Google's Project Suncatcher reaches orbit: in-orbit chip lifespan becomes a performance metric

Chips in Space: Lifetime Is Performance

Source: Semiconductor Horizons Published: 2026-10-26 17:56 Chip Design Google generated poster

Part of a processor's value depends on how much useful computing power it can produce for a spacecraft before its hardware ages or fails.

On October 1, a prototype satellite of Google's "Project Suncatcher" rode a SpaceX launch mission into orbit. The satellite, jointly developed by Google and Planet, has now established a communications link with the ground. Its mission is to test how commercial AI hardware performs in the space environment — once an in-orbit processor fails, it cannot simply be replaced at the next maintenance window like ground-based equipment.

The appeal of in-orbit computing power is not hard to understand. Some orbits can receive sunlight for long periods, and launch costs continue to fall, so many proposals call for deploying computing infrastructure beyond Earth. However, this satellite's successful arrival in orbit still cannot prove that in-orbit data centers are commercially economical.

A commercial in-orbit computing facility needs to generate power, distribute it, dissipate heat, transmit data at high speed, and operate continuously and stably for many years. All subsystems must be launched into space with the satellite; once hardware fails, it shortens the service life of the entire high-value payload, and its manufacturing cost is already very high.

For chip designers, this changes the logic of performance evaluation. Part of a processor's value depends on how much useful computing power it can produce for a spacecraft before its hardware ages or fails.

The supporting infrastructure behind the chip

First, ensuring continuous chip operation requires a whole set of supporting facilities.

A spacecraft needs power generation and distribution systems, energy storage solutions for periods of insufficient sunlight, and radiative radiators to expel waste heat. The vacuum of space does not allow convective cooling; heat must be conducted or pumped to a radiating surface and released outward. High-speed inter-satellite and satellite-to-ground data links add further hardware weight, and every kilogram of payload drives up the launch budget.

A study published by Slava Turyshev in April 2026 analyzed the engineering conditions and economic thresholds required to realize in-orbit computing. In one representative 1-megawatt computing scenario, the model requires about 5,640 square meters of photovoltaic arrays and 2,500 square meters of radiative cooling surface.

The above is only a model calculation result, dependent on a series of assumptions such as solar power generation and heat dissipation capacity. It intuitively shows the hardware scale of a 1-megawatt in-orbit computing facility; even so, compared with ground-based hyperscale data centers, this scale is still on the small side.

Turyshev's model also accounted for launch costs, satellite manufacturing costs, equipment utilization, and service life. Under one set of assumptions, for this scheme to be competitive, the combined cost of launch and satellite manufacturing must be below the baseline corresponding to the dedicated launch price of Falcon 9. This comparison result depends heavily on the choice of launch scheme and cannot predict future launch price trends.

In an assessment report in April, the U.S. Government Accountability Office pointed out that heat dissipation, communications, and deployment economics remain unresolved challenges. An early scenario that is easier to realize is processing observation data directly on the satellite locally: only less raw data needs to be transmitted back to the ground. A general-purpose in-orbit data center, by contrast, faces much more complex infrastructure challenges.

Processor service life becomes a performance evaluation metric

The processor itself also determines the scale and cost of the supporting facilities.

Accelerators with more computing power and higher power consumption require larger photovoltaic arrays, power distribution equipment, and radiative cooling surfaces. These supporting structures increase payload weight and may offset the benefits brought by high-throughput computing power. The ultimate trade-off depends on the actual workload. Software efficiency, bit error rate, equipment utilization, and expected service life all change the comparison between two chips with similar benchmark scores.

For example, two chips run the same business workload: one has higher operations per second but requires a massive cooling system; the other has lower computing power and lower power consumption, can be built into a smaller satellite, and may have a longer mission cycle.

What really matters is the total computing power produced over the entire mission cycle, which must also account for the probability of interruptions, performance degradation, and premature failure. A processor with impressive benchmark results may actually have fewer effectively usable hours before the satellite is retired. For operators, failed boards are difficult to replace, and the loss of computing lifetime translates directly into economic loss.

