Briefing: Beijing's AI Policy and Saaspocalypse is Not Real.
Meta released its commercial model; the benchmark dilemma continues, and my prediction about the SpaceX-Cursor deal has come true.
Beijing’s AI export control
Chinese officials are apparently kicking around a new idea: limiting overseas access to China’s most advanced AI models.
→ Reuters
Trump is a pioneer of many things (I’m practicing positive thinking these days). The latest one, he initiated export controls on AI models, and others soon followed suit.
That’s the latest lesson Beijing learned, and they couldn’t wait to put it into practice.
But the funny part is that the Chinese have gradually found ways to work around the lack of access to the best Nvidia chips. And you’d be so wrong if you think they’re the only ones who need this trick because they don’t have access to those chips.
The fact is, everyone, everyone in the industry would want to learn this workaround, because it’d immediately lower their costs.
Not to mention that when most people hear “Chinese AI,” they immediately go: stealing, distillation, copy-of-a-copy. Sure, that happens. OpenAI has accused DeepSeek of distilling U.S. models.
But if you think American AI companies built their models w/o learning from globally published research and open-weight releases, you’re living in a fairy tale. The fact is that everyone is now distilling everyone else’s model.
But I can’t say this is for the better or worse; to be honest, I’m torn.
On one hand, AI used to be sold as a universal tool. Now it’s starting to look like a permissioned strategic resource: chips, model weights, cloud access, data, deployment rights for the rich and powerful (countries and companies)
On the other hand, export controls in economics typically mean limited economic activity, as they hinder innovation. Isn’t this exactly what Altman and Amodei have been begging for all along? A pause in AI development? This is a great start!
I just wouldn’t have imagined that in the end, it was Trump and Xi who delivered what the AI labs advocated for.
About That ‘Saaspocalypse’
The Information claimed that software stocks have collapsed 30% to 40% this year, because they’ve cited examples of small businesses that are saving a boatload by replacing name-brand enterprise software with custom apps developed with help from AI.
I’m honestly quite disappointed that the Information still humors topics like this.
Having read this article, one serious mistake they made (I could point out 10, but I wrote about this here already):
The journalist treats “we replaced the tiny slice of Salesforce we actually used” as evidence that Salesforce itself is easy to replace.
Meta AI is now ‘Closed’!
Meta Platforms on Thursday unveiled its latest AI model, Muse Spark 1.1.
→ Meta
This is HUGE!
Sorry, that wasn’t shouting, and not excitement either. Combined what I’ve read via the last four years’ Meta earnings calls … Open-weight and Llama were mentioned less as the year went on, from 19 times in early 2023, to 0 times in 2026.
If you get the value of open source, you know how much of a pity this is.
That said, it’s not like the technology itself gave Zuckerberg any choice. To understand why this is the case, stay tuned; the deep dive is coming soon.
The Thing About Benchmarks
Two things happened in the last two days
Claude proves that Fable 5 can achieve 92% of the task in the SWE-bench with just 63% of the cost.
In the meantime, OpenAI audited SWE-Bench Pro, the benchmark everyone used to rank AI coding models, and found it is no longer trustworthy. This one benchmark is basically a standardized test for AI; however, human reviewers found issues in a third of the tasks.
→ OpenAI
You likely have smelled the irony here by now.
So benchmark is no longer trustworthy (as I’ve always said, read my scrutiny at the 2026 Stanford report)
In this case, how do we test AI models? The most effective way is still to just try it out and implement it in your workflow.
A Follow-up of the SpaceX Cursor Deal
SpaceXAI—as the unit is called—and its soon-to-be subsidiary Cursor jointly introduced Grok 4.5, a model “built for coding, agentic tasks and knowledge work.”
→ X
As I predicted in this article, Musk didn’t waste any time releasing an enterprise coding model with Cursor.
I can’t wait to read SpaceX’s next few earnings transcripts to see how this marriage goes and how much value xAI got from plugging themselves into Cursor.
What’s even more interesting is that, now that we know we have xAI, Anthropic, OpenAI, Deepseek, and so on, Meta is also joining the war, fighting for enterprises’ developers’ attention. Curious to see who won in 2 years’ time.
Samsung’s pre-earnings guidance
Normally, a pre-earnings announcement is associated with unexpected bad news.
But in this one, they’ve announced a second-quarter operating profit of about $59 billion, up 1,800% from the year-earlier operating profit, even more impressive than their American opponent Micron's 1,435% growth.
→ Samsung
If you aren’t familiar with either one, think of Samsung and Micron as companies building the plumbing behind AI (of course, Samsung also sells phones, but not in this context). NVIDIA’s GPUs are where the action happens, but Micron’s memory is what moves the data and feeds it to the GPU. And one doesn’t work w/o the other.
And as AI shifts from training models to actually serving answers all day, memory becomes even more important.
As to why both have gone mushroom cloud boom, you only need to know economic 101: when supply scares, prices go up. In this case, the price, the terms, and the customers all try to play nice with the memory companies, so the recent craziness in memory stocks.
It’s good to know more about this topic, even if hardware isn’t your thing. Read this.