Briefing: Washington wants open-weight AI, just not China’s
Kimi K3 lands as Meta tests a cloud move, DeepSeek weighs a listing, and SK Hynix cashes in on the memory squeeze.
Kimi K3, one that puts pressure on Fable 5 and GPT 5.6
You likely have some vague impression of a company called Moonshot and their AI model. Don’t worry about it, because you’ll know of them soon enough. They’ve recently released Kimi K3, an open-weight model that topped multiple coding benchmarks and closed in on Anthropic’s and OpenAI’s flagship models in blind developer testing.
It’s not yet another “gap closing” story; I believe this framing undersells what’s moving underneath it. This drop triggered a few ripples:
Axios reports K3’s release already restarted talk in Washington about banning foreign open-weight models outright.
Chinese models have accounted for 30% of the tokens used by U.S. firms since February. Basically, companies are routing around cost with their own wallets
And another reason companies go for Chinese models is the more relaxed guardrails, raising the question of whether we asked for too much safety alignment that might hinder the US models' growth?
The topic is still brewing; there are some extremely interesting threads to pull. I’m working on an analysis of Kimi and its ripple effect as this week’s deep dive. Stay tuned!
Xi’s attitude towards open-weight AI
An extension of the first story: the same week, Xi Jinping addressed China’s premier AI conference, calling for “open source, openness, collaboration and sharing” and pitching Beijing as an alternative center of AI governance.
If you’ve been following Trump’s admin attitude about open weight, you’d see the irony… While the United States wants broad adoption of U.S. technology and backs open models at home, Chinese vendors’ open-weight releases have been consistently cheaper and more advanced than the US open weight ones.
Makes me wonder if chip export controls play a huge role in pushing Chinese open-weight firms to aggressively pursue a very different AI strategy than the US?
Meta wants a piece of the cloud pie!
Meta Platforms plans to hire Dave Brown, a senior executive at Amazon Web Services, the strongest indication yet that Meta is considering renting out its excess data center capacity to other companies.
→ WSJ
Remember what I said about Meta’s entering the hosting business?
A few clues in the past earnings hinting that if Meta overbuilds, the simplest version is to rent capacity.
But renting out capacity could mean raw GPU access, hosted model access, API access, or some bundle of all three.
My analysis covers each scenario, so bookmark and read it!
Could DeepSeek go public ahead of OpenAI and Anthropic?
DeepSeek is planning to raise funding again! Just this time, the AI firm is also preparing to go public.
I’ve been watching the OpenAI IPO rumors since last winter. Anthropic has also filed confidentially, but still no indication of the timing.
While the US Frontier Labs’ drama drags on, someone finally can’t wait and thought to dip their toe into the water. With a $74 billion valuation and weighing a Shanghai listing, potentially with a filing this year.
If that happens, it will be under scrutiny as the first real public-market test of what a frontier AI lab is worth.
Memory chips continue to harvest their money tree
South Korean memory chipmaker SK Hynix raised $26.5 billion in the largest IPO ever by a foreign company in the U.S., surpassing the record set by Alibaba’s 2014 IPO ($25 billion).
→ BBC
And this is why the SK Hynix listing matters. In my Micron piece, I explained why that memory (not just GPUs) is becoming the bottleneck that the AI industry has to pay for.
SK Hynix is even more exposed to that thesis: it is the HBM (a type of memory that’s key to running AI models) leader, owning half of the market share.
So the institutional investors are not really buying “a memory company,” but the same shortage story: AI needs more memory → HBM is hard to make → capacity takes years to build → and everyone downstream ends up paying for it.
An AI CEO’s side quest
Enough of the serious topics.
Here’s an interesting look at an AI CEO’s worldview and expectations.
Junjie Yan, founder and CEO of Chinese AI developer MiniMax, told employees in an internal memo that he will forgo his salary until the company achieves artificial general intelligence (AGI), while also promising to allocate some of his personal shares to employee incentives and to open source.
The “no salary until AGI” is just the person being dramatic; simply, no founders got rich from salaries. This isn’t the part I find interesting.
The story I want you to see past the headline is how Yan Junjie gave up shares: 4% of MiniMax for employee incentives, plus 1% for open source.
With AI talent being fought over globally, and every AI company keen to establish its market fit, this is MiniMax saying it needs people to think like long-term owners, and stating their contribution to the ecosystem.