What happened
Chinese companies, exemplified by DeepSeek AI, are developing efficient AI models, such as DeepSeek-V3, which utilise Mixture of Experts (MoE) architectures. DeepSeek-V3, with 671 billion parameters, activates only 37 billion per token during inference, enabling comparable performance to larger models like GPT-4 at a fraction of the training cost and computing power. This contrasts with the US focus on large, complex models and AGI, as China prioritises practical, state-driven AI deployment.
Why it matters
The proliferation of highly efficient AI models, developed with significantly reduced computational and financial overheads, introduces a new competitive constraint for organisations relying on traditional large-model development paradigms. This increases due diligence requirements for R&D and procurement teams to assess the true cost-performance ratio of AI solutions, and for strategic planning to adapt to a rapidly evolving, resource-optimised AI landscape. It also creates a visibility gap regarding the optimal investment in AI infrastructure and model development, as established benchmarks for scale and cost are being redefined.
