The number nobody is pricing: 23.
On October 16, 2026, China's Ministry of Commerce and Ministry of Science and Technology published Draft Announcement No. 48 โ a proposed export control regime covering 40 technologies, of which 14 are outright prohibited and 26 restricted. Twenty-three of those 40 are artificial intelligence. The headline clause that got scraped, retweeted, and grimly quoted in Telegram channels is the threshold: any model with two trillion parameters, or any model ranked in the global top ten on performance and safety benchmarks, requires an export license before its weights can cross a border.
Read that again and notice what it does not say. It does not say "any model." It says "weights."
The nuance is where the entire strategic design lives, and where almost every commentary I have read in the last forty-eight hours has walked straight past it. This is not a firewall around Chinese AI. It is a router. It redirects the global flow of Chinese machine intelligence into precisely one channel โ the channel that Chinese cloud providers own โ and it does so with language that is calm, technical, and, at first pass, almost reassuring.
Start with the arithmetic, because the arithmetic is where the draft reveals its own modesty. Across all publicly disclosed Chinese models to date, none has reached two trillion total parameters. DeepSeek-V3 and R1 sit at 671 billion total with roughly 37 billion active. Kimi K2 crossed into the trillion-parameter class โ a textbook mixture-of-experts design with around 32 billion active parameters. Qwen3-235B-A22B and GLM-4.5 at 355 billion round out the frontier. Every single one of them is an order of magnitude or more below the prohibition line. If the threshold is computed on total parameters, then today the prohibition clause captures essentially nothing. The draft has drawn a red line and placed it above the current ceiling.
That fact invites two readings and the industry has not yet decided which to adopt. Either the regulator is reserving a threshold for models already in training โ because governments rarely legislate against things that do not exist โ or it is deliberately setting a bar high enough that nothing moves today, giving the ecosystem a buffer period before the rules bite. The two readings imply opposite policy intentions and the draft, at least in the summarized form circulating, provides no distinguishing signal. For what it is worth, my experience in 2017 auditing 45,000 lines of Solidity for Paragon Coin taught me that the line a designer draws says more about the designer than about the system. When a team writes a transfer function with a check for a condition that cannot currently occur, it is usually because they know something about the next release that they are not telling the auditor. I suspect the same here. The two-trillion line is a placeholder for a model that exists in a lab and not yet on a card.
The second half of the threshold โ "ranked in the global top ten on performance and safety benchmarks" โ is the more interesting engineering artifact, because it is not engineering at all. It is a legal instrument disguised as a metric. Which benchmarks? MMLU-Pro is a proven deflation target. LMArena is a human-preference ranking with known contamination problems. A bespoke safety evaluation would require the regulator to publish the rubric, and the draft, from what has circulated, does not. Who evaluates? A third party, the regulator itself, or a self-certification regime? At what point in the lifecycle โ post-training, pre-release, post-deployment?
Any one of those unanswered questions makes the compliance decision indeterminate. And indeterminate compliance combined with criminal liability produces a predictable economic response: firms over-report. They file licenses for models that clearly do not qualify, because the cost of filing is administrative and the cost of not filing is potentially prosecutorial. You do not need a draconian enforcement apparatus to produce a chilling effect. You need ambiguity and a penalty. The draft supplies both.
Now the load-bearing exemption, the one nobody is quoting. The draft explicitly states that the new licensing requirements do not apply to cloud API services. DeepSeek and Alibaba, per the text, can continue serving global customers through cloud interfaces without triggering a new export permit. Separately, technologies already published under open-source licenses or disclosed in academic papers are excluded โ which is less a concession than an acknowledgment of physics. You cannot recall a weight file. Once it is on HuggingFace it is in the noosphere permanently, and no ministry writes regulations that pretend otherwise.
Put the two exemptions side by side and the architecture becomes legible. Sell the model as a product โ as a downloadable artifact, a private deployment, a sovereign AI file transfer โ and you enter a licensing regime. Sell the model as a service โ as inference tokens billed by the million โ and you remain untouched. The regulator has not blocked China's AI export. It has selected which export format survives.
