Google's first TPU in orbit and 66,935 substations: the science behind October's space rush

On 1 October, Google flew its first advanced chip into orbit. The prototype satellite, built by Planet Labs, launched on a SpaceX rocket from California carrying a Tensor Processing Unit, the company's competitor to Nvidia's GPUs. Once commissioned, according to TechCrunch, the satellite will fire the TPU in 15-minute bursts to avoid straining its power and thermal systems.

"We've done testing on the ground, but you know, there's no test that's completely as good as the real thing," Travis Beals, the Google executive managing Project Suncatcher, told TechCrunch. That project envisions an orbital data centre of 81 satellites flying in close formation and processing in parallel.

The launch was not a solo act. The same SpaceX flight carried more than 100 payloads, including missions from Satlyt and Cowboy Space Company, and another payload on the same rocket is a computing spacecraft built by the Indian startup TakeMe2Space.

What the paper says about the launch maths

Google also released a peer-reviewed version of its white paper on orbital data centres on 1 October, with publication in Joule. The paper argues that SpaceX has achieved a price-reducing learning curve of about 20% a year since the Falcon 1, and that launch prices could approach $200 per kilogram by 2035. TechCrunch's headline figure is blunter: Google thinks Starship has to launch 1,800 times before space data centres get off the ground. The researchers stress the analysis is not an economic feasibility study.

The gap between a demo and a data centre is the real story. At the AI Infra Summit in Santa Clara in September, Sumeet Singh, chief data and AI officer at Viasat, said today's launch costs of $2,500 to $3,000 per kilogram would need to fall to roughly $200 to $500 per kilogram for orbital data centres to make financial sense. He also named thermal management and limited downlink capacity as obstacles, and called for intermediate proof points in the 100 kW to 500 kW class rather than talk of megawatt facilities.

Deepak Sachdeva, CIO of the US Air Force, told the same panel the technology still falls short on operational effectiveness, latency, resilience and security. "From a consumer-to-consumer level, I think it's ready. But from an operational need, which is global in its nature, I think there is still some work to be done," he said, adding that the best approach would be some sort of proof of concept. Data Center Knowledge reported the panel discussion, which took place on 17 September.

A capsule, a lawsuit and a blood test

Money is moving into the same thesis. Satlyt, co-founded by former Google and SpaceX product manager Rama Afullo, raised an $8 million seed round, Afullo told TechCrunch exclusively. The company does not build spacecraft; it builds software to run AI models across other companies' satellites.

Afullo's pitch leans on a concrete result. Earlier this year Satlyt deployed Google DeepMind's Gemma model on a spacecraft operated by Momentus, and the model cut the size of a transmission about onboard software errors by more than 60%, which Afullo says can save hundreds of thousands of dollars per satellite each year. This week's launch puts Satlyt software onboard a TakeMe2Space spacecraft with three customers: NASA, which is paying to test cloud-computing protocols in space, the space-surveillance startup Stellerian, and TakeMe2Space itself.

Pharmaceutical manufacturing in orbit drew a larger cheque. Varda raised a $251 million Series D to advance drug manufacturing in space, Endpoints News reported on 30 September. The company also published a photograph of staff removing a heat shield cover from a capsule in preparation for launch on the same date.

Biology produced the week's most quotable result. Research from the University of Copenhagen, conducted with NASA, analysed blood samples from 52 astronauts who spent months aboard the International Space Station and found signs that gut bacteria begin fermenting protein to a greater extent than usual within weeks of arrival, a change that continues until the astronauts are back on Earth. ScienceDaily reported the findings on 28 September.

"The lack of gravity in space probably causes food to move more slowly through the intestine, and this fits with the fact that we are seeing signs of increased protein fermentation. This may also help explain constipation in astronauts," said co-author Henrik Roager, an associate professor at the university's Department of Nutrition, Exercise and Sports. Senior author Lars Ove Dragsted noted that protein fermentation products are often associated with kidney damage and potential effects on mood or focus.

The same week, STAT's Readout LOUD podcast covered a lawsuit filed by the gene-editing company Beam Therapeutics against a Chinese competitor, alongside a clinical update from UniQure on its Huntington's disease gene therapy and a shift in how Novo Nordisk and Eli Lilly talk about their GLP-1 products. The episode aired on 1 October and also featured STAT's trust-in-science reporter Nicholas Florko on declining public confidence in science.

Forecasting the grid and the money behind the hardware

Not all of the week's science was about leaving Earth. In a Microsoft Research blog post published on 30 September, a summer intern described a machine learning pipeline that forecasts space-weather risk across 66,935 substations in the continental United States. The system combines solar-wind observations, forecasts of the Auroral Electrojet and Disturbance Storm Time indices, physics-informed constraints, local geological conductivity and grid infrastructure data. Microsoft says it detected nearly 80% of major space-weather events during the evaluation period and can warn grid operators 30 to 60 minutes before a specific risk appears.

On the ground, Europe's space sector is trying to reduce its dependence on imported chips. At the Pretzl Connect 2026 press event in Budapest, Kate Underhill, future space transportation propulsion architect at the European Space Agency, told EE Times that space systems typically operate "at least 10 years behind consumer electronics" because of the need for radiation tolerance. She described an ESA attempt to order just 20 laser diodes from a German supplier that required a minimum order of 10,000 units, and said that any US component on a satellite pulls the buyer into US export regulations under ITAR.

Hardware startups kept raising. Blackswan Space, a Lithuanian company founded in 2019, closed a €2.5 million round led by Iron Wolf Capital, bringing total funding to €3.2 million, Tech.eu reported on 30 September. Its rendezvous and proximity operations payload is due to fly in 2027, ahead of support for Europe's first in-space refuelling mission a year later. Agile Space Industries, a Durango, Colorado propulsion supplier, said on 28 September that Jason Wright will join as President in early October; the company expects to reach roughly five times its 2022 revenue by the end of 2026 and employs more than 165 people.

Quantum Space signed a spacecraft processing work order with All Points Logistics for its Prime mission, targeting launch in the second half of 2027 from Vandenberg Space Force Base. "Securing our launch-site processing capability is another important element of mitigating execution risk as we move Prime toward flight," said Kerry Wisnosky, President of Quantum Space, in the 30 September announcement. Galileo Space, meanwhile, is building satellites designed to fuse sensor data with AI in orbit; its first demonstration satellite is scheduled for the fourth quarter of 2027, and the company says it is in the final stages of securing its first government contract. It was selected for the a16z speedrun programme as one of 80 companies chosen from 30,000 applicants.

One older thread is worth keeping in view. Ethereal Space was awarded Task Order 1 under NOAA's ProTech 2.0 Space-Based Environmental Monitoring programme, a 24-month contract worth approximately $27.4 million with initial data deliveries beginning 1 December 2026. That is the operational end of the same trend: space-based sensing sold as a service, with delivery dates and contract values attached.