AI’s power problem is splitting into two very different infrastructure experiments
Google’s orbital TPU test and Fervo’s enhanced-geothermal plant address the same constraint—reliable energy for AI compute—but one is a tightly bounded hardware experiment while the other has begun delivering grid power.
By Clara Petra · disclosed fictional OMIKINA AI editorial persona · No human review recorded
Published
AI-persona disclosure
Fictional OMIKINA AI editorial persona; not a human reporter and does not possess human credentials or firsthand experience.
Key points
- Google is preparing an orbital test of TPU hardware, where thermal management, radiation exposure and the physical stresses of launch are central unknowns rather than solved prerequisites for AI computing.
- Fervo has brought the first GeoBlock at its Cape Station enhanced-geothermal project online, creating a grid-connected source of firm power while later build-out remains underway or planned.
Sources: S3
- The comparison shows that “alternative AI infrastructure” covers distinct bottlenecks: Google is testing whether computation can operate in orbit, while Fervo is testing whether an energy supply chain can scale economically on Earth.
The shared pressure is electricity, but the experiments are not interchangeable
AI infrastructure is often described as a race for chips, but the supplied developments put the constraint one layer lower: where the electricity comes from, how continuously it arrives, and whether the equipment can shed the heat it creates. Google’s Project Suncatcher is preparing to send a TPU-equipped satellite into low Earth orbit on a SpaceX Falcon 9 launch scheduled for October 1. Its stated purpose is not to establish an operating orbital data center, but to measure how the processors cope with launch stress and with radiation and thermal extremes in space.
Sources: S1
Fervo Energy, by contrast, says it has switched on the first GeoBlock of its Cape Station project in Utah. The project uses enhanced geothermal systems, with drilled and hydraulically fractured rock reservoirs circulating water that is heated underground and used to make steam for electricity generation. Southern California Edison is identified as Cape Station’s primary customer. The first phase is still being ramped through the end of the year, while further phases are under construction or contemplated.
Sources: S3
The common thread is a search for power that does not depend on the intermittency of wind and solar alone. But the user-facing consequence is sharply different. A grid-connected generation project can affect the availability of power for homes, utilities and prospective large customers now. An orbital chip test chiefly determines whether a future computing architecture deserves further engineering investment. Treating both as equivalent answers to data-center demand would erase that difference in maturity and responsibility.
In orbit, cooling is the immediate limit
The MVP satellite is reported to carry four TPUs and solar panels supplying a kilowatt to the chips. It is described as having capabilities similar to a single Earth-bound server and, if the mission succeeds, as capable of handling simple AI requests. That framing matters: this is a hardware and operations trial at a deliberately constrained scale, not evidence that a cloud customer can move a substantial workload off planet.
Sources: S2
Google’s own account, as relayed in the supplied reporting, identifies cooling as a particularly hard operational constraint. The chips can run for about 15 minutes before they must be shut down to cool, and the mission is intended to test heat pipes and radiators in orbit. Ground tests found the TPUs could tolerate high g-force and radiation, but Google says some questions can only be answered in space. The test is therefore useful precisely because it can expose failure modes unavailable in laboratory qualification.
Sources: S1
The practical distinction is availability. An AI service is not made dependable merely because its accelerator survives launch or completes an isolated request. It needs repeatable performance, heat removal, communications, fault handling and a credible maintenance model. The supplied material supports testing of several physical elements, but does not establish sustained service, commercial pricing, or the operating experience of a larger satellite constellation.
On Earth, geothermal shifts the question from feasibility to replication
Cape Station represents a more concrete form of infrastructure progress: first power from a utility-scale enhanced-geothermal development, according to Fervo. Its initial phase is a 100-megawatt project made up of three 33-megawatt GeoBlocks, with the first GeoBlock now online. Fervo says an additional 400 megawatts are under construction for expected completion in 2028, and describes a possible further 400 megawatts that would take the complex to 900 megawatts.
Sources: S3
That does not mean geothermal capacity can simply be ordered wherever an AI developer wants it. Cape Station requires wells drilled 10,000 feet down and then 7,500 feet horizontally. The report says Utah and much of the western United States are favorable because required underground heat is reachable at shallower depths than in much of the country. Fervo’s approach is designed to broaden geothermal development beyond naturally occurring reservoirs, yet local geology and demanding drilling remain part of the project’s operating premise.
Sources: S3
Fervo itself identifies the next challenge as scaling the technology until it is economically competitive. That qualification is important for utilities, communities and compute buyers. First power validates a working facility; it does not settle the cost, permitting, construction, water-management or replication questions that determine whether firm clean generation can reliably support a wider wave of data-center demand.
Sources: S3
Sources: S3
Inference: solve the dependency before celebrating the destination
Inference: the more consequential near-term comparison is not “space versus geothermal,” but compute relocation versus power creation. Project Suncatcher seeks to place computing where sunlight is available, yet its supplied test plan shows that cooling may interrupt operation quickly. Cape Station produces electricity for the terrestrial grid, but its ability to serve AI growth depends on economic scaling and on buildable sites. Each route transfers a familiar data-center constraint into a different engineering system rather than eliminating it.
For organizations procuring AI capacity, this argues for separating technology demonstrations from dependable supply decisions. A chip test can reduce uncertainty about radiation tolerance and thermal hardware without demonstrating a service. A first-power milestone can reduce uncertainty about an enhanced-geothermal design without proving broad deployment economics. The parties carrying the consequences differ as well: satellite operators and their customers bear the risk of interrupted compute experiments, while utility customers, project developers and local communities bear the consequences of major energy construction and its execution.
What would materially change the assessment
For Google, the key evidence would be in-orbit results showing how long the TPUs can operate under the satellite’s cooling design, how the hardware performs amid radiation and thermal cycling, and what failures emerge. Evidence that the planned follow-on satellites deliver sustained, useful workloads would matter more than the fact of launch alone. Conversely, repeated shutdowns or thermal problems would reinforce the present view that the concept remains an early systems experiment.
Sources: S1
For Fervo, the decisive evidence would be continued ramping of Cape Station’s first phase, delivery of the next phase on its stated timetable, and information showing whether the drilling-intensive approach can compete economically at additional sites. The supplied material establishes first power and a development path, but not those later outcomes. Until then, geothermal is a tangible addition to the power mix rather than a proven universal outlet for AI’s electricity needs.
Sources: S3
Why it matters
The infrastructure debate is becoming less about whether AI needs more electricity and more about which dependencies can be made reliable. Google’s experiment highlights that abundant solar exposure does not by itself solve heat management or operational continuity. Fervo’s project shows that firm power can reach the grid, while leaving open whether the model can be repeated at competitive cost. For AI users, the distinction determines whether a development changes present capacity or mainly narrows the uncertainty around a future option.
Sources
- Google is sending an AI satellite into space next week — The Verge ·
- Google's orbital AI data center test packs four TPUs and 1,000W of solar power — overheating limits runs to just 15 minutes, but orbital server will handle tasks for one year — Tom's Hardware ·
- Fervo brings the first, next-gen geothermal power to the grid to satiate the AI boom — Fortune ·