Two conversations used the word “rules” inside the same seven days. They were not proposals about the same object — and conflating them is the most available error in reading this week.
Timeline — September 2026
15 Sep 2026
Beijing Xiangshan Forum opens. Delegates from nearly one hundred countries and regions attend.
The Beijing Xiangshan Forum ran from 15 to 17 September 2026, drawing delegates from nearly one hundred countries and regions. At the forum, China‘s Defence Minister Dong Jun called for intensified security dialogue and clearer rule-making across three emerging domains: artificial intelligence, outer space and the deep sea. He said countries should “strengthen risk assessment, experience-sharing and rule-making, so that emerging technologies can better benefit all of humanity.” He also said countries should “actively engage in security dialogue in areas including artificial intelligence, outer space and the deep sea.”
Published accounts of who led the United States delegation differ, so no claim about that is made here. On the diplomatic backdrop: reporting indicates US and Chinese negotiators are set to meet in New York this weekend on a set of issues that includes AI, and that Xi is set to meet Trump on 24 September, which China has not confirmed. Both are reported rather than confirmed arrangements, and nothing here predicts either outcome.
In the same week, a distinct conversation was running in Western laboratories: whether commercial frontier model releases should be paced by agreement among the developers themselves. The two conversations both used the word “rules.” They were not proposals about the same thing.
The key insight: A release-rate agreement governs what a company ships. A state-to-state security dialogue on military AI governs what a state fields. Neither instrument reaches the other’s object. The Xiangshan proposal is not an answer to the commercial pacing argument, and it does not fill the gap that argument has where China is concerned. It is a different instrument aimed at a different layer.
Grouping artificial intelligence with outer space and the deep sea treats it as a domain rather than as a product. The other two domains acquired their founding instruments decades after the capability arrived.
The Structural Read
The most common error available in reading this week is to treat the two “rules” conversations as addressing the same problem from different angles. They do not. A defence-ministry channel does not constrain a commercial training run. A developers’ agreement does not constrain a weapons programme. These are not gaps waiting to be filled by the other instrument — they are different instruments, operating at different layers, with different counterparties.
The second structural observation concerns venue. This proposal arrived from a defence ministry at a security forum — not from an industry ministry, not from a standards body. Where a government chooses to raise a question is evidence about which institution it believes owns that question. A speech establishes what was said and where. It does not establish policy intent, internal process, sincerity or strategy, and no claim about any of those is made here.
Permission Layer — BE Framework
Two Layers, Two Permission Problems
Every commercial pacing mechanism discussed this week required a permission held by someone other than the proposing party — an antitrust waiver from a government, auditing staff a profession does not yet have, participation from a jurisdiction outside the arrangement. A state occupies a different position on that single dimension: it can bind conduct inside its own jurisdiction without requiring anyone’s waiver. That is a statement about instruments, not outcomes. It does not mean state channels work better, move faster, or produce agreement. International dialogue has its own permission problem — a more familiar one: it requires the other state to agree.
The third structural observation concerns the bundle itself. Artificial intelligence, outer space and the deep sea share one property: in each domain, capability has repeatedly arrived ahead of doctrine. Two of the three acquired foundational instruments after that happened — the Outer Space Treaty was concluded in 1967, and the UN Convention on the Law of the Sea was adopted in 1982. The third has no equivalent foundational treaty. That is not the same as saying AI is ungoverned: national rules, export controls and international declarations all apply to it. What is absent is an instrument of the same character as the other two.
Grouping all three treats AI as a domain question — something to be governed the way a commons is governed — rather than a product question, which is the framing the commercial pacing debate operates in. Whether that reframing is right, whether a treaty would help, and whether the existing two founding instruments have worked are separate questions this piece does not attempt.
