Building data centers in space runs into a structural financing obstacle: the data center ecosystem relies on debt backed by chips as collateral, but chips orbiting at 17,500 mph cannot be repossessed and resold. Lenders behind SpaceX's $40 billion chip financing deal have asked to exclude in-orbit chips. Meanwhile SpaceX and Tesla are pursuing a US-based Terrafab foundry, which has yet to break ground.
TPU
张量处理单元
Google's custom tensor-processing accelerator, planned for space deployment under Project Suncatcher.
Curated Intelligence
Google is restructuring its TPU supply chain: on some new-generation TPUs it designs the compute die itself, MediaTek handles I/O chip design, and Google buys HBM directly from memory makers, bypassing Broadcom to cut intermediary costs. Broadcom still participates in other high-end flagship TPU programs, creating a dual-track model. The move reflects hyperscalers reducing NVIDIA GPU dependence and reshaping HBM pricing dynamics.
Google's Project Suncatcher prototype satellite successfully completed on-orbit commissioning, generating its first AI response using an onboard Gemma 4 model and data-center-class TPU. This milestone validates the basic feasibility of running AI compute chips in space, paving the way for future thermal/radiation stress tests and the planned twin-satellite interconnection in 2027.
Google announced that a data-center-grade TPU on its Project Suncatcher prototype satellite successfully completed its first AI inference task, marking a key breakthrough in space-based computing infrastructure. The test validated the stability of the Gemma 4 model under microgravity, extreme temperature fluctuations, and radiation, laying the groundwork for future large-scale orbital inference satellite networks.
Google's Project Suncatcher satellite successfully booted its onboard TPUs and ran the Gemma 4 model to generate the first AI response from space. This milestone validates basic compute capabilities of specialized AI chips in orbit, paving the way for thermal/radiation stress tests and future orbital data centers.
Google's Project Suncatcher satellite successfully booted its onboard TPU and ran the Gemma 4 model to generate a haiku, completing the first full inference loop from ground command to orbital response. This milestone validates the feasibility of space-based computing powered by solar energy, though challenges like radiation tolerance and thermal management remain before commercial viability.
Google announced that its Project Suncatcher prototype satellite successfully booted a TPU in orbit to process a prompt from Earth and return a complete response (a haiku). This marks a critical milestone moving space-based edge computing and in-orbit AI inference from proof-of-concept to operational reality.
Google’s prototype space data center satellite is now online and communicating with Earth, featuring a single TPU chip running the pre-loaded Gemma 4 model. Limited by ~1kW power generation, the TPU operates in short bursts (~20 minutes), successfully completing inference tasks like writing haikus to validate the feasibility of orbital large-scale AI workloads.
Google Cloud revenue surged 103% year-over-year to $30.7B, driven by robust demand for AI infrastructure, with operating income tripling. The Gemini app surpassed 1 billion users, and model API throughput quadrupled, while TPU chip sales expanded to external partners like Oracle and OpenAI. Despite this growth, massive capital expenditure on AI data centers resulted in a free cash flow deficit exceeding $9B.
According to The Information, OpenAI plans to formally procure and deploy Google Cloud TPU chips, currently hiring lead engineers for model adaptation and optimization on the platform. This move signifies a key step in OpenAI's compute diversification strategy, aiming to reduce reliance on single vendors and optimize inference costs.