Below is now the consensus on DeepSeek. Nvidia, trading over 7% today is the final nail on the coffin as the confirmation signal:

With advancements in distillation research, it’s becoming increasingly evident that optimizing for cost efficiency such as reducing resource requirements or tolerating higher latency does not necessarily lead to models with enhanced performance or capabilities. While the ability to serve models at a lower cost is valuable for scalability, it should not be mistaken for the breakthroughs required to achieve higher capabilities or improved outcomes. This growing decoupling emphasizes the need to treat cost-efficiency and capability development as distinct and separate challenges in AI model innovation.

Bottom line: Lowering costs in model serving does not inherently result in better model capabilities, as these two aspects are becoming increasingly independent.

That said, a model that delivers 90% of the performance of o1 but comes with no usage limits is exponentially more useful in a knowledge worker’s daily routine than an o1-tier model restricted to approximately 10 queries per day.