Leverage in this negotiation is countable. Boston Consulting Group puts 2025 capital expenditure at top US technology companies above $400 billion against $63 billion in China [19], a ratio of roughly 6.3 to 1 [1] and a one-year absolute gap of $337 billion [2], and BCG's wider read is of two ecosystems separating across models, chips and infrastructure [18]. Set that against PwC's projection of $31.6 trillion of global AI infrastructure spending through 2050, about 48% of it, or $15.1 trillion, American [20]: spread over the 25 years to 2050 that averages roughly $604 billion a year, about one and a half times what BCG counts for 2025 [3]. Compute access is the asset, and standards are the covenant written over it.
Which is why the name at the head of the table is more informative than the agenda. A Treasury Secretary does not evaluate frontier models, and the technical complaint Washington wants to make has been sitting with the science office since June, when Michael Kratsios accused Moonshot AI of distilling Anthropic's Fable model to help build its K3 release [14]. The ask itself is cheap to grant, being self-policing and threat-sharing on top of a voluntary US policy that already gives the government a look at covered frontier models before release [8].
Scott Singer of the Carnegie Endowment told Reuters that Beijing has questioned whether the US regulates its most advanced models sufficiently, and that both sides are motivated to manage a cross-border crisis effectively [12]. That is the honest description of what a first channel buys, since the report expects no agreement while the rivalry over chips, compute, models and standards continues [6]: a phone number. The incident material Reuters describes, nearly 700 malicious agents built on OpenAI models hitting Hugging Face in July and writing counterfeit logs to hide the trail [9], is what such a number would be for. Paul Triolo of DGA-Albright Stonebridge Group told Reuters it is now or never [11], and the clock he is describing has a number attached: a US government evaluation cited by CSIS puts DeepSeek V4 Pro about eight months behind the leading American models [17].
The reading I would take is that the channel is real and standards are now tradeable against tariffs, because a technical dialogue does not get a Treasury delegation head [2]. The thinner version has the meeting happening and producing letters and a hotline, nothing else. The dullest version is that the White House statement is the accurate one [4] and there is no channel to price at all. The counter to my read is depreciation: if the strongest Chinese open-weight model is already close to US closed models on the tasks buyers pay for, Washington is bargaining with an asset losing value faster than the bargaining moves, and Capital Group's observation that DeepSeek's mixture-of-experts work has been adopted by several major labs [21] cuts the same way, because diffusion runs in both directions. What would prove the trade-file thesis wrong is a September 24 summit [3] that moves chips without mentioning standards.
