Geopolitics & Sovereign Positioning

ByteDance's Seed AI team has published research identifying a new scaling law for AI agents: systems can double their effective learning speed every three months by accumulating experience through real-world task execution, rather than requiring ever-larger pre-training compute runs. South China Morning Post reports the finding arrives precisely as traditional compute-scaling approaches face diminishing returns.

The strategic significance is substantial. The foundational logic of US export controls on advanced semiconductors to China rests on the assumption that frontier AI capability is inseparable from access to cutting-edge training hardware. An agentic scaling pathway that derives performance gains from deployment interaction rather than pre-training compute does not eliminate the relevance of chips, but it introduces a second axis of capability growth that controls cannot directly target. If validated at scale, this would mean Chinese labs could compound capability in deployed systems even while remaining constrained on training infrastructure — a meaningful shift in the effectiveness calculus of the current sanctions regime. The finding is a preprint-stage research claim and requires independent replication, but its policy implications warrant immediate attention.

The US Department of Defense's GenAI.mil platform has reached approximately 1.7 million users, with the Pentagon's chief AI officer signalling plans to expand the model roster available on the platform. Defense One frames this as a moment of operational momentum rather than pilot-phase experimentation.

The user scale matters less as a capability metric than as an indicator of institutional embedding. At 1.7 million users across a workforce of roughly 3.5 million DoD personnel, generative AI is no longer confined to innovation units — it is becoming part of routine military staff work. This creates structural demand that will drive procurement, security architecture decisions, and eventually doctrine. The comparison that matters geopolitically is with the PLA's parallel efforts to integrate AI into command and staff functions, where public data is sparse but investment signals are consistent. The US is now accumulating operational learning at scale; the question is whether that translates into doctrinal and procurement advantage faster than adversaries can close the gap through their own institutional integration.

Alibaba has formally added Anthropic's Claude Code to its internal high-risk software list, prohibiting employee use and citing a previously identified instance of hidden tracking code targeting Chinese users. South China Morning Post reports the ban follows broader industry backlash in China over the incident.

This is a consequential data point in the longer-term bifurcation of enterprise AI tooling. Whether or not the original Anthropic code instance was intentional surveillance or a security misconfiguration, the political effect is the same: it provides a credible, publicly defensible rationale for Chinese enterprises to systematically audit and exclude US AI development tools. The cascade risk is significant — if major Chinese technology companies standardise on domestic equivalents for coding assistants and AI development infrastructure, they accelerate the maturation of those alternatives while simultaneously shrinking the footprint of US AI firms in the Chinese enterprise market. This is a trust-based market exit that export controls did not mandate but that US actions have enabled.

Zhipu AI is raising HK$31.4 billion through an accelerated secondary placement after shares surged 9 percent, while Iluvatar CoreX — a domestic AI chip designer operating under US sanctions — is conducting a parallel raise for hardware expansion. Separately, Unitree Robotics has received regulatory approval for a Shanghai IPO that will set valuation benchmarks for China's embodied AI sector. South China Morning Post and South China Morning Post report both market moves on the same day.

The simultaneous capital formation across model development, chip design, and physical AI represents a maturing of China's sovereign AI financing ecosystem. Iluvatar CoreX's ability to raise billions in Hong Kong public markets while under US export controls is particularly notable: it demonstrates that sanctions have not impaired access to growth capital for sanctioned entities operating within Chinese and Hong Kong market structures. The robotics IPO pipeline, with Unitree leading, will test whether public market valuations can sustain the VC flood into embodied AI — but the more durable signal is that China is building a self-reinforcing domestic capital cycle for AI infrastructure that does not depend on US institutional investors or NASDAQ listings.

India's Bhashini initiative is sponsoring a hackathon explicitly oriented toward offline, multilingual AI tools — a design choice that prioritises functionality in low-connectivity environments and across India's linguistic diversity over integration with dominant cloud-based Western AI platforms. Rest of World frames this as a direct challenge to the Silicon Valley model of centralised, cloud-dependent AI delivery.

The strategic logic is sound: India's dependency on Western AI cloud infrastructure is a sovereignty risk, and building offline-capable, locally deployable AI in Indian languages creates capabilities that US or Chinese platforms have limited commercial incentive to develop at equivalent quality. Bhashini represents a state-anchored demand signal for a specific architecture of AI — distributed, open-source, linguistically sovereign — that could become an export model for other large multilingual emerging economies. This positions India not merely as a rule-taker in global AI governance but as a potential originator of an alternative development paradigm attractive to the Global South, where connectivity constraints and linguistic diversity create similar dependency dynamics.