Two trillion parameters is a line. Cross it, and you have to apply for an export license. Unless you make it public, so anyone can download it.
On October 16, the Ministry of Commerce and the Ministry of Science and Technology released a draft for public comment. Forty technologies are proposed for addition to the catalog of technologies prohibited or restricted from export: fourteen prohibited, twenty-six restricted. Twenty-three of them are AI-related.
The notice gives three ways a large model can cross the line: computing power reaching 10^26 floating-point operations; parameters reaching two trillion; or ranking among the global top ten in public evaluations.
Most people reading this would come to the same conclusion: the more powerful it is, the less likely it is to be allowed out.
That’s backwards.
There’s another sentence tucked into the same document: technologies that have already entered the public domain through open-sourcing or published papers are not subject to these controls.
Add that sentence, and the first three criteria take on a different meaning. What’s being measured is ownership. Capability comes second.
Here’s a real-world example. Guess which one gets held up?
DeepSeek’s V4-Pro has 1.6 trillion parameters—40 billion short of two trillion, nowhere near the line. Even if it did reach the threshold, its weights are posted in full on Hugging Face under an MIT license. Anyone can download and run it, so those three criteria simply don’t apply. Baidu’s ERNIE 5.0, on the other hand, exceeds two trillion parameters and is closed-source. That exemption doesn’t really work in its favor.
A bigger model can go overseas freely. A smaller one, by contrast, may have to get approval one item at a time. Strange, isn’t it?
Why?
Because of whether it has an owner who can be identified.
Bring Manus into the picture, and the logic clicks into place.
In 2025, the company moved its headquarters to Singapore, shut down its services in China, and changed its corporate shell so thoroughly that it looked squeaky clean. At the end of the year, Meta announced its acquisition for $2 billion, the third-largest deal in Meta’s history. Yet regulators didn’t step in until two months later. On April 27, the Office of Foreign Investment Security Review under the National Development and Reform Commission issued an announcement prohibiting the investment and requiring it to be unwound and restored to its original state, leaving no room for maneuver. In August, Manus announced it was returning to independent operations.
The basis is laid out in the review rules: substance over form. The technology was developed in China, the algorithms were developed in China, and the people writing the code were in China. Only the place of registration had changed on paper; nothing underneath had changed. So the review went past the business license and straight to the core. It didn’t even ask which flag the company was registered under.
Changing the shell doesn’t work. It never looks at the shell. Really.
Here’s the point I’m trying to make: this line measures ownership; technological strength comes second. The dividing line is whether there’s an entity behind it that can sign, be held accountable, and hand over the underlying code. With an entity behind it, technology becomes an asset that can be transferred. And transferable things are what need licenses.
The same thing plays out on-chain every day. Exactly the same.
Bitcoin has no issuer, no shareholders, and no intellectual property rights that could be presented to the authorities. For more than a decade, regulators around the world have gone after its on-ramps: exchanges, OTC trading, fiat gateways, custodial wallets. None of them can escape. But the asset itself has never once been shut down. There’s no one in the dock. Really, no one.
Stablecoins are different. They have issuers, so whenever something goes wrong, the first thing regulators check is the reserves; the second is whether they can freeze those few addresses on the list. Permissioned chains are even more straightforward: their operators can shut them down and change the rules, so they have to get licensed and verify users’ identities. Public blockchains can’t be shut down, so regulators can only stand guard at the entrance. There’s nothing anyone can do to the chain itself.
To figure out how easy something is to regulate, first ask who will be called in for questioning when things go wrong.
The same goes for people holding crypto.
Can the rules governing the coin in your hand be changed in the middle of the night? Can the issuer quietly mint more? Can someone hit pause on the contract? None of that is captured by the adjectives in a white paper. It’s a list of names. The list says who’ll be sitting in the meeting room when something goes wrong. If the list is empty, the rules can’t be changed. If there’s a name on it, that person can make decisions for you.
The public comment period closes on November 15. There’s only a month left. We don’t yet know which items will make the final cut or when the rules will take effect. Nobody can guarantee whether the openings created in this round will be narrowed again. At least, I haven’t seen anyone make that promise.
If you want to see more breakdowns like this, follow me.