What Google's experiment can verify

"Project Suncatcher" can gradually accumulate the in-orbit operating data necessary for such comparisons. The current experimental satellite's hardware scale is far below that of a commercial data center.

Before launch, Google completed vibration testing and proton beam irradiation testing on the Trillium TPU. Google says the chip's tolerance to cumulative ionizing radiation performed as expected, and the team is also studying various errors that occur when AI workloads run. Laboratory irradiation can only preliminarily verify radiation tolerance; the frequency of failures and the impact of failures during years of in-orbit operation still need to be measured in space.

Google also tested a thermal management system consisting of heat pipes plus radiative radiators in a vacuum simulation environment. Whether the hardware can stably carry business workloads when temperature and operating conditions change is also part of what this in-orbit experiment verifies. The longer the satellite operates in orbit, the more useful data we can obtain on device aging, error recovery, and sustained computing output.

Communication links are another constraint. Google plans to conduct a dual-starlight link experiment in 2027 to verify high-bandwidth inter-satellite connections. Distributed AI tasks often require high-frequency data exchange between processors, and the network architecture directly determines the economics of the entire system.

The TPU's ability to operate stably in orbit is an important technological milestone. To build a commercially competitive in-orbit computing service, the entire system—power generation, heat dissipation, and communications—still needs to work together stably at a larger scale.

NASA's Space Processor Approach

NASA's High Performance Spaceflight Computing (HPSC) project represents another processor design approach. The project is developing the HPSC processor in collaboration with Microchip Technology, targeting demanding spacecraft missions and focusing on meeting the requirements of autonomous operation, radiation tolerance, power management, and fault-tolerant design.

NASA published preliminary test results in May 2026: this chip's performance is approximately 500 times that of existing radiation-hardened space processors. The project's early goal was to increase computing power to 100 times that of existing space computers. The two figures correspond to different comparison bases and test stages, and the chip's qualification work is still underway.

HPSC targets spacecraft autonomous control and scientific exploration missions; Google's TPU is a commercial AI accelerator. The two cannot be directly compared on benchmark performance metrics, and the workloads they carry are completely different.

It is worth noting that NASA made fault tolerance, adaptive power consumption, and radiation tolerance core metrics from the very beginning of the design. This also explains why the evaluation criteria for in-orbit processors differ from those for terrestrial chips.

The Full Lifecycle Cost of Computing Power

The core investment criterion: how much reliable computing power can an in-orbit facility produce over its full lifecycle.

The accounting must include equipment utilization, the power generation system, heat dissipation hardware, the communications system, launch and satellite manufacturing costs, the expected failure rate, and spare parts replacement costs. It also depends on where the data is generated and where the computing results must be delivered.

Local processing of satellite imagery can avoid transmitting massive amounts of raw observation data back to the ground. Training large models in space, however, places much higher demands on bandwidth between accelerators and continuous power supply capacity. Before arguing whether a plan is economically viable, these two types of workloads must be modeled and evaluated separately.

This is very similar to the classic trade-off logic in the semiconductor industry: moving to a more advanced process node is only worthwhile when improvements in performance, density, or energy efficiency can cover production costs and product demand. In the in-orbit scenario, expected lifespan and the weight of accompanying payloads carry much greater weight in the trade-off.

Before that, chipmakers need to focus on one key question: before a spacecraft reaches end of life, how much reliable computing power can a single processor stably deliver. In the in-orbit scenario, effective service life is itself a performance metric.

This article is reproduced from media reports or online platforms and represents the author's personal stance or views. Our reproduction is solely for sharing and does not mean we endorse or agree with it. If the source attribution is incorrect or your lawful rights and interests are infringed, please contact customer service promptly, and we, as a neutral platform service provider, will promptly correct, delete, or handle it in accordance with the law. Comments: No user comments yet. Latest News More Related Recommendations

SK Hynix plans to launch a new shareholder return plan, up to 60 trillion won 2026-10-26 18:10

Cambricon's Chen Tianshi: The company's next-generation intelligent processor microarchitecture will focus on optimizing training and inference scenarios 2026-10-26 18:10

Changchuan Technology: Customer demand is currently strong, and product orders are plentiful 2026-10-26 18:09