Artificial intelligence’s newest contest is forming around control points: who may release a model, who receives early access, and where the machines are installed [E1][E3][E5][E7]. This month, Moonshot rationed subscriptions, Xi Jinping sold openness abroad, 29 countries signed onto a Shanghai institution, Washington built a government review channel, and Hyundai moved to consolidate Boston Dynamics [E2][E3][E4][E5][E7]. These acts arise from different systems and incentives, yet each converts technical capability into a governed route to users, states or factories [E2][E4][E5][E6][E7]. The inference is that power is accumulating at the pipes surrounding the model, alongside the model itself [E1][E2][E3][E5][E7].
Kimi K3 made the physical constraint unusually visible. Moonshot described a 2.8-trillion-parameter, native-vision model with a one-million-token context and promised full weights by 27 July [E1]. Demand then outran available compute, prompting the company to pause new subscriptions while preserving service for existing users [E2]. The launch’s first public emergency was congestion, a reminder that open weights still need scarce serving infrastructure to become usable access [E1][E2].
Xi Jinping used the World AI Conference to turn openness into foreign policy. He called for open source, collaboration and sharing, opposed an expansive use of national security, and paired that pitch with training programmes and regional cooperation centres [E3]. Twenty-nine countries also signed an agreement to create the World AI Cooperation Organization with headquarters in Shanghai [E4]. Taken together, the speech and institution offer the Global South a channel into Chinese models, standards and technical relationships, with access itself serving as diplomacy [E3][E4].
Washington is constructing a more selective route. A June executive order created a classified advanced-cyber benchmark and a voluntary framework under which laboratories may give the government up to 30 days of pre-release access to covered frontier models [E5]. The July launch of GOLD EAGLE added an AI cyber-vulnerability clearinghouse for coordination around discovered weaknesses [E6]. The result is a privileged state channel at the release perimeter, turning model publication into a checkpoint for security testing and government visibility [E5][E6].
Hyundai’s Boston Dynamics move shifts the same contest onto concrete. Its shareholders are pursuing SoftBank’s entire remaining stake, which would concentrate ownership of the robotics company inside the industrial group [E7]. Atlas is planned for Hyundai’s US Metaplant from 2028, placing the humanoid inside a named factory and capital plan [E7]. Embodied AI becomes more consequential when the owner of the robot also controls the plant, deployment schedule and balance sheet that can absorb it [E7].
Capability remains a hard limit on this thesis. Kimi K3’s scale and context window are competitive claims, and Washington’s classified benchmark exists because differences in cyber capability still matter [E1][E5]. The language of “open” and “gated” also works as branding over two states pursuing advantage through standards, access and security policy [E3][E4][E5][E6]. A coherent grand design should not be read into every move: Moonshot faced a capacity shortage, governments followed institutional incentives, and Hyundai followed an industrial ownership logic [E2][E5][E7].
Even with that caution, the control points now align across software, diplomacy, security and machinery [E2][E4][E5][E6][E7]. Moonshot decides when weights and subscriptions move, Beijing builds an international access network, Washington reserves a pre-release window, and Hyundai is placing Atlas inside its own production geography [E1][E2][E3][E4][E5][E7]. Each actor is securing the route between capability and real-world use, where scarcity, permission and ownership become leverage [E2][E4][E5][E7]. The decisive score may be written at the release gate, the government channel and the factory floor [E1][E2][E5][E7].