Put computing power into orbit, and solve the power problem in a more thorough way. The timeframe for this initiative is measured in decades—ten years.
A technology company has announced a plan to mount its in-house inference chips on satellites and form a solar-powered data center in low Earth orbit. The first batch of two prototype satellites will be manufactured and operated by a satellite imaging company, with launches expected in early 2027. Each of the two test satellites will carry a set of chips, using a dawn-dusk orbit to achieve nearly continuous sunlight.
The plan itself is restrained. The two prototype satellites will first validate communications and thermal management. All economic feasibility is not expected to be possible until the middle of the 2030s. In the near term, it cannot replace terrestrial data centers. The company defines it as a validation project.
The motivation is an energy constraint. Power consumption for training and inference is growing faster than the expansion of the power grid. On the ground, projects are becoming increasingly difficult in terms of site selection, power supply, and approvals. By swapping the power source to sunlight, the most expensive cost item is essentially moved into orbit.
For the industry, the implication lies in how the cost curve is shaped. If the challenges of cooling and communications can be solved through engineering, the marginal cost structure of computing power will be rewritten, and the basis of comparison will shift to the electricity prices and land costs on the ground. Otherwise, this initiative will remain limited to the stage of two test satellites.
Moving the data center into the sky requires first calculating the return on investment.
#人工智能 #算力
A technology company has announced a plan to mount its in-house inference chips on satellites and form a solar-powered data center in low Earth orbit. The first batch of two prototype satellites will be manufactured and operated by a satellite imaging company, with launches expected in early 2027. Each of the two test satellites will carry a set of chips, using a dawn-dusk orbit to achieve nearly continuous sunlight.
The plan itself is restrained. The two prototype satellites will first validate communications and thermal management. All economic feasibility is not expected to be possible until the middle of the 2030s. In the near term, it cannot replace terrestrial data centers. The company defines it as a validation project.
The motivation is an energy constraint. Power consumption for training and inference is growing faster than the expansion of the power grid. On the ground, projects are becoming increasingly difficult in terms of site selection, power supply, and approvals. By swapping the power source to sunlight, the most expensive cost item is essentially moved into orbit.
For the industry, the implication lies in how the cost curve is shaped. If the challenges of cooling and communications can be solved through engineering, the marginal cost structure of computing power will be rewritten, and the basis of comparison will shift to the electricity prices and land costs on the ground. Otherwise, this initiative will remain limited to the stage of two test satellites.
Moving the data center into the sky requires first calculating the return on investment.
#人工智能 #算力
