Chinese startup Moonshot AI has garnered international attention and shattered confidence in US AI leadership by releasing the powerful open-weight LLM Kimi K3 in July. K3 matches – and by some benchmarks outperforms – the most advanced US models. And it makes its weights public. Despite the media hype, K3 is not another DeepSeek moment. Moonshot has not demonstrated any real breakthroughs like DeepSeek did in January 2025 when it unveiled  its R1 model. Kimi is competing with US models on their own turf, using size as a measure of the best model. This shows that China is still a fast follower in a race whose terms were set by US companies like OpenAI and Anthropic. 

DeepSeek demonstrated that optimizing design can drastically reduce the financial and computing resources required to train an advanced, price-competitive AI model. Moonshot, like all other AI labs currently, deployed a Mixture-of-Experts (MoE) architecture for K3, which splits a model into multiple “expert” sub-models and only activates the parts of the network that matter for a given task. But with its 2.8 trillion parameters, K3 is the largest open-weight model ever built, a different beast from small, efficient reasoning models like DeepSeek’s R1. Models this large require a lot more hardware (hence also money and electricity) to train and run.

To move from being a fast follower to a leader, China would need to overcome its current computing power gap, due to its short supply of advanced chips, or move toward a different paradigm where bigger does not equal better. China already has some structural advantages. Lower operational costs and cheap and abundant electricity, if paired with efficient architectures and resource allocation, could allow Chinese AI firms to compete on price and speed up adoption. 

Yet, as DeepSeek’s CEO admitted in July during a leaked investor call, Chinese AI companies lack the computing resources (because of US export controls and Huawei’s production constraints in indigenous chips) to properly train large frontier models. Indeed, Anthropic and the US government have accused Moonshot of illegally copying US models using fraudulent accounts. After releasing K3 in July, Moonshot ran into a capacity crunch which forced it to temporarily suspend new subscriptions.

MERICS Head of Program Science, Technology and Innovation (Co-lead) Rebecca Arcesati: “Regardless of the merits of the US accusations, bigger currently equals better in advancing AI. The DeepSeek moment was so transformative because it proposed a different approach, one that is more efficient, sustainable, and accessible. Speed, cost, and energy efficiency will continue to matter, but overall, it seems that China is now playing the same hugely resource-intensive game as the US.”