sonya huang

Application companies are vertically integrating to own their intelligence layer — driven by cost pressure at scale, latency needs, domain-specific performance gains from open models, and proprietary data ownership — turning every AI-native company into an applied research lab.

New open weight models from Moonshot AI (Kimi K3) and Zhipu AI (GLM) now provide baselines close to closed-model frontier, making it feasible for companies to achieve better-than-frontier performance through post-training, harness engineering, and online learning on their own data.

New open-weight models from Moonshot AI and Zhipu AI have achieved near-frontier performance while providing full weight access, making them superior starting points for custom post-training compared to closed APIs. This unlocks a new technical roadmap where companies can surpass frontier performance by owning their stack.

Building sovereign AI demands a complex infrastructure stack: a production harness on top of a custom model, plus a development stack for evals, high-quality post-training data (expert trajectories, synthetic data, RL environments), and online learning loops. Best-in-class tools like Fireworks, LangChain, Mercor, and Trajectory are emerging for each layer.

Owning intelligence transforms the stack from simple API calls to a complex production harness (model + context + tools) plus a development stack (evals, data, RL environments, online learning) — creating massive demand for specialized tooling at each layer.

For companies with zero or negative margins on AI features, sovereign AI is a must-have not a nice-to-have. The more successful the AI product, the higher the API costs, creating a paradox where early AI adopters are the first to vertically integrate. This economic imperative drives the build vs. buy decision more than performance today.

Tab autocomplete in coding and security detection are early sovereign AI winners because latency is a P0 and API costs scale with usage — small distilled custom models beat large general ones on speed, while post-training on domain data beats closed APIs on performance.

The build-vs-buy decision for AI capabilities hinges on four factors: COGS impact, latency sensitivity, whether domain fine-tuning beats closed models, and data proprietary-ness — with the balance shifting toward owning as open models improve and deployment scales.