At the recent Daytona Compute Conference, Lin Qiao, co-founder of Fireworks AI, delivered a compelling session on why 2026 is the "year of the agent" and how developers can build a sustainable competitive advantage by owning their AI stack.
The AI Flywheel: From API to IP
The core message was clear: don't treat AI models as mere utility APIs. To build a real "moat", companies must move toward a model where the AI is an integral part of their product IP. While foundation models are built on public data, over 90% of the world's information is locked inside private enterprises. By continuously tuning models on this domain-specific data, businesses create a "data flywheel" where the model learns specific business logic and vocabulary, making the intelligence itself a proprietary asset.
Disruption Across the Board
Qiao highlighted how agents are fundamentally shifting several key sectors:
Coding: Software development has been significantly disrupted by agents like Cursor, which built its fast code generation product on Fireworks to scale from a small startup to a massive user base.
Search: The industry is moving from traditional keyword-based engines to NLP and LLM-based search, allowing for complex, natural language queries in e-commerce and beyond.
Workflow Automation: Using Vision Language Models (VLMs), companies are now automating the extraction of intelligence from non-text sources like PDFs and physical documents: tasks that were previously slow and expensive for legal firms and insurance companies.
Specialized Domains: From healthcare diagnostics to legal research, domain specialty that once required years of training is being semi-automated by specialized agents.
Performance at Massive Scale
Fireworks AI isn't just about experimentation; it’s built for production. Qiao shared some staggering metrics regarding their infrastructure:
Massive Throughput: The platform serves over 13 trillion tokens per day and handles 180,000 requests per second, a volume comparable to Google Search.
Global Virtual Cloud: To combat the "capacity crunch" of GPUs and hardware, Fireworks operates a virtual cloud across 10+ providers and 20+ regions, offering global routing and disaster recovery.
Quality & Speed: By utilizing Reinforcement Fine-Tuning (RFT), Fireworks enables users to tune smaller models that can reach 10% better quality than closed models while being 8x faster and 5-8x cheaper.
For those looking to build, customize, and scale their AI agents, the message was simple: the faster you can iterate and the more you own your intelligence, the stronger your moat will be.
