Google runs AI in orbit and Korea's chip exports jump 259%

Over the weekend the AI buildout ran into physics, accounting and the law, all at once. Google got a chip to think in orbit and then had its own energy chief explain why that may never make financial sense. South Korea published export numbers so large they look like a typo. A federal judge told xAI that 35 gas turbines are not a mobile source. And two companies started changing the rules for whoever uses their AI: OpenAI merging its usage meters, Apple renting a Chinese model to get into China. Four of these affect what you pay and what you can build.

A Google TPU wrote a haiku from orbit — and Google says orbital data centers cost up to 100x more

Google's Project Suncatcher prototype, built with Planet and launched October 1 on SpaceX's Transporter-18 mission, ran the open model Gemma 4 on a data-center-class TPU in low Earth orbit. On October 30, Travis Beals, Senior Director of Paradigms of Intelligence at Google, announced the "Hello World" milestone: the team uplinked a prompt asking for a haiku about being in space and the satellite answered as it passed over Svalbard, Norway — "Blue marble below / Data streams in silent grace / Starlight feeds my core." Radiation tests on Trillium TPUs showed they survive a total ionizing dose equal to more than five years in space. Google plans twin interconnected satellites in 2027 to test laser links.
Days later, Amanda Peterson Corio, Google's head of data center energy, told The Information that building data centers in orbit could be "as much as 100 times costlier than building them on the ground," and added: "Let's explore all the terrestrial options before we go extraterrestrial." An expert quoted by NPR put the cost of servicing hardware in orbit at roughly 10x to 100x the ground equivalent. SpaceX is also developing orbital AI compute, with deployment expected as early as 2028. Google expects capital spending of $180 billion to $190 billion in 2026.

Why it mattersThe same company that just proved AI inference works in space told the industry, in the same week, not to count on it. That is the most honest thing anyone has said about the data center crunch: the bottleneck is power and land, and orbit is a research answer, not a cheap one. What it means for Latin America: the 100x reality check makes the region's real asset more valuable, not less — Brazilian, Chilean and Paraguayan renewable power and cool land are the cheap places to put compute, and that is the negotiation the region should be having now rather than waiting for a satellite constellation.

South Korea exported $115.3 billion in one month and chips were almost half of it

The Ministry of Trade, Industry and Resources reported on November 1 that October exports hit a record $115.3 billion, up 93.7% year over year. Semiconductor exports alone were $56.4 billion, up 259% — roughly 49% of everything the country shipped. Imports rose 36% to $72.78 billion, leaving a trade surplus of $42.5 billion.
By destination, exports to China rose 133.6% to $26.98 billion and exports to the United States rose 185.8% to $24.9 billion. Outside chips, petroleum products rose 80% to $6.92 billion, cosmetics 57.5% to $1.44 billion and automobiles 18.1% to $6.55 billion. South Korea is on track for its first trillion-dollar export year, expected to be crossed in November.

Why it mattersThis is the AI boom showing up in a country's trade balance. Memory is the scarce input of this cycle, and Korea is the one selling it, which is why a single sector now carries half of a G20 economy's exports — and why that economy is now exposed to a single demand curve. What it means for Latin America: those prices land on your invoice. Memory shortages are already pushing up the cost of RAM, phones, laptops and servers, and the region imports every one of those in dollars. If you're buying equipment for a business or a team, buy it this quarter rather than next, and quote clients with a hardware clause.

A judge ruled 35 gas turbines are not "mobile" and xAI must face the NAACP

Federal judge Debra M. Brown denied xAI's motion to dismiss the NAACP's Clean Air Act lawsuit over the Colossus data center complex near Memphis, reported October 31. The suit alleges xAI ran at least 35 combustion turbines and other pollution sources without the required preconstruction or operating air permits. Nothing was dismissed; the claims proceed. A similar motion by the Department of Justice had already been denied earlier in the month.
The court rejected all of xAI's arguments: that the turbines are mobile sources exempt from permitting (they are not, under Mississippi's State Implementation Plan), that the plaintiffs lack standing (they have statutory standing under the Clean Air Act), and that citizen suits are unconstitutional. Brown wrote that citizen plaintiffs sue in their own name for harm to themselves, so they are not exercising federal executive power. The NAACP is represented by the Southern Environmental Law Center. Separately, New York recently passed a moratorium on new data centers.

