Geopolitics & Sovereign Positioning

OpenAI's decision to slash pricing for its GPT-5.6 Luna model by 80% — bringing API costs to $0.20 per million input tokens — is not a routine commercial adjustment. It is a strategic retreat forced by Chinese competitors, particularly DeepSeek and its successors, who have systematically undercut Western AI pricing while matching or approaching frontier performance. As South China Morning Post frames it, OpenAI is 'blinking' in a face-off it did not anticipate fighting on these terms.

The geopolitical consequence is significant. Low API pricing compresses the commercial moat that US AI firms relied upon to lock in enterprise and developer ecosystems globally, particularly in the Global South where cost is a primary adoption barrier. If Chinese firms can offer comparable capability at lower cost — and increasingly do so through open-weight releases that eliminate API dependency entirely — the infrastructure-layer advantage the US sought to preserve through export controls on chips is being partially circumvented at the software layer. A CFR discussion featuring President Froman flags this dynamic, noting US concern about how to maintain technological advantage as Chinese frontier models close the gap.

RedNote's reported plan to build a 600MW data centre in Ulanqab, Inner Mongolia — budgeted at $2.2 billion — is one data point in a broader pattern of Chinese private-sector firms internalising AI infrastructure at scale, partly in response to US export controls limiting access to leading-edge Nvidia chips. Inner Mongolia's renewable energy base makes it a strategic location for power-hungry AI compute, and the project reflects Beijing's sustained policy push to concentrate sovereign AI infrastructure domestically. As South China Morning Post reports, this is one of several such projects underway across Chinese tech firms.

Simultaneously, Chinese MLCC manufacturers — producers of multilayer ceramic capacitors, a foundational passive component in all electronic devices — are posting record earnings driven by AI infrastructure demand, with stocks hitting daily trading limits. South China Morning Post notes that Guangdong Fenghua and peers are benefiting from a supply chain that US controls have not meaningfully constrained. This matters because component-level self-sufficiency is a prerequisite for China to build AI hardware at scale without foreign dependency — and that capacity is clearly advancing.

An Atlantic Council analysis argues that the decisive variables in Africa's AI infrastructure competition are financing access, energy permitting, and physical infrastructure — not model quality or algorithmic superiority. The piece presents African deployments as evidence that China's infrastructure-first approach, backed by state financing and established construction relationships from BRI, gives it structural advantages that US firms competing on capability benchmarks alone cannot easily overcome.

This framing has direct policy implications. If the Global South selects AI stacks primarily on the basis of who builds the power plants, lays the fibre, and provides concessional financing for data centres — rather than who has the best frontier model — then US AI dominance in benchmark leaderboards translates poorly into geopolitical influence. The US response, including initiatives like the Partnership for Global Infrastructure and Investment and recent bilateral AI agreements, is attempting to match China's infrastructure financing, but the Atlantic Council assessment suggests the gap remains. This is a geopolitical commentary piece, not a confirmed policy finding, but the analytical framing is consistent with observable deployment patterns.

South Korean President Lee's itinerary — semiconductor investment pledges from Silicon Valley firms followed immediately by trade and mineral agreements in Brazil and Chile — illustrates a sophisticated hedging strategy by a mid-tier power with acute AI supply chain exposure. South Korea is home to Samsung and SK Hynix, two of the world's critical HBM and DRAM producers, making it a linchpin in both US and Chinese AI hardware supply chains. As The Diplomat reports, Lee is simultaneously deepening the technology partnership with the US while securing upstream mineral supply — lithium, copper, and rare earths — that underpin next-generation chip and battery production.

The strategic logic is clear: South Korea cannot afford to be a passive participant in US-China technology decoupling. By locking in US investment in Korean semiconductor capacity while diversifying mineral sourcing away from Chinese-controlled supply chains, Seoul is positioning itself as an indispensable node in the US-aligned AI hardware ecosystem while reducing its own vulnerability to Chinese resource leverage. The specific terms of the Silicon Valley pledges and the South American mineral agreements have not been fully disclosed, so the binding force of these commitments remains unclear from available reporting.

MiniMax's release of the H3 video generation model with open weights — topping video editing benchmarks according to Artificial Analysis — and ByteDance's concurrent Seedance 2.5 release represent a deliberate Chinese strategy of open-source proliferation in AI media generation. As South China Morning Post reports, MiniMax explicitly frames open weights as a challenge to closed-source dominance. Open-weight releases cannot be effectively controlled by US export restrictions once published — they are downloadable globally and require only sufficient local compute to run.

A separate South China Morning Post piece on visits to Shanghai AI labs frames this as a soft power contest, with Chinese labs actively cultivating foreign audiences and positioning their AI development as an alternative paradigm to US-centric approaches. The combination of open-weight releases, competitive pricing, and deliberate international engagement creates a multi-layered influence strategy that operates largely outside the perimeter of US export controls, which are designed around hardware access rather than software distribution.