The wire arrived through a Web3 feed. That is the first thing I want on the record, because it tells you what kind of document this is before you read a single clause.
Forty items. Fourteen prohibited. Twenty-six restricted. Twenty-three of the forty are AI technologies. The arithmetic closes — 14 plus 26 equals 40, and 23 sits inside 40. Internal consistency is not proof, but it is a floor, and floors are worth marking.
Then the number that actually sets the ceiling. Two trillion parameters.
I pulled the public model sheet the way I pull a contract before I size a position. DeepSeek-V3 and R1: 671 billion total parameters, 37 billion active. Kimi K2: one trillion total, 32 billion active. Qwen3-235B-A22B. GLM-4.5 at 355 billion. Nothing on the Chinese public record crosses two trillion. Nothing within a factor of two.
So the first hard fact about Beijing's proposed AI export control is that its headline threshold sits above every model the country has actually shipped. A dam built above the waterline still matters — but it matters for what it tells you about the river, not for what it holds back.
I count the cracks before the dam breaks.
Provenance before policy
Before I read the policy, I read the pipe it came through.
In 2017 I audited three mid-tier ICO contracts instead of reading their whitepapers, and I found an integer overflow in CoinDash's fundraising logic that the team had missed. The lesson stuck harder than the trade did: verify the artifact, not the narrative around the artifact. So when a blockchain outlet publishes a story about AI export controls with zero blockchain content in it, my first move is to grade the source, not the claim.
Grade: medium-low. No link to the original announcement text. No clause-by-clause list of the 23 technologies. No penalty schedule, no licensing channel, no timeline for approval. The publication date on the event is 16 October 2026, attached to what is described as Joint Announcement No. 48 from the Ministry of Commerce and the Ministry of Science and Technology. Public comment runs to 15 November. Temporary measures take effect immediately. Transactions already initiated but not closed are captured by the new regime.
Those last two sentences are the ones with teeth. Everything else in the document is a category. Those two are enforcement.
Strip the framing and the structure is this: 23 AI technologies pulled into a 40-item control list, split between outright prohibition and licensing requirement. Two additions to the registry matter more than the rest — compute clusters, and chip design methods. IP licensing across borders must be registered. And two carve-outs are written explicitly into the text: technologies already released through open source or academic publication are not restricted, and delivery through cloud API or SaaS does not trigger a new export licence.
Read those two carve-outs again. Then read the threshold again.
A door wider than any house in the country
The two-trillion-parameter line is a placeholder. It is not a filter.
Consider the mechanics. Under a mixture-of-experts architecture, total parameters and activated parameters can diverge by a factor of thirty. Kimi K2 is one trillion total and 32 billion active. If the threshold is measured on total parameters, the binding constraint is roughly three times the largest Chinese model in production. If it is measured on activated parameters, the constraint is sixty times. The announcement does not say which. That single undefined variable changes the regulated population by an order of magnitude.
A regulator does not usually set a bar that nothing in the jurisdiction can reach by accident. Two readings are available. Either the bar is forward-looking — a reserved slot for a model class that is being trained right now and has not surfaced — or it is deliberately slack, designed to avoid an immediate shock to an industry the state has spent a decade subsidising. Those two readings imply opposite policy intent, and the source material offers no way to separate them.
That is not the interesting crack. This is: the binding constraint was never the parameter count. It is the second clause. Performance and safety benchmarks, top ten globally. Undefined benchmark. Undefined evaluator. Undefined measurement moment. Is it MMLU-Pro, an arena-style human preference ranking, or a state-built safety suite? Is the assessment done by a third party, by the regulator, or self-declared and audited? Before release, or after training completes?
Any one of those left open makes the compliance question unanswerable at the desk. And an unanswerable compliance question, attached to criminal liability, does not produce caution. It produces over-reporting. Firms file on everything because the cost of filing is bounded and the cost of a wrong judgement is not.
That is the real tax in this document, and it is not levied on capability. It is levied on decision-making. I have watched this exact mechanism run in crypto under MiCA. The regulation promised clarity. What it delivered was a compliance cost curve that only balance sheets above a certain size could climb, and a reserve requirement schedule that quietly deleted the small issuer tier. Clarity that costs more than the ambiguity it replaces is not clarity. It is a filter wearing clarity's clothes.
Two open lanes
Here is the structural read, and it is the whole article in one line.
The control regulates the asset and exempts the pipe.
Weights are a bearer instrument. Copy them once and the copy is as good as the original, and the original cannot be recalled. An API is a custodial wrapper. You never hold the asset. You hold a claim on the issuer's runtime, revocable at the issuer's discretion, logged at the issuer's ledger.
