China Considers AI Export Controls

Plus: SK Hynix prepares IPO backed by Situational Awareness, SpaceX rebrands to SpaceXai, Patrick O’Shaughnessy releases second Jeremy Giffon interview

Jul 07, 2026

Happy Tuesday.

We’re live on YouTube and 𝕏.

The current thing in tech and business is China considering limiting the release of domestic AI models to the rest of the world.

Today’s Lineup

Acquisition.com Co-Founder Alex Hormozi at 11:30 AM
CIV Investor Josh Zoffer at 12:30 PM
American Turbines Founder John McElhone at 12:40 PM
Lux Capital Partner Deena Shakir at 12:50 PM
Super.com Co-Founder & CEO Hussein Fazal at 1:00 PM
Norm AI Co-Founder & CEO John Nay at 1:10 PM

Run of Show

China Considers AI Export Controls

Chinese authorities have reportedly held meetings with some of the country’s top tech firms, including Alibaba, ByteDance, and Z.ai, about potentially restricting overseas access to China’s most advanced AI models.

The discussions are still preliminary, and it isn’t clear how Beijing could claw back access to models that have already been released. Once open-weight models are posted publicly, they can be copied, mirrored, downloaded, and reused anywhere. But there are several ways China could restrict future frontier models before they leave the country.

At the lightest end, Beijing could require companies to file future releases with regulators or submit them for national-security review before publishing model weights. A more aggressive version would require government approval before any advanced model could be released overseas. And at the most restrictive end, China could simply ban the public release of frontier open-weight models, limiting them to state-approved customers or Chinese government priorities.

The talks come as the U.S. is also moving toward tighter control over frontier AI. The Trump administration recently restricted foreign access to Anthropic’s most advanced models, Fable 5 and Mythos 5, before later lifting some of those controls after additional safety mitigations. OpenAI has also reportedly faced government pressure around the rollout of its next frontier models.

The issue has become more urgent because Chinese open-weight models are increasingly competitive. Z.ai’s GLM-5.2, in particular, recently impressed Silicon Valley by performing close to leading U.S. models at a much lower cost. Alibaba’s Qwen family has also become one of the most important open model ecosystems in the world, while ByteDance’s Doubao is one of China’s most widely used domestic AI systems.

If China places export controls on future frontier models, it could raise costs for American AI users and companies. Many startups and enterprises rely on open-source or open-weight models for cheaper, lower-stakes tasks, while reserving expensive U.S. frontier models for their most important workloads. If the best Chinese models are no longer freely available abroad, some of that cheap model supply disappears, and more demand gets pushed back toward the expensive frontier labs. — Brandon

Leopold Aschenbrenner piles into SK Hynix’s massive US listing

SK Hynix is set to raise around $28B in a Nasdaq share sale this week, potentially one of the largest-ever New York listings by an Asian company. The South Korean memory chipmaker is already public in Seoul, where its stock is up more than 750% over the past year, but the US listing gives American AI investors a much easier way in.

The story is big on the timeline because Situational Awareness, the hedge fund run by Leopold Aschenbrenner, may participate alongside Baillie Gifford and Coatue. Together, the firms could take as much as $7B of the deal. SK Hynix has become one of the purest AI infrastructure plays because it leads in high-bandwidth memory, has become a key Nvidia supplier, and now has a market cap around $1.1T.

The business is ripping: 2025 revenue rose 47% to $63B, profit more than doubled to $28B, and Q1 revenue nearly tripled year over year. But the stock still trades at only around 7x forward earnings because memory is brutally cyclical. The bull case is that AI demand makes this cycle longer and stronger than the old ones; the bear case is that it’s still memory, and memory cycles eventually end.

Big banks are trying to buy their way around debit-card economics

JPMorgan, Bank of America, Wells Fargo, PNC, and other big banks are reportedly exploring a deal to buy a Fiserv debit-card network, which would let them own more of the payment rails their customers’ debit cards run on. The basic idea is pretty simple: banks hate the debit-fee limits created by the Durbin Amendment, and they want a way to turn debit back into a more valuable business.

Fiserv owns major debit networks including STAR and Accel, which already connect to millions of cardholders and thousands of financial institutions. So this is not some science-project payments startup — it’s a real rail with existing bank and merchant connectivity. The strategic playbook looks a lot like Capital One buying Discover: own the issuer, own the network, move your own volume onto your own rails, and capture more of the economics.

The likely losers are merchants, Visa, and Mastercard. Merchants would probably see higher effective debit costs if banks find a way to route more volume through bank-owned networks, while Visa and Mastercard would lose some leverage over U.S. debit. The bigger story is that payments are getting vertically re-bundled: banks want the account, the card, the network, the wallet, the fraud layer, and eventually the AI-agent checkout surface all under one roof.

Clip Spotlight: Replit Head of AI says we’re entering the “post-prompting era” of AI

“I think we’re going to be seeing a shift very quickly away from apps. They won’t disappear overnight. They still play a role. SaaS will play a role in companies for a long, long time. But it’s also true that some software can be replaced by agents accomplishing tasks end-to-end.”

“LLMs have become much better at long-horizon tasks. And you see them working for hours in a row. The additional layer that we put on top of them — the agent harnesses — are such that you can introduce verification criteria. So that, for example, you come up with a high-level goal. Say you built an app on Replit. Then you want to optimize the conversion rate. And you set 5% as a goal.”

“Guess what? You can have an agent doing that for you. You don’t have to prompt it on the specifics for how to make it happen. You will connect your app to the agent, the agent will decide how to improve conversions, and then you keep measuring until it accomplishes the goal.”

“That’s what you’ve been hearing, more and more, from the industry when they talk about loops. I would call this the ‘post-prompting era.’”

“We don’t put much effort into prompting anymore. Even when you hear Anthropic researchers and engineers explaining how to use Fable 5, the main insight is, stop giving it prescriptions. Stop giving it rules. Just tell the model what you want to accomplish and trust the process.”