971: 90% of The World’s Data is Private; Lin Qiao’s Fireworks AI is Unlocking It

Episode Summary

Executive Summary: Fireworks AI CEO Lin Chow argues that most valuable AI intelligence remains trapped in private enterprise data and that the future is continuous, automated customