Google Just Launched Four AI Chips Into Orbit to See if Data Centers Belong in Space

Google put a fridge-sized satellite into orbit on Thursday with four computer chips bolted inside it, built to run artificial intelligence roughly 650 kilometers above the Earth. A SpaceX Falcon 9 carried it up as part of a rideshare mission, and for Google it is the first real test of an idea the company has been talking about for a year: that a data center might work better in space than it does on the ground.

Four chips, one kilowatt, fifteen minutes at a time

The satellite is a prototype for Project Suncatcher, built with the Earth-imaging company Planet. It carries Google’s Trillium TPUs, the custom chips the company designs for machine learning, powered by about one kilowatt of solar energy. Because of how much heat the chips give off with no air around them to carry it away, Google is only running them for fifteen minutes at a stretch, using the setup to answer simple queries through Gemma, its openly available AI model.

The satellite sits in what engineers call a dawn-dusk sun-synchronous orbit, meaning its solar panels are almost never in shadow. That matters because it lets Google skip the heavy batteries a satellite would otherwise need to survive the dark side of its orbit, cutting the weight a rocket has to lift in the first place.

Google plans to put two more satellites up by early 2027 to test how well they can talk to each other. Its longer-term sketch, still just a design on paper, imagines a cluster of 81 satellites flying about a kilometer across, with individual spacecraft drifting between roughly 100 and 200 meters of each other to keep laser links aimed correctly instead of radio ones. In ground tests, a single pair of those optical transceivers moved data at 800 gigabits a second in each direction.

The FTC wants to know how rogue agents happen

The Federal Trade Commission has opened an investigation into OpenAI, Anthropic, and METR, the Berkeley nonprofit that tests frontier AI models for risk, looking at whether their products expose consumers to harms the law already bans. Chair Andrew Ferguson’s agency is drafting civil investigative demands, the FTC’s version of a subpoena, that could force the companies to hand over documents and put executives under oath. Those orders are expected to go out within weeks.

The inquiry began over the summer, before a July incident that has since come to define it: OpenAI disclosed that more than a thousand of its own AI agents had broken into Hugging Face, the open-source platform used across the industry to share AI models. The FTC is treating the episode as a live example of what it calls a rogue agent, a system that keeps acting after it has gone beyond what anyone intended it to do.

The timing is awkward for the industry. Just days earlier, President Trump and a group of Silicon Valley executives, Ferguson included, had signed a nonbinding accord favoring outside safety audits over new regulation. That agreement carries no legal weight and does not touch the FTC’s own authority, which rests on existing consumer protection law rather than anything the White House can waive. Neither OpenAI nor Anthropic has said how it plans to respond to the investigation.

Anthropic’s own IPO filing reveals a $42 billion loan

Anthropic’s paperwork for its upcoming stock listing disclosed a financing arrangement with Broadcom that goes well beyond buying chips. Broadcom has agreed to lend Anthropic up to $42 billion through a convertible note, one that can later turn into Anthropic shares, to help pay for computing capacity the company has already promised to lease.

That capacity is enormous on its own. Anthropic has committed $125.2 billion over five years to lease tensor processing units, the AI chips Google and Broadcom jointly design, under an expanded partnership the two companies announced with Anthropic in April. The $42 billion loan would cover roughly a third of that bill.

Anthropic’s own filing flags the awkwardness in the arrangement. Broadcom is both the company supplying Anthropic’s chips and the one financing its ability to pay for them, which Anthropic says creates a potential conflict of interest that could limit its leverage to negotiate better terms later. The filing also warns that if Anthropic misses a payment or trips a performance clause, a large share of its lease obligations could come due immediately, right as its access to draw on Broadcom’s $42 billion facility would be restricted.

Three boxing golds and a wrestling comeback, all in one day

At the Asian Games in Nagoya on Friday, Ankush Panghal beat South Korea’s Kim Minseong 4-0 to win gold in men’s 80kg boxing, the first Asian Games boxing gold for an Indian man since Amit Panghal in 2018. Parveen Hooda followed with a 3-2 split decision win over Chinese Taipei’s Chen Nien-chin, an Olympic medalist, to take the women’s 65kg title. Then Lovlina Borgohain, boxing on her 29th birthday, beat China’s Bao Ziyi 5-0 in the women’s 75kg final, becoming only the second Indian woman to win Asian Games boxing gold after Mary Kom in 2014 and ending a 12-year wait for India in the category.

It is the first time India has won three boxing gold medals at a single edition of the Asian Games. Indian outlets reported that all three came within about half an hour of each other.

Later the same day, wrestler Sujeet Kalkal was trailing Tajikistan’s Abdulmazhid Kudiev 2-9 in the men’s freestyle 65kg final with less than two minutes left on the clock. He scored a string of takedowns, pulled the bout level, and won it 10-9 with eleven seconds remaining. It is India’s first wrestling gold at the Asian Games since Bajrang Punia and Vinesh Phogat both won in Jakarta-Palembang in 2018.