$BTC $USDC $TAO
#开源 #出口管制 #稳定币 #AI
On October 16, the Ministry of Commerce and the Ministry of Science and Technology released a draft for public comment. Forty technologies are proposed for addition to the catalog of technologies prohibited or restricted from export: fourteen prohibited, twenty-six restricted. Twenty-three of them are AI-related.
The notice gives three ways a large model can cross the line: computing power reaching 10^26 floating-point operations; parameters reaching two trillion; or ranking among the global top ten in public evaluations.
Most people reading this would come to the same conclusion: the more powerful it is, the less likely it is to be allowed out.
That’s backwards.
There’s another sentence tucked into the same document: technologies that have already entered the public domain through open-sourcing or published papers are not subject to these controls.
Add that sentence, and the first three criteria take on a different meaning. What’s being measured is ownership. Capability comes second.
Here’s a real-world example. Guess which one gets held up?
DeepSeek’s V4-Pro has 1.6 trillion parameters—40 billion short of two trillion, nowhere near the line. Even if it did reach the threshold, its weights are posted in full on Hugging Face under an MIT license. Anyone can download and run it, so those three criteria simply don’t apply. Baidu’s ERNIE 5.0, on the other hand, exceeds two trillion parameters and is closed-source. That exemption doesn’t really work in its favor.
A bigger model can go overseas freely. A smaller one, by contrast, may have to get approval one item at a time. Strange, isn’t it?
Why?
Because of whether it has an owner who can be identified.
Bring Manus into the picture, and the logic clicks into place.
In 2025, the company moved its headquarters to Singapore, shut down its services in China, and changed its corporate shell so thoroughly that it looked squeaky clean. At the end of the year, Meta announced its acquisition for $2 billion, the third-largest deal in Meta’s history. Yet regulators didn’t step in until two months later. On April 27, the Office of Foreign Investment Security Review under the National Development and Reform Commission issued an announcement prohibiting the investment and requiring it to be unwound and restored to its original state, leaving no room for maneuver. In August, Manus announced it was returning to independent operations.
The basis is laid out in the review rules: substance over form. The technology was developed in China, the algorithms were developed in China, and the people writing the code were in China. Only the place of registration had changed on paper; nothing underneath had changed. So the review went past the business license and straight to the core. It didn’t even ask which flag the company was registered under.
Changing the shell doesn’t work. It never looks at the shell. Really.
Here’s the point I’m trying to make: this line measures ownership; technological strength comes second. The dividing line is whether there’s an entity behind it that can sign, be held accountable, and hand over the underlying code. With an entity behind it, technology becomes an asset that can be transferred. And transferable things are what need licenses.
The same thing plays out on-chain every day. Exactly the same.
Bitcoin has no issuer, no shareholders, and no intellectual property rights that could be presented to the authorities. For more than a decade, regulators around the world have gone after its on-ramps: exchanges, OTC trading, fiat gateways, custodial wallets. None of them can escape. But the asset itself has never once been shut down. There’s no one in the dock. Really, no one.
Stablecoins are different. They have issuers, so whenever something goes wrong, the first thing regulators check is the reserves; the second is whether they can freeze those few addresses on the list. Permissioned chains are even more straightforward: their operators can shut them down and change the rules, so they have to get licensed and verify users’ identities. Public blockchains can’t be shut down, so regulators can only stand guard at the entrance. There’s nothing anyone can do to the chain itself.
To figure out how easy something is to regulate, first ask who will be called in for questioning when things go wrong.
The same goes for people holding crypto.
Can the rules governing the coin in your hand be changed in the middle of the night? Can the issuer quietly mint more? Can someone hit pause on the contract? None of that is captured by the adjectives in a white paper. It’s a list of names. The list says who’ll be sitting in the meeting room when something goes wrong. If the list is empty, the rules can’t be changed. If there’s a name on it, that person can make decisions for you.
The public comment period closes on November 15. There’s only a month left. We don’t yet know which items will make the final cut or when the rules will take effect. Nobody can guarantee whether the openings created in this round will be narrowed again. At least, I haven’t seen anyone make that promise.
If you want to see more breakdowns like this, follow me.
$BTC $USDC $TAO
#开源 #出口管制 #稳定币 #AI