- South Korea's sovereign AI plan carries a price tag of $919 billion and a target of 18.4 gigawatts by 2035, split across SK Group, GS Group and Naver.
- The US Justice Department filed a statement of interest backing OpenAI and Microsoft's fair use argument in the New York Times case, days before summary judgment motions land. Useful context if you handle contracts or content rights, alongside our legal ops automation playbook.
- Meta's Muse Spark 1.3 is built to ask before it acts. Meta says it burns fewer tokens; an independent index says each task costs more.
- Anker put a 26-TOPS chip and support for 48TB of storage in a home camera hub, so the footage gets analyzed without leaving the house.
- OpenAI's Astra model uses a technique called recurrent depth, and safety researchers say it leaves far less readable thinking to audit.
Nobody moved a price today. What moved was access: who gets the electricity, who gets the copyrights, who gets to read the model's thinking.
Read Thursday that way and a scattered news cycle turns into one argument. The models are drifting toward commodity. The keys around them are not.
The Front Page: South Korea Buys 18.4 Gigawatts
SemiAnalysis mapped South Korea's sovereign AI program on September 3, and it reads more like an industrial plan than a product launch. The targets: 8.4 gigawatts of data center capacity by 2029, 18.4 gigawatts by 2035, $919 billion of planned investment. Phase one parcels capacity among SK Group, GS Group and Naver, while the government runs a tournament to pick a national foundation model champion.
A gigawatt is roughly the output of a large nuclear reactor, so 18.4 of them is an energy decision wearing a technology label. Chips ship in weeks. Transformers and grid connections take years, and the countries ordering them now are the ones with somewhere to put a model in 2031.
The skeptical read is short. A 2035 number is a plan, not a build, and plans this size slip or quietly shrink when rates or elections move. Anointing a national champion has a mixed record everywhere it has been tried, and picking a winner in 2026 locks in an architecture bet. Worth noting who benefits today: the groups already holding the land and the power contracts.
What it means: Sovereign AI has stopped being a slogan and become a procurement schedule. For anyone building on models, that makes where your compute physically sits a policy question rather than only a latency one. Residency rules, national model preferences and export restrictions all attach to buildings, and the buildings are being poured now.
Releases & Features
Meta shipped Muse Spark 1.3, a model designed to interrupt itself. The September 2 release asks clarifying questions, asks for help when stuck, and confirms before consequential actions. Meta reports about 20% fewer tool calls and 25% fewer tokens than version 1.2, both measured in-house. Independent numbers disagree: Artificial Analysis, whose figures sit in Digital Applied's September release ledger, puts it at roughly 57% more input tokens per task and $0.55 per task against $0.40 for its predecessor, at an unchanged list price. Mark Zuckerberg also said open weights are coming, per LLM Daily.
Anker moved the camera brain into the house. At IFA on September 3, it unveiled the Eufy MindBase, a hub with a 26-TOPS chip, 64GB of flash and support for 48TB of external storage, running its agent locally instead of shipping video to a cloud. TOPS is trillions of operations per second, a rough measure of how much inference a chip can chew through. The "AI brain" branding is the usual fare. The architecture underneath is the interesting bit.
What it means: Both land in the same place from opposite directions. A model that stops to ask a question and a camera that never uploads your hallway are trading raw throughput for control. That trade pays off only if you know what the job is supposed to produce, which is the boring work nobody wants to do. Our workflow directory is a decent place to start writing it down.
In the Lab
OpenAI's forthcoming Astra model uses a technique called recurrent depth, and safety researchers are not thrilled about it. The plain version: most reasoning models think by writing tokens you can read, a visible scratchpad called chain of thought. Recurrent depth loops computation through the network's own layers instead, so more of the reasoning stays as numbers inside the model and never becomes words.
That matters because reading the scratchpad is one of the few practical safety tools anyone has. Watching a model talk itself into a bad plan is how you catch the bad plan. If the thinking never becomes text, the monitor has less to monitor. It lands alongside OpenAI's own statement that Astra is the first model to cross the critical cybersecurity threshold in its Preparedness Framework, with a perfect score on ExploitBench and two zero-day vulnerabilities found and exploited in modified tests, according to the company. The company's own evaluations, on its own benchmark.