Beijing just wrote a rule that prohibits exporting the bearer instrument and permits exporting the custodial claim. In traditional finance terms: ban the physical gold shipment, exempt the ETF share. Every compliance officer in the world knows which lane the flow will take, and it is not the one with the paperwork.
This is not a loophole. It is the design. The exemption list maps almost perfectly onto how Chinese labs actually earn foreign revenue today — cloud inference and hosted model access, not weight delivery. Whoever drafted this understood the industry's cash flow. That is the tell that separates a real policy instrument from a press release.
Which means the practical consequence of the entire announcement is a routing decision. Sovereign AI buyers in the Gulf and Southeast Asia who wanted to run Chinese weights inside their own perimeter now have three options: accept a licence process of unknown duration, buy the service through a Chinese cloud region instead, or switch stacks. The rule does not stop the capability from leaving. It changes the venue it leaves through, and it hands the venue to Alibaba Cloud, Tencent Cloud and Volcano Engine.
The exemption is an admission, not a concession
Open source being carved out is not leniency. It is a recognition of physics.
Once a checkpoint is on a public hub, it is a public good. It cannot be un-published, un-downloaded, or un-fine-tuned. A rule that restricted already-open weights would be unenforceable on day one, and unenforceable rules are worse than no rules because they teach the market that the regulator's stated boundary is decorative. Code is law until the miners decide otherwise, and here the miners are several hundred thousand developers with local disks.
So the exemption converts a permanent leak into a declared strategy. Follow the incentive: closed-weight export now carries licence friction, open-weight export carries none. Rational labs respond to that gradient. My base case is that Chinese open-weight release cadence accelerates after this regime lands, not because of ideology but because the compliance cost of the alternative went up.
That is a testable prediction, and I intend to test it. Monthly count of new open-weight releases from Chinese labs, tracked against the pre-announcement baseline. If the curve bends upward after November, the routing thesis holds.
It also creates a strange tension the drafters may not have priced. If the goal is to prevent frontier capability from diffusing, open source is the widest aperture in the system. If the goal is to protect commercial IP and steer demand toward domestic cloud, open source is fine and arguably helpful. The text does not say which goal governs. That omission is the single largest source of uncertainty in the whole document, and it is the thing most likely to be closed in a later revision.
The cluster clause is the real bite
Everyone is arguing about two trillion parameters. Almost nobody is arguing about compute clusters, and that is where the damage sits.
Read the clause technically. A compute cluster in this context is not a card. It is a rack-scale or supernode-scale system — the turnkey intelligent compute centre that Chinese vendors have spent two years selling into the Gulf, Southeast Asia and Central Asia. That export line was the commercial answer to silicon sanctions. If the system itself now requires a licence, the answer gets withdrawn by the same government that wrote it.
Two policy goals, one instrument, mutually exclusive. Goal one: keep advanced compute architecture inside the border. Goal two: push domestic accelerators into global markets to break the CUDA lock. A cluster export control achieves the first and forecloses the second. Build the cage, then watch the beast jump in.
The second-order effect is domestic. If cluster-scale systems cannot be exported, that capacity has to clear inside China. More supply chasing the same domestic demand is a price war, and price wars in compute look exactly like price wars in anything else — margin compression, consolidation, and a scramble for any external venue that remains legal.
Chip design methods and EDA registration are the mirror image of the US tooling restrictions, and I want to be precise about the asymmetry. Chinese EDA vendors hold single-digit global share. A registration requirement on their outbound work is a rounding error in revenue terms. The actual friction lands on multinational EDA firms' China-based research centres and joint labs, where the technology-return question becomes a filing question. This clause is a negotiating chip, not a weapon.
Deal flow dies on timing, not price
This is the part I have P&L memory for.
The announcement states that transactions already initiated but not yet completed are captured. In a functioning M&A market, that sentence is a freeze order. Every cross-border AI asset deal in the pipeline stops, because neither side can price an approval of unknown duration with an unknown pass rate.
Liquidity is just borrowed time with a premium. When the premium is an unpriceable option on a regulator's calendar, the market does not reprice. It withdraws. Buyers will demand a condition precedent — unconditional termination if the licence does not issue — and sellers will discover that the condition precedent is worth more than the discount they were prepared to give. Deals will die on the calendar, not on the valuation.
Then the structure adapts, the way it always does. Equity purchases become asset purchases. IP licences become service agreements. Technical assets get split from customer contracts and brand. Licensing gets restructured as support and tuning, which may or may not be inside the perimeter depending on wording nobody has published.
Every one of those restructurings creates a billable hour. Compliance, legal, audit and transaction advisory are the only unambiguous winners in this document, and they are barely mentioned in it.
The same logic runs through DeFi, and I have seen the mechanism up close. In 2020 I ran arbitrage across Uniswap and Sushiswap during the UNI airdrop volatility, moving size within seconds of a pool imbalance. What that taught me is that liquidity does not disappear when a rule bites. It migrates to the venue with the lowest friction, and the migration is fast enough that the original venue never gets to observe the loss in real time. Subsidised TVL teaches you the same lesson in slower motion. Pull the incentive and the number falls, because the number was never users. Here the incentive is a licence exemption, and the venue that gets it is the cloud.
The read-through to the on-chain stack
So what does this mean for the part of the market I actually trade?
The naive read is bearish for decentralised compute. Weights get restricted, therefore open inference networks get squeezed. That read is wrong, and it is wrong for a mechanical reason.
Split the market in two. Closed frontier weights delivered as a service. Open weights delivered anywhere. The control makes the first category more expensive and more centralised — it now must travel through a licensed cloud region, which means the inference endpoint is a permissioned counterparty. The control makes the second category structurally advantaged, because open weights carry no licence and no counterparty.
That is a bifurcation, not a contraction. Decentralised inference over open checkpoints gets a relative boost. Decentralised inference that was pitched as a way to move closed frontier weights gets a relative hit. Most DePIN compute narratives blur those two, and the blur is about to become expensive.
For my own book, the implication is narrow and concrete. I run options strategies on decentralised derivatives venues and I price greeks off realised volatility in the underlying. When a policy announcement changes the distribution of where inference capacity sits — not how much exists, but where it is allowed to be — the volatility that gets repriced first is venue risk, not model risk. Venue risk is the part of the surface most traders ignore because it does not appear in any standard model. Risk is not a number; it is a feeling you ignore.
The Gulf picks a side
The macro consequence is larger than any of this and gets a fraction of the coverage.
A sovereign AI programme needs three things: silicon, weights, cloud. The United States controls the first with conditions. China is now proposing to control the second with licences. A buyer who wants both has to satisfy two incompatible review regimes, and the compliance cost of mixing is climbing toward unaffordable. So they pick.
The 2022 LUNA unwind taught me to look for the failure in the incentive structure rather than in the sentiment. The death spiral was not a mood. It was a design flaw that became visible only under load. The global AI stack has the same shape now: two control regimes compressing from opposite ends, with third-country builders in the middle carrying the cost of both. When the load arrives, they will not hedge. They will choose, and the choice will be sticky for a decade.
That is the actual event here. Not a two-trillion-parameter line. A fork in the infrastructure stack, executed by paperwork.
The contrarian read
The consensus is treating this as a hardening. I think that is a category error, and it comes from pricing the announcement rather than the exemptions.
The headline says export controls on AI. The text says export controls on AI assets, with explicit exemptions for AI services and published research. That is not a wall. That is a toll booth with two lanes marked open. Markets do not stop at toll booths. They pick a lane.
Where the consensus is right, and where I will concede ground, is the chilling effect. Vague standards plus immediate effect plus retroactive capture of pending deals is a genuine tax. But notice who pays it. A large platform has compliance staff and a government affairs desk. A twenty-person model shop has neither, and its entire overseas business may have been one sovereign deployment. That asymmetry means this document is consolidation policy wearing security policy's jacket, and the consolidation will be read as strength by people who do not look at the tail of the distribution.
The second thing the consensus is missing is that the open-source exemption is unstable. It is the widest aperture in the system and it was written into a document whose stated purpose is preventing diffusion. Watch the revision cycle. If the aperture narrows in the final text, the routing thesis breaks and the whole analysis reprices.
What I am watching
Four signals, all cheap to monitor, all falsifiable.
First, the final text after the 15 November comment window. Specifically whether the 23-item list is published in full, and where the compute cluster boundary is drawn — compute scale, interconnect bandwidth, rack count, cooling architecture. That boundary decides whether the domestic compute export strategy survives.
Second, the first licence application and its turnaround. One data point tells you whether this is formal review or substantive review, and that single distinction determines whether cross-border AI deal flow is impaired or dead.
Third, delivery slippage on Chinese compute vendors' Gulf and Southeast Asian order books, tracked quarterly. That is the cleanest available proxy for how much the cluster clause actually binds.
Fourth, monthly open-weight release counts from Chinese labs. If the curve bends up after the regime lands, the routing thesis is confirmed and the market has been arguing about the wrong clause for a year.
The ledger bleeds faster than the logic holds. The logic here is a two-trillion-parameter door that nothing in the country can walk through, attached to a benchmark nobody has defined, enforced by a licence nobody has issued. The ledger is the flow, and the flow is already moving to the two lanes that were left open.
The open question is not whether China can stop model weights from leaving. It cannot. The open question is whether the exemption that lets them leave through the cloud was written as strategy, or written as an admission — and whether anyone in the room understood that those two things produce the same policy until the revision cycle arrives.
Survival is the only alpha that compounds.