HPE’s failed disks expose a reliability cost alongside the launch bill. Heat rejection and limited repair access add engineering costs; ground communications impose workload constraints. NASA has also warned that a proposed broadband mega-constellation carried unusually high collision risk. Moving a facility beyond a local permitting process does not move it beyond regulation: satellite operators still need spectrum access and a workable place in orbit.
Frequently asked questions
How much computing capacity will Google’s Q1 2029 Project Suncatcher launch carry?
The piece does not specify a TPU count, power capacity, satellite count, or operating lifetime for the planned dedicated New Glenn mission. Google’s earlier prototype had four TPUs and was designed for a one-year mission, but that does not establish the scale of the 2029 deployment.
What is the output or capacity of Rune’s off-grid compute modules?
The article says Rune raised $40 million to deploy modular compute at renewable-energy sites, but it gives no per-module GPU count, megawatt rating, or deployment schedule. Those figures would determine whether the modules can serve isolated workloads or larger clusters.
Has the FCC approved Blue Origin’s proposed Project Sunrise constellation?
The piece says Blue Origin has sought approval for nearly 52,000 satellites; it does not report an FCC approval or a deployed constellation. The proposal would still face spectrum and collision-rule requirements.
Does Exowatt’s one-cent-per-kilowatt-hour target represent the delivered cost of running AI workloads?
No such all-in cost is established in the piece. Exowatt’s stated target concerns concentrated solar power for data centers, and the article says it is an ambition rather than a demonstrated cost across deployments.
What workloads did Starcloud run on its H100 in orbit?
Starcloud reported running an Nvidia H100 in orbit, querying Google’s Gemma, and training an LLM there. The piece does not identify the model size, training duration, workload throughput, or network performance.
Project Suncatcher: prototype to planned dedicated launch
- October 1, 2026 — Google planned to launch Project Suncatcher’s experimental four-TPU MVP, designed to operate for one year.
- November 20, 2026 — Google agreed with Blue Origin on a dedicated New Glenn launch for TPU-powered Project Suncatcher satellites.
- Q1 2029 — The dedicated heavy-lift Project Suncatcher launch is scheduled to occur.
Google’s planned New Glenn mission can test components and integration that a Nevada battery site cannot. Rune’s modules can seek usable power without paying to launch, cool and communicate with a spacecraft. An operator can justify moving a workload to orbit only if its total cost and network performance beat the terrestrial options available on that workload’s schedule. The estimated 44 GW in Big Tech’s plans needs terrestrial addresses by 2028, ahead of Google’s dedicated orbital launch in Q1 2029.