What it means: Capability and auditability are pulling apart. A more efficient way to reason is a less legible one, and the industry keeps discovering that its safety story rests on an implementation detail nobody promised to keep. Assume the vendor sees more of the model than you do, and build your checks around outputs you can verify yourself.
The Oversight Desk
On September 2 the US Justice Department filed a statement of interest supporting OpenAI and Microsoft's position that training on copyrighted text can qualify as fair use, in the New York Times lawsuit. A statement of interest is the federal government telling a court what it thinks without being a party to the case. It decides nothing. The judge still rules, and the brief lands days before summary judgment motions, which is no accident of the calendar.
Follow the money. Training data is the largest unpriced input in the industry. If it stays free, model economics work as advertised. If it becomes licensed, every training run gains a line item that flows straight to publishers. Washington has now signaled which outcome it prefers, in the case most likely to set the precedent, the same week its officials were at a G20 ministerial arguing against AI-specific rules.
What it means: Two jurisdictions are moving opposite directions on one question, with Europe's top court weighing training and reproduction while Washington leans toward fair use. If they split, what a model costs starts to depend on where it was trained and where it is served. That belongs in the same spreadsheet as your token costs.
Today's stories all come down to something you rent rather than own. Worth knowing which parts of your setup sit on someone else's terms.
On the Radar
Smaller moves worth a glance, with the sources if you want to go deeper.
- Meta stopped grading employees on how much AI they use. Adoption dashboards and token counts are coming out of performance reviews after staff challenged the practice in court, even as the company pushes a new internal agent called Hatch. Wired.
- Uber is lobbying alongside driver unions. The company that spent a decade fighting organized labor now backs rules keeping humans in the network, including a proposed 85% human-driver quota in New Jersey. Uber's own estimate: one autonomous vehicle displaces roughly four drivers. Financial Times.
- A law firm decided to build rather than buy. Kirkland & Ellis committed $500 million to custom AI systems with Palantir, over $100 million of it in year one, starting with private equity fund formation paperwork. Financial Times.
- A 4.5 billion record TikTok scrape is now public. An independent researcher posted a 289GB dataset, acknowledging the collection broke TikTok's terms and adding a card forbidding profiling or targeting. Those restrictions are policy, not code. Hugging Face.
- A one-developer memory layer for coding agents. Dejavu is MIT licensed, runs entirely on your machine with no account and no backend, and gives tools like Claude Code memory between sessions. Small, free, fixes a real annoyance. LLM Daily.
The Bottom Line
Thursday's through-line was ownership. Korea buying gigawatts, Washington weighing in on who owns training data, OpenAI building reasoning that leaves no transcript, Meta backing off counting its employees' tokens. Each is a fight about control rather than capability, and control gets settled years before it shows up in a product. You cannot influence any of it. You can decide how much of your own work sits on terms someone else writes, and the cheapest time to find out is before they change.
Frequently Asked Questions
-
Sovereign AI is the idea that a country should own the compute, the data centers and at least one serious model inside its own borders instead of renting all three from foreign companies. South Korea's program, as reported by SemiAnalysis, targets 18.4 gigawatts by 2035 and $919 billion of investment across SK Group, GS Group and Naver, plus a government tournament to pick a national foundation model. Most of that money buys electricity, land and supply contracts, because those are the parts you cannot order at short notice. The smaller version of the same question, where your own agents run and who controls that runtime, is in hosted AI agents.
-
Because every clarifying question adds a turn, and every turn re-sends the conversation so far. Meta says Muse Spark 1.3 makes about 20% fewer tool calls and uses 25% fewer tokens than version 1.2, both from its own testing. Artificial Analysis, measuring independently, puts the same model at roughly 57% more input tokens per task and $0.55 per task against $0.40. Both can be true: fewer calls, more context carried through each one. It is a good argument for tracking cost per finished task rather than cost per million tokens, which you can only do once the task is written down. That is the exercise in how to write a workflow spec.
-
AI Daily Newsstand is BYOBot's daily AI news brief, published every night. It covers the day's model releases, new features and capabilities, research, and oversight news, then tells you what each move means for people building with AI, in plain English and without the hype.