Why it mattersThe energy cost of AI just became a legal cost. "Move fast" works on software; it does not work on an air permit, and a court has now said that a hyperscaler cannot reclassify its power plant as a truck. Expect permits, water and emissions to become the real timeline on every data center announcement. What it means for Latin America: the same buildout is arriving in Querétaro, northern Chile and Brazil, usually announced as jobs and almost never as megawatts and liters. This ruling is the template communities and local governments will copy — and the reason any company signing a data center deal in the region should get the environmental permits in writing before the press release.

OpenAI is merging ChatGPT and Codex into one usage allowance in December

Per OpenAI's updated documentation, starting in December ChatGPT and Codex will draw from a single shared usage allowance for Plus, Pro and Business subscribers. Chats in Instant mode are excluded. The remaining allowance and the reset time will appear in the Usage section of Settings. OpenAI has not disclosed the allowance size per plan, the reset cycle, consumption rates by model, or the exact December start date.
What burns the allowance is tokens, not requests: the prompt, attached files, conversation history, tool outputs and the model's response all count, along with the model and reasoning settings. Long-running agent tasks consume far more than short questions. As of November 1, the Fast speed setting costs 2.5x plan usage versus 1x for Standard. When you run out, the options are to wait for the reset, upgrade, or buy credits — credits do not restore the monthly allowance.

Why it mattersThis is the quiet part of the agent era: agents are metered, and the meter is now shared with the chat window you use all day. One long autonomous task can eat the allowance you were counting on for work, and "Fast" costs two and a half times as much for the same answer. What it means for Latin America: for anyone here paying a subscription in dollars, the margin on AI work is now a usage-management problem. Before December, check which of your workflows actually need an agent and which only need a short answer, move the long jobs to Standard speed, and build the cost of tokens into what you charge clients instead of absorbing it.

Apple Intelligence reaches China on November 4 — powered in part by Alibaba's Qwen

Apple confirmed on its Chinese newsroom, announced October 30, that Apple Intelligence testing in China begins November 4 as part of the next iOS developer beta, with public availability in December alongside the iOS 27.3 release. Certain requests, plus Writing Tools and Visual Intelligence, will route to Alibaba's Qwen; users can ask about photos, PDFs and presentations and generate content, after signing in to an account or using Sign in with Apple.
The China package includes an improved Siri with Type to Siri, Live Translation in Messages, FaceTime and Phone, Writing Tools, Genmoji, Clean Up in Photos, natural-language photo search, priority notifications, notification summaries and Smart Reply. The AI Siri shown at WWDC is not included and stays unavailable in China for now. It is iPhone-only at launch — iPhone 15 Pro and Pro Max, iPhone 16 and later, and iPhone Air — with Siri and device language set to English or Simplified Chinese. Reuters reported Apple reached an agreement with Chinese regulators in mid-July.

Why it mattersApple just conceded that there is no single global AI product. To sell in China it had to rent a Chinese model, ship a shorter feature list, and leave its flagship Siri at the border — the first clear case of a US company splitting its AI stack by jurisdiction because the rules required it. What it means for Latin America: the same logic cuts the other way for us. iOS 27.2, released October 26, added Siri AI in Spanish and Portuguese, which means the region gets the full product rather than a local substitute — but it also shows that whoever writes the rules gets to choose the model. If you build products, assume your AI layer will have to be swappable by market.

The big picture

Put these four days together and the story is not about models getting smarter. It is about the AI industry discovering its own constraints and the price of each one. Physics: a Google chip proved it can think in orbit, and Google's own energy chief priced that at up to 100 times the ground. Supply: a single month of Korean chip exports, $56.4 billion and 259% growth, shows the memory everyone is fighting over and where the money actually goes. Law: a judge said 35 turbines need a permit, which turns electricity from an engineering problem into a litigation timeline. And economics: OpenAI put one meter on chat and agents, while Apple paid the price of entry to China in the only currency the regulator accepted — using a local model. The pattern is a maturing industry learning that compute is land, power, permits and metered consumption. If you're building from the region, the two levers you control are picking workflows that don't burn tokens for nothing and locking in hardware and energy costs before the next quarter does it for you.

Go deeper

Sources & further reading

Topics:AI Chips Space Regulation