Conference Presentation, Keynote, Fireside Chat, Tutorial
How Companies Are Building Their Own Intelligence | Sonya Huang, Sequoia Capital
- Shift toward sovereign AI is driven by companies seeking vertical integration to own intelligence layers, fueled by cost efficiency for low-margin operations and speed advantages from distilled custom models over general ones in specific domains.
- Future market expectations anticipate a decentralized ecosystem where bespoke intelligence triumphs over centralized black-box models, with product definition shifting from UI to the underlying intelligence layer in a "race for the intelligence layer."
- Strategic decisions regarding ownership versus rental will be dictated by factors including cost, latency, performance, and proprietary data, with coding agents likely remaining rented while tab autocomplete and bio-sector applications move to sovereign models.
- Organizations are advised to establish de novo teams with frontier-level research capabilities rather than repurposing existing infrastructure, potentially starting with small groups capable of significant output similar to Harvey's seven-person team.
- Operational roadmaps include a progression from strategy definition and evaluation setup to model routing, post-training, and online learning feedback loops to achieve frontier-level performance by 2026 using open weight models like Kimi K3 and GLM 5.2.
- Performance differentiation is expected to rely on novel context approaches, high-taste research publication for vendor legibility, and a production stack functioning as a harness over models supported by technical expertise from entities like Fireworks, Langchain, and Harvey.
- The competitive battleground emphasizes the "not your weights, not your product" dynamic, with application companies emerging as new "Neolabs" to drive innovation and allow distinct independence from partners like Anthropic and OpenAI.