Geopolitics & Sovereign Positioning

A detailed analysis in War on the Rocks challenges the dominant Western framing of China's AI rise as a state-orchestrated campaign. The piece argues that China's 2030 AI leadership target has become a rhetorical fixture in policy circles that obscures a more complex reality: the firms actually driving Chinese AI capability — MiniMax, ByteDance, DeepSeek, Zhipu AI — are being forged through ferocious commercial competition, not bureaucratic coordination. The 15th Five-Year Plan does direct Party organs to take a more active role, but the competitive dynamics are market-driven in ways that make Chinese AI harder to target through export controls or technology denial strategies.

This matters for how Western policymakers design countermeasures. If Chinese AI capability is primarily a function of domestic market competition rather than state subsidy and directed research, then controls aimed at choking off state-linked entities will have limited effect on the broader ecosystem. MiniMax's M3 model — which reportedly reduces computational requirements to one-twentieth of prior levels and processes 1 million token contexts — is a commercial product built for coding agents and enterprise workflows, not a state procurement project. South China Morning Post reports the efficiency gains are substantial, meaning Chinese firms are actively engineering around compute constraints imposed by export controls — adapting rather than halting.

DeepSeek ranked first on Ramp's June trending software vendors list — a tracker of first-time corporate purchases — displacing established US providers including OpenAI and Anthropic on cost grounds, according to South China Morning Post. This is a qualitatively different problem from chip export controls or military technology transfer: Chinese AI software is penetrating US enterprise markets through price competition, and there is no enacted policy mechanism that currently prevents it. The Biden-era and Trump-era export control frameworks were designed to deny China access to advanced semiconductors and restrict outbound technology transfer — they were not built to stop American firms from buying Chinese AI inference services.

The strategic implication is a two-track dependency problem. US policymakers are focused on preventing China from acquiring American technology; simultaneously, American enterprises are adopting Chinese AI at scale, creating data exposure and vendor dependency risks that intelligence agencies have flagged but Congress has not resolved legislatively. A DeepSeek ban has been proposed in specific federal contexts but no comprehensive restriction on private-sector use is enacted as of this date. The commercial momentum is running ahead of the policy response.

Rest of World reports that compute scarcity in India, Brazil, the UAE, and across Africa is driving the development of local AI infrastructure stacks — hardware configurations, model architectures, and data pipelines designed specifically to operate under resource constraints. This is not catch-up mimicry of Silicon Valley approaches; it is differentiated architecture built for different input conditions. The UAE's positioning is particularly notable: Abu Dhabi has invested heavily in sovereign AI infrastructure through G42 and related vehicles, positioning itself as a neutral AI hub capable of serving both Western and non-Western clients.

For foreign policy strategists, the significance is that these countries are transitioning from passive recipients of AI technology — dependent on US or Chinese cloud infrastructure and foundation models — toward builders of indigenous capacity with genuine strategic optionality. This shifts the geopolitical calculus: nations with their own AI stacks are less susceptible to technology leverage, less likely to align exclusively with either the US or Chinese AI ecosystems, and more capable of setting domestic standards. Asia-Pacific enterprise investment data from IDC, cited by South China Morning Post, shows 37 percent of regional firms investing aggressively in AI with minimal outcome evaluation — nearly double the global average — reflecting the FOMO-driven capital mobilisation that is funding much of this infrastructure buildout.

TSMC chairman C.C. Wei publicly dismissed concerns about competition from Huawei and Chinese foundries at the company's annual shareholders' meeting, saying the company was 'not afraid' and that competition had always existed, according to South China Morning Post. The posture is defensible at leading-edge nodes: TSMC's 2nm and below process technology remains years ahead of anything SMIC or Huawei can produce, and advanced AI training workloads still require that capability. However, the strategic question is not whether Chinese foundries can match TSMC at 2nm — it is whether Chinese chipmakers can produce sufficient volumes at 7nm and above to sustain domestic AI inference deployment at scale, which US export controls are less effective at preventing.

The distinction matters operationally. AI training — building foundation models — requires the most advanced chips and is where TSMC's lead is most consequential. AI inference — deploying models at scale — is less compute-intensive and can be done on older process nodes. If China can sustain inference-capable chip production domestically, the export control strategy succeeds in slowing Chinese frontier model development while failing to prevent Chinese AI deployment at national scale. MiniMax's M3 efficiency gains, reducing compute requirements by a factor of twenty, directly expand the range of chips on which Chinese AI products can run competitively.