This is where the strategic reading gets uncomfortable for the people who wrote the draft, and it is the part I have not seen discussed anywhere else.
For the last two years, Huawei, Inspur, and H3C have been running a deliberate commercial offensive across the Gulf, Southeast Asia, and Central Asia, selling turnkey intelligent computing clusters โ not single accelerators, but rack-scale supernode architectures, the CloudMatrix-class designs that compete with Nvidia on system integration rather than on silicon. That is the export channel for Chinese compute. And "computing clusters" appear, per the circulated summary, on the controlled list of 23 AI technologies, alongside "chip design methods."
The draft may have just cut off China's own compute export strategy with Chinese scissors.
The mirror logic is apparent. Washington restricts advanced silicon and EDA tooling flowing into China; Beijing responds by restricting Chinese design methodologies and cluster architectures flowing out. Symmetric in principle. But the magnitudes are not symmetric at all. Chinese EDA firms โ Huada Empyrean, Primarius โ hold single-digit global market share. Registering their intellectual property for export review is a diplomatic gesture with limited commercial bite. Cluster architectures are different. That is where the actual revenue and the actual ecosystem ambitions sit. Every sovereign AI project in Riyadh or Abu Dhabi that was going to be built on Ascend silicon and a Chinese cluster design now faces a permit process whose duration is undefined and whose approval rate is unknown.
And there is the deeper structural problem. CUDA's moat is not a chip. It is a decade and a half of developer familiarity, a documentation corpus, a toolchain, a habit. You break that moat by getting your alternative into as many hands as possible, as fast as possible, in as many jurisdictions as possible. Ecosystem requires scale. Scale requires export. Efficiency is the enemy of resilience, and here resilience โ the strategic goal of compute independence โ is being sacrificed to an efficiency argument about controlling outbound capability. The two objectives are not reconcilable inside a single export-control framework. One of them will lose. I would not have expected the draft to choose this one.
The regional consequence is the thing that should genuinely worry anyone modeling global AI infrastructure. The United States has spent four years throttling the inbound flow of advanced accelerators to China. China now proposes throttling the outbound flow of frontier model weights. These are not competing policies. They are complementary halves of a single bifurcation, and the seam runs through third countries.
Consider the position of a sovereign AI program in Malaysia, or Kazakhstan, or Indonesia. To build national capability you need three things stitched together: compute, model, and cloud. Take American silicon and you accept American export conditions โ end-use monitoring, end-user declarations, the whole apparatus. Take Chinese weights for private deployment and you now require a Chinese export license. The mixed-stack approach, which was the pragmatic path for most mid-tier nations eighteen months ago, is becoming the expensive path. You are not choosing a vendor. You are choosing a bloc, because the compliance overhead of hedging has exceeded the cost of commitment.
Who actually bears this cost? Not OpenAI, not Anthropic, not DeepSeek, not Alibaba. The people who pay for bifurcation are always the ones without the legal department to route around it.
I want to be precise about what is solid here and what is inference, because the gap matters and because a great deal of what is being written about this draft is inference dressed as fact.
Solid: the numbers reconcile. Forty total, fourteen prohibited, twenty-six restricted, twenty-three AI-specific. In a policy document scraped from a secondhand aggregator โ and I should note the source, a blockchain and Web3 news outlet that somehow published 2,000 words on AI export controls without mentioning blockchain once โ internal arithmetic consistency is the only reliability signal available. It holds. Also solid: the API exemption is explicitly stated, and it maps precisely onto how Chinese AI companies actually earn foreign currency, which is overwhelmingly through cloud inference billing and model hosting rather than weight delivery. A regulator who did not understand the industry would not have drawn the line there. Somebody in that ministry has read a revenue breakdown.
Inference, and I want to flag it as such: everything about magnitude. How much of Chinese AI export revenue runs through private deployment versus API? No public figure. How large are the Gulf sovereign AI contracts at risk? Estimates in the hundreds of millions to low billions, but nothing auditable. What is the licensing timeline? Unstated. Whether the permit regime is a formality or a substantive review is the single variable that determines whether this policy is a speed bump or a wall โ and the draft does not say.