Three Implications
IMPLICATION 1 — Institutional Ownership Is Being Contested
The choice to raise AI rule-making at a defence-ministry security forum rather than at an industry or standards venue is itself a positioning act. It signals — without proving — that at least one major government frames advanced AI primarily as a security-domain question. Whether that framing holds, spreads, or shapes any eventual instrument is not predictable from a single speech. But the venue is evidence, and evidence is worth recording.
IMPLICATION 2 — The Commercial Pacing Debate Has an Unresolved Gap Xiangshan Does Not Close
Proponents of commercial pacing agreements have acknowledged that any arrangement confined to Western laboratories leaves a jurisdiction gap. The Xiangshan proposal operates at state-to-state level and concerns military AI. It addresses a different object. Analysts and observers who read this week’s two conversations as mutually reinforcing should state explicitly which layer they believe connects them — because the connection is not structural and is not established here.
IMPLICATION 3 — Domain Framing Versus Product Framing Will Shape What Governance Looks Like
The Outer Space Treaty and UNCLOS governed commons and domains. Commercial AI governance proposals — release pacing, audit requirements, compute thresholds — govern products and companies. These are different architectures of rule. Whether AI ultimately acquires governance instruments that resemble the 1967 and 1982 models, or instruments that more closely resemble product-safety or financial-stability regimes, is one of the genuinely open structural questions in this space. Xiangshan’s bundling of the three domains pushes toward the former. The commercial pacing debate pushes toward the latter. Neither has resolved the question.
Business Engineer Framework
The Permission Layer
Every governance mechanism — commercial or state-level — runs into a permission problem. For industry agreements, it is antitrust clearance, audit capacity, and cross-jurisdictional buy-in. For state dialogue, it is the other state’s agreement. The Map of AI traces where these permission dependencies sit across the nine layers of the stack — and which actors hold the keys at each one. Understanding the layer determines what instrument can actually reach it.
Two conversations used the word “rules” in the same week. One governs what a company ships; the other governs what a state fields. A defence ministry channel does not reach a commercial training run, and a developers’ agreement does not reach a weapons programme — which means the Xiangshan proposal is not a substitute for the commercial pacing mechanism, and the commercial pacing mechanism is not a substitute for the Xiangshan proposal. They are instruments for different layers. Reading them as a single conversation is the error most available this week, and the one most worth resisting.
The two quotations are Dong Jun’s, verbatim; no other quotation appears. The level of United States representation at the forum is reported as a matter of record and nothing is inferred from it. The expected Trump-Xi summit and the reported US-China AI talks are exactly that — expected and reported. Neither is an arranged or published schedule, nothing here states what any talks would cover, and no outcome is predicted. The characterisations that China softened its criticism and warned of militarism appear in the coverage of the forum; they are not findings of this publication. Saying that artificial intelligence has no foundational treaty equivalent to the Outer Space Treaty or UNCLOS is not a claim that it is ungoverned. National rules, export controls and international declarations all apply to it; what is absent is an instrument of the same character. Nothing here argues for or against any treaty, or assesses whether the existing two have worked. A release-rate agreement between model developers and a state-to-state security dialogue on military AI govern different objects. Nothing here presents the Xiangshan proposal as an answer to the commercial pacing argument or as filling the gap that argument has where China is concerned, and nothing here says whether it would work. A speech establishes what was said and where. No claim is made about the policy intent, internal process, sincerity, motives or strategy of any government, and none is made about the military AI capabilities, doctrine, deployments or intentions of any country. No outcome, agreement or timeline is predicted. No company is named as a beneficiary or a loser, and no claim is made about any company’s corporate status, valuation, share price or market capitalisation. This is businessanalysis, not investment advice, no view is expressed on any security, and no recommendation is made.
Gennaro is the creator of FourWeekMBA, which reached about four million business people, comprising C-level executives, investors, analysts, product managers, and aspiring digital entrepreneurs in 2022 alone | He is also Director of Sales for a high-tech scaleup in the AI Industry | In 2012, Gennaro earned an International MBA with emphasis on Corporate Finance and Business Strategy.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
