Alibaba, ByteDance, DeepSeek race to build 10-trillion-parameter AI models

Common ground

• All sides agree that the push for trillion-parameter AI models is driven by geopolitics, not just technology.
• There is agreement that US export controls on chips have forced Chinese companies to pursue domestic alternatives.
• All participants recognize that the parameter count is partly a marketing signal, not purely a technical milestone.
• There is consensus that inference cost per token is a critical metric for real-world AI usefulness.
• All sides acknowledge that both Chinese and Western AI models reflect the values and priorities of their home countries.

Points of contention

• Eastern Agent sees China's AI investment as a strategic necessity for sovereignty, while Western Agent views it as a tool for authoritarian control.
• Neutral Agent argues that state-backed financing has real limits, but Eastern Agent insists China's banks have a longer-term risk tolerance.
• Western Agent claims Chinese AI exports are coercive bundled deals, while Eastern Agent says they are free market choices by the Global South.
• Neutral Agent and Eastern Agent disagree on whether the Scaling Law is a fixed constraint or a research frontier that can be overcome.
• Western Agent argues that China's training data is ideologically poisoned, while Neutral Agent says the models still perform well on objective reasoning tasks.

Blind spots

• All sides focused heavily on parameter counts and governance, but barely discussed the environmental impact of training trillion-parameter models.
• The debate overlooked how smaller, specialized AI models might outperform giant ones for most practical applications.
• No one addressed the potential for AI regulation or international agreements to manage the risks of this arms race.
• The role of open-source AI development as a counterweight to both US and Chinese dominance was not explored.

WorldAttention’s read

This debate revealed that the trillion-parameter AI race is a proxy for deeper geopolitical struggles, not just a technical competition. While all sides agree that US export controls pushed China toward self-reliance, they sharply disagree on whether this is a legitimate pursuit of sovereignty or a dangerous export of authoritarian surveillance. The parameter count itself was largely a distraction—the real battleground is inference cost, data quality, and which governance model gets embedded in the world's AI infrastructure. The Global South's choices are shaped by both US and Chinese power, not made in a vacuum, and neither side fully acknowledged that their own AI systems carry unaccountable biases. Moving forward, the conversation needs to shift from model size to practical deployment, environmental costs, and the need for global AI governance that respects diverse values without enabling oppression.