OpenAI said on August 7 it has paused any work on Astra, its next major unreleased model, that does not meet stricter safeguards, because early tests suggest it may reach Critical, the level in OpenAI's risk framework for a model that can find and build working attacks on many well-defended real systems without human help. No earlier OpenAI model has been rated Critical, and OpenAI says only that it cannot rule this out, not that Astra crossed the line. New measures include isolated testing, restricted network access, encrypted model weights and live monitoring that can halt risky behavior.
Why it matters
This is the same Astra presented days earlier as a math breakthrough, now the reason its own maker is slowing down. It is the first time a frontier lab has publicly held back its flagship on offensive hacking grounds rather than as a general precaution.
Google said on August 5 that Demis Hassabis, DeepMind's co-founder and leader since 2010, is stepping back from day-to-day leadership. He becomes chair of Google DeepMind and Alphabet's chief scientist, and stays CEO of drug discovery company Isomorphic Labs. Koray Kavukcuoglu, previously DeepMind's chief technology officer, takes over daily leadership, reporting to Sundar Pichai. The same day Jeff Dean said he is leaving Google after 27 years to co-found Discovery Loop, a startup building models that improve themselves with little or no human feedback, with Google as a founding investor.
Why it matters
Two of the most recognizable names in AI research changed jobs on the same day, at the company under the most pressure to keep pace with OpenAI and Anthropic. Investors noticed: Alphabet shares fell around 5 percent on the news.
Meta Superintelligence Labs released Muse Spark 1.2, an update to its coding model, on August 5, alongside Muse Code, a tool that runs in a terminal (the text-based window programmers type commands into) and can plan changes, write code and check its own results across large codebases while working in the background. Muse Code launched in beta for macOS and Linux. Independent evaluator Artificial Analysis scored Muse Spark 1.2 at 54 on its intelligence index, up from 51 for the previous version and 43 for the April version, Meta's third model release in four months.
Why it matters
Coding agents that work unattended are the fastest-moving category in AI products right now, and Meta makes a fourth serious competitor alongside OpenAI, Anthropic and Google. If you are learning to program, these tools are changing what a first job actually involves.
SK Hynix approved about 54 trillion Korean won, roughly $38 billion, for two new memory chip factories. The larger one, in Yongin, breaks ground in July 2027 and will focus on high-bandwidth memory and advanced DRAM, the fast memory that sits next to AI chips. The second, in Cheongju, breaks ground in February 2027. The company pointed to rising memory prices caused by a supply shortage and by demand from AI data center builders and chip firms such as Nvidia.
Why it matters
Memory, not graphics chips, has become the pinch point in AI infrastructure. Amazon blamed a $20 billion increase in its spending plan on memory prices alone last week, and the same shortage feeds through to what you pay for laptops, phones and game consoles.
A New Mexico judge ordered Meta on August 6 to pay $567 million over five years into a fund for young people's mental health, with $420 million of it going to treatment. This is the remedies stage of a case Meta lost in March, when jurors found it knowingly harmed children's mental health and hid what it knew about child exploitation on its platforms, and imposed $375 million in penalties; the total is now roughly $942 million. The court also ordered changes for users under 18: no Like counts, no notifications between 10 p.m. and 7 a.m., and a cap of 90 hours of use a month.
Why it matters
The product changes matter more than the money. A court is now specifying how an app must behave for under-18s, down to notification hours, and other states can copy the template.
In its first earnings report since listing on the Nasdaq in June, SpaceX posted quarterly revenue of $7.8 billion, up 92 percent from a year earlier and about $900 million above what analysts expected. Starlink brought in $4.29 billion of that, up 66 percent, and its subscriber count doubled to 12 million. The company narrowed its net loss to $541 million from roughly $1 billion a year earlier. Shares still fell more than 7 percent after hours, which coverage attributed to the rising cost of building AI infrastructure.
Why it matters
This is the first time anyone outside the company has seen SpaceX's real numbers. It is also a clean example of something that confuses people about markets: a company can beat every expectation and still see its shares fall, because the price already assumed the good news.
Base Power closed a $1 billion round on August 3 at a $13 billion valuation, led by Ribbit, Addition, Valor Equity and JPMorgan's Strategic Investment Group, taking its total funding past $2.5 billion. It installs large batteries at customers' homes for a subscription rather than thousands of dollars up front, and says it installs about 100 a day. The batteries charge when demand is low and feed power back when the grid runs short, which is how the company earns money alongside selling electricity.
Why it matters
This is the other half of the data center story. Three weeks after New York paused new data centers over strain on the grid, a billion dollars went into thousands of small batteries designed to relieve exactly that strain.
The clearest signal this week came from OpenAI holding back its own most capable system, and it is worth sitting with what that means. The model presented days earlier as a mathematical achievement is now being handled as a potential weapon by the people who built it, before anyone outside forced the issue. Read alongside the leadership churn at the company that most needs to keep pace, it suggests the frontier labs are running into limits of judgment rather than engineering.
Everything else this week was about the physical bill: tens of billions for memory factories, a billion for home batteries that exist to relieve a strained grid, and a launch company whose shares fell despite a good quarter because building AI infrastructure costs so much. A pattern runs across the last month: the AI stories keep arriving as questions about control, and the money stories keep arriving as answers about capacity. The money is a bet that the control questions get resolved.
What to watch
OpenAI says Astra still gets a broad release once it meets the safety requirements, so watch whether it arrives with the Critical rating confirmed or cleared.
Test yourself
A few quick questions on this week. Just for you, nothing is recorded.
1. Why did OpenAI slow internal work on Astra?
2. What changed at the top of Google DeepMind on August 5?
3. Why is SK Hynix spending $38 billion on new factories?
4. SpaceX beat analyst expectations, but its shares still fell. What does that illustrate?
