Conference Presentation, Fireside Chat
Fireside Chat with Lin Qiao, CEO & Co-Founder of Fireworks AI | RAISE Summit 2026
- Event Overview: The session featured a discussion between Matt Miller, Managing Director at Avantik VC, and Lin Chao, CEO and Co-founder of Fireworks AI, hosted at the RAISE conference.
- Avantik VC Status:
- Matt Miller formerly served as a partner at Sequoia Capital for 13 years before launching Avantik VC over a year ago.
- The firm closed its first fund on June 30th of the previous year.
- Lin Chao's company, Fireworks AI, was the first founder contacted after the fund closed, establishing a pre-existing relationship from Miller's time at Sequoia.
- Fireworks AI Value Proposition:
- Fireworks operates as a specialized intelligence platform focusing on customized models built using a company's private data.
- The core mission is to enable companies to own their intelligence rather than relying on external APIs, thereby creating durable business moats.
- The platform acts as an enabler for open-source models, contrasting with "closed" foundational model providers akin to "Snowflake" versus Fireworks' role as a "Databricks."
- Founding Team and Origins:
- The founding team consists of seven members, including Lin Chao, who previously built AI infrastructure at Meta for seven years, including leading PyTorch efforts.
- The company was founded to address the high barrier of capital and talent density required for pre-GAI specialized intelligence (e.g., recommendation systems, self-driving cars).
- Market Trends and Strategic Shifts:
- Transition to Fine-Tuning: The rise of foundation models has shifted the requirement from building models from scratch to post-training on private data, making specialized intelligence significantly less capital intensive.
- Democratization: Specialized intelligence is now accessible to more companies due to reduced talent density requirements and the availability of pre-trained foundation models.
- Software Development Disruption: Application development has become commoditized; a single non-technical user can build and scale an application in weeks using AI, reducing the barrier to entry and eroding the defensibility of product ideas.
- Data as the Moat: With application replication becoming easy, competitive differentiation now relies on proprietary data (product analytics, user intent) used to train customized models.
- Key Decisions and Strategic Directives:
- Specialization over Generalization: There is no such entity as a "specialized general company"; differentiation requires owning a specialized intelligence model specific to the use case.
- Model Evolution: Customization is a continuous, iterative process, not a one-time event, as applications evolve and base models update weekly.
- Budget Allocation Strategy: Companies are advised to allocate 5% to 20% of their total AI budget toward training/inference customization to achieve a 5x to 10x reduction in long-term inference costs.
- Future Architecture: The industry will move toward a blend of models with intelligent routing, rather than relying on a single general-purpose model for all tasks.
- Cost and Efficiency Metrics:
- Customized models trained on private data can reduce inference costs by a factor of 5 to 10 times compared to using generic API wrappers.
- This efficiency extends company cash flow and improves long-term business health.
- Open vs. Closed Models:
- Open-source models have reached a quality threshold where they can easily beat frontier closed models when tuned and customized.
- Full control over model weights allows for deeper customization and better alignment with specific product metrics.
- Customer Onboarding and Support:
- Fireworks offers a self-serve public platform with comprehensive documentation for immediate use.
- For serious deployments, the company provides researcher support to assist with model tuning and platform optimization.
- CEO Operational Tactics:
- Lin Chao uses AI agents to summarize high-volume internal communications (e.g., Slack) to maintain organizational awareness and agility.
- The primary use case for AI in her own workflow is information synthesis and real-time signal detection regarding product performance and customer feedback.
- Forward-Looking Statements:
- The future of AI will not be dominated by a few AGI models but will consist of millions of specialized models tailored to specific applications and use cases.
- Both specialized intelligence strategies and AGI will co-exist, but the immediate path to durable business value lies in owning and customizing models.
- The industry trend points toward "no specialized general company," emphasizing that every business must build unique intelligence to survive commoditization.