Own Your Intelligence: How to Do It (7-minute read)

The competition in the AI application layer is transforming into a battle over who owns the intelligence layer itself, rather than UI or workflows. Recently, the performance of open-weight models such as Kimi K3 and GLM 5.2 has improved rapidly, allowing them to surpass frontier models in specific domains.

The reasons for directly owning intelligence are as follows:

• Cost: As AI products succeed, inference costs surge, so model ownership protects margins.
• Speed: In domains where speed is critical, such as coding or security, small custom models are advantageous over large general-purpose models.
• Proprietary Data: If customer interactions or domain data determine system performance, utilizing them for training is effective.
• Control: As the app layer and intelligence layer converge, there is a need to directly control the learning loops that define the product mindset.

According to the roadmap revealed at an event hosted by Sequoia, the key is forming small 'sovereign AI' teams specialized in evaluations and data shaping, rather than platform teams. Harvey is cited as a case where a team of 7 achieved significant research results. Now, when choosing an AI champion, published research and benchmarks are becoming important criteria.