Sequoia's Sonya Huang on Building Sovereign AI

Sequoia Capital partner Sonya Huang kicked off an event focused on "Sovereign AI," a movement for companies to own their artificial intelligence capabilities, down to the model weights, without external dependencies. Addressing an audience of portfolio company founders and AI leaders, Huang emphasized that this isn't about abandoning existing closed models like OpenAI's GPT-4 or Anthropic's Claude, but rather about strategically integrating owned AI into specific product areas.

Huang highlighted that the trend towards sovereign AI is gaining significant momentum, with prominent figures like Alex Karp of Palantir Technologies and Jensen Huang of Nvidia (NASDAQ:NVDA) voicing support. The core message is clear: companies want to own, not rent, their intelligence, recognizing it as too fundamental to outsource.

The Shift from Centralized to Decentralized Intelligence

Huang presented a dichotomy between centralized intelligence, where a single powerful AI system dominates, and decentralized intelligence. The latter, she argued, fosters an ecosystem where individual companies build bespoke AI tailored to their specific data and needs, leading to greater innovation and individuality. This decentralized approach is seen as a more optimistic vision for the future of AI.

Why Companies Are Pursuing Sovereign AI

Several key drivers are pushing companies towards building their own AI stacks:

• Cost: For companies with low or negative margins, AI costs can become prohibitive. Owning models can offer significant cost optimization, especially for those already deploying AI at scale.
• Speed: In critical domains like coding and cybersecurity, a small, custom-distilled model can outperform large general models due to its speed and latency advantages.
• Performance: While historically open models might have lagged behind closed ones, the landscape is changing. Companies can now achieve superior performance by fine-tuning models on their proprietary data.
• Controlling Destiny: Building independent AI capabilities provides companies with greater control and autonomy, reducing reliance on external providers.

Huang drew a parallel to the crypto adage, "Not your keys, not your crypto," applying it to AI with "Not your weights, not your product." This highlights the idea that true ownership of a product is intrinsically linked to control over its underlying intelligence.

A Framework for Building Sovereign Intelligence

Huang outlined a four-step framework for companies embarking on the journey of building their own AI:

1. Strategy: Own vs. Rent: Companies must strategically decide which AI capabilities are core to their business and should be owned, and which can be outsourced. This decision is guided by factors like cost, speed, performance requirements, and the proprietary nature of their data.
2. Team: Assembling the right talent is crucial. This involves identifying leaders with either a strong research or engineering background, depending on the focus of the AI efforts. Huang stressed the importance of creating dedicated, agile teams for sovereign AI initiatives, rather than shoehorning them into existing platform teams.
3. Legibility: Building internal expertise is not enough; companies must also control the narrative and ensure their capabilities are legible externally. This involves excellent technical marketing, potentially branded research groups, and high-quality publications to demonstrate sophistication and expertise to potential customers.
4. Technical Roadmap: The path involves defining strategy and rigorous evaluation, followed by experimentation with model routers and harnesses. Companies may then move to post-training, mid-training, or even pre-training, ultimately aiming to create feedback loops where live customer data continuously improves the AI.

The Race for the Intelligence Layer

Huang concluded by emphasizing that the competitive battleground is shifting from the application layer to the intelligence layer itself. The companies leading this charge are those that can not only control product surfaces but also shape and improve the core intelligence powering their products. She pointed to examples like Harvey AI, which has demonstrated significant research and development in this space, highlighting the growing importance of applied research from companies themselves.