How Companies Are Building Their Own Intelligence | Sonya Huang, Sequoia Capital

Event Context & Strategic Goal

• The gathering included approximately 80 portfolio company founders and AI leaders to discuss "Sovereign AI" (companies owning their own intelligence down to the model weights).
• The agenda combines strategic rallying with technical "how-to" sessions, demos, and workshops on building custom AI capabilities.
• Definition: Sovereign AI refers to the vertical integration of owning intelligence rather than renting it via external APIs, distinct from the debate on whether to leave closed APIs entirely.
• Market Sentiment: Support for sovereign AI has escalated sharply in the last month, with endorsements from industry leaders like Alex Carp, Satya Nadella, and Jensen Huang regarding the availability of open-weight models in the U.S.

Core Drivers for Adopting Sovereign AI

• Cost Efficiency: For companies with low or zero-negative margins, owning models is a financial necessity rather than a luxury; high AI deployment costs correlate directly with advanced product usage.
• Speed & Latency: Custom, distilled models outperform large general models in P0 domains like coding and cybersecurity where sub-second latency is critical.
• Performance Optimization: Open-weight models can now outperform closed models in specific, proprietary domains through fine-tuning, challenging the historical assumption that closed APIs offer superior performance.
• Strategic Independence: Companies seek to control their own destiny ("not your weights, not your product"), reducing reliance on external providers like Anthropic or OpenAI.
• Decentralized Intelligence: The shift favors a decentralized ecosystem where companies build bespoke intelligence on their own data, preventing a single centralized AI from controlling global GDP.

Strategic Framework for Ownership

• Strategy (Step 1): Organizations must define which capabilities to own versus rent based on four factors: cost impact, latency requirements, performance needs, and the proprietary nature of the data.
  • Example: Coding agents remain mostly rented for out-of-the-box performance, while tab autocomplete models are often sovereign due to high frequency and latency sensitivity.
  • Example: Cybersecurity and biotech firms increasingly own models to leverage proprietary domain data and specific performance tuning.
• Team Assembly (Step 2):
  • Leadership profiles vary between research-heavy (e.g., Apple/Google Brain backgrounds) and engineering-heavy (e.g., Microsoft/Ramp backgrounds).
  • Organizational Structure: Companies are advised to build dedicated, de novo "lab" teams rather than shoehorning sovereign AI into existing platform teams, as it requires "offense" research rather than "servicing" capabilities.
  • Benchmark: Harvey, a cybersecurity AI company, achieved significant research output with a team of only seven people.
• Legibility (Step 3):
  • External legibility via published research is critical for buyer confidence and distinguishing sophisticated vendors in a crowded market.
  • Strategies include establishing branded research groups, publishing high-taste technical papers, and practicing excellent technical marketing.
• Technical Roadmap (Step 4):
  • Phase 1: Define strategy and establish robust evaluation frameworks (evals).
  • Phase 2: Implement model routers and harnesses; move from out-of-the-box performance to post-training or mid-training.
  • Phase 3: Establish online learning loops where live customer data continuously improves model intelligence.
  • Goal: Leverage open-weight models (e.g., Kimi K3, GLM 5.2) to reach or exceed frontier-level performance through strong post-training, prompt engineering, and online learning.

Architectural Stack & Infrastructure

• Production Stack: Consists of a custom harness layered over an open-source base model to drive user-facing tokens.
• Development Stack: Becomes critical and complex, requiring tools for monitoring evals, managing data drift, sourcing high-quality data (expert trajectories, synthetic data, RL environments), and configuring post-training pipelines.
• Context Management: Performance is heavily influenced by context handling via vector databases (e.g., TurboPuffer), knowledge graphs (e.g., Glean), and novel memory encoding techniques (e.g., Ngram's research on encoding context into weights).
• Closed vs. Open Comparison:
  • Closed Stack: High floor, low ceiling; simple API calls but limited ability to improve intelligence via proprietary data.
  • Open Stack: Lower floor due to complexity, but offers a higher ceiling by enabling full data utilization and continuous model improvement.

Upcoming Workshops & Partners

• Post-Training: Led by Lynn (Fireworks).
• Harnesses & Evals: Led by Harrison (Langchain).
• RL & Synthetic Data: Led by Brendan (Mercore).
• Online Learning: Led by Trajectory.
• Full Stack Strategy: Led by Harvey (who recently announced Harvey Research).