There is a procedural detail in the circulated summary that deserves far more attention than it has received: the measures take effect immediately, the public comment period runs to November 15, and transactions already initiated but not completed are subject to the new rules. Read those three clauses together. The regulator has locked in review authority before the rules are finalized, and has explicitly closed the window in which firms might have raced to close pending deals during the comment period. That is not a drafting oversight. That is a designed enforcement posture, and the summary treated it as a minor administrative note.
What it means practically: the cross-border AI asset transactions sitting in term sheets right now are frozen. Buyers will demand material adverse change clauses keyed to license approval. Sellers will accept discounts to close before the rules tighten further. Every deal involving Chinese AI intellectual property acquires a new due diligence line item โ is this asset an export-controlled technology, and if so, has the license been granted โ and that line item is now billable to somebody. The compliance and legal advisory sector gets a structural tailwind from a policy document it had no hand in writing. I have watched this pattern before; in 2022, after Terra, the demand for independent reserve attestation and on-chain forensics exploded within a quarter, and the firms positioned for it captured more value than most of the protocols did.
And there is one asymmetry in this draft that I find genuinely elegant, whether or not it was intended.
The exemption for open-source weights means that the one path to global distribution that survives the new regime is the one that distributes capabilities most freely. A Chinese lab that publishes weights under a permissive license faces no licensing requirement, no review, no timeline. A Chinese lab that tries to sell those same weights to a sovereign customer faces all three. Whatever the stated intent โ protecting intellectual property is the charitable reading โ the operative effect is to make open publication the cheapest route to global reach. The math was sound; the trust was the variable. Beijing is being pushed, by its own control architecture, toward the open-weight strategy that has already made Qwen and DeepSeek and Kimi the default substrate for fine-tuning across the developing world.
If that reading holds, the next twelve months should show an acceleration in Chinese open-weight releases, not a deceleration. That is a falsifiable prediction, which is more than most of the commentary on this draft has offered. Track HuggingFace release cadence from Chinese labs on a monthly basis. If publication frequency rises after November 15, the thesis holds. If it falls, I am wrong and the chilling effect dominates.
The version of this story that the market is currently pricing is a story about restriction, and restriction is legible โ it makes for clean headlines and clean risk-off positioning. But the restriction here is narrow. Two trillion parameters captures nothing currently in production. Top-ten benchmarks is a discretionary standard, not an automated cutoff. Open-source weights are exempt. Cloud APIs are exempt. The binding constraints are on a specific set of delivery formats โ weight transfer, private deployment, IP licensing, cluster architecture export โ and those formats, while the highest-margin, are not the highest-volume.
The version of this story the market is not pricing is the one about the cluster architecture, and that is where I would focus attention. If the final catalog, due after November 15, defines "computing clusters" in a way that catches rack-scale systems, then China has voluntarily surrendered its most credible channel for competing with CUDA, at precisely the moment when compute ecosystem expansion was the strategic priority. That is either a deliberate trade โ accepting ecosystem loss in exchange for capability control โ or it is a drafting error with nine-figure consequences for the firms holding Gulf orders. The difference between those two explanations will be visible in the final text.
Watch the first license application. Not the rules, the application. Whether the first filed request gets substantive technical review or a routine signature tells you whether this regime is real. My expectation is that we see that first application within ninety days of the final catalog, and that its processing timeline function as the true pricing mechanism for every Chinese AI asset on the planet.
Until then, what we have is a draft that sits sideways relative to its own stated purpose: tall enough to signal, wide enough to accommodate, and pointed โ perhaps accidentally โ straight at the one export channel China cannot afford to lose. Liquidity is not a floor; it is a horizon. Capital, models, and talent all move toward the path of least resistance, and this draft has drawn a very clear map of where the resistance is low.
Correlation is the smoke. The fire is in the cluster clause.