Google is putting its first TPUs in orbit next week
Orbital compute stopped being a slide deck this week: Google is flying its own AI silicon, and Shanghai's banking regulator drew a lane for models that talk to customers.
Google will launch a prototype satellite on SpaceX's Transporter-18 rideshare next week carrying its first Tensor Processing Units to orbit — the first in-orbit test of Project Suncatcher, its research bet that space can host large-scale AI computing. The mission was built with satellite operator Planet, and Google says it is designed to collect data on how the chips handle launch forces, radiation and thermal extremes rather than to demonstrate a working orbital data center. That framing is deliberate. The company has been studying the idea since announcing Suncatcher last year, and the whole economic case rests on one physical fact: in low Earth orbit, satellites see near-constant sunlight and can generate up to eight times more solar power than the same panel on the ground.
The engineering details are the part that decides whether any of it scales. A rocket ride to low Earth orbit lasts about ten minutes and subjects the spacecraft to sustained loads of up to 10 g, with individual TPU chips seeing 50 to 100 g; Google says the hardware survived vibration testing on all three axes. On radiation, the team ran Trillium TPUs through a proton beam at UC Davis's Crocker Nuclear Laboratory while running AI workloads, and reports the chips withstood a total ionizing dose greater than a five-year mission would deliver. Cooling is the hardest problem, because a vacuum has no airflow and heat has to leave through radiators — Google is testing a heat-pipe and radiator design in a thermal vacuum chamber. The 2027 milestone is two satellites testing the high-bandwidth laser links that would let a cluster of them behave like one machine; future satellites are meant to carry dozens of chips each.
Keep it in perspective. Google is starting with the least glamorous question — does the chip survive — while the commercial side of in-orbit inference is already moving. We covered the other end of this race on Tuesday — Orbital inference starts selling: Yusuan Xinghe ships in October, where a Guangzhou startup says customers have already paid in advance for models running on its own satellites. Experts quoted by Reuters still put commercial orbital compute years away on launch costs, engineering constraints and satellite production bottlenecks. But a company bringing its own chips, its own launch partner and a five-year radiation budget is a different kind of entrant than a startup renting rides.
Shanghai's financial regulator issued 16 measures for AI in banking and insurance, including a pilot mechanism that would let institutions put large models directly in front of customers — in a controlled environment. The document, from the Shanghai bureau of the National Financial Regulatory Administration, says the city will seek inclusion in a regional pilot run by the national regulator, and that customer-facing model applications would be aligned with the Cyberspace Administration's generative-AI service filing and application registration regime. It also promises a risk-tiered, "prudent and tolerant" supervisory framework, with a fault-tolerance mechanism and differentiated treatment by risk level.
Two words matter here: the text says pilot mechanism and controlled environment, not regulatory sandbox, and it names no institution — this is a lane, not a licence. It is Shanghai's implementation of the national regulator's June guidance on AI in banking and insurance, which set approval and reporting duties for high-risk uses such as credit approval and claims, and it goes one step further by opening a path to customer-facing models. For Chinese banks that is the line that has been missing. Models have been allowed inside the back office for years; letting one answer a customer without a human in the loop has been the unresolved question.
What to watch: whether Transporter-18 lifts on schedule and whether the TPU downlink shows anything beyond survival, and whether Shanghai's pilot gets the national sign-off it is asking for.
Should regulators require a human in the loop before a bank's model answers a customer — or is that just friction? Tell us in the comments.
Sources: Google — Behind Project Suncatcher · Google Research — Exploring a space-based AI infrastructure system design · Reuters — Google plans first test of AI chips in space · Unite.AI — Project Suncatcher prototype heads to orbit · Cailian Press via Sina Finance — 上海金融监管局 16 条措施 · Jiemian News via Sina — full text of the 16 measures · Beijing Business Today via Eastmoney — 推动上海银行业保险业人工智能应用的若干措施 · China Daily — NFRA guidance on AI in banking and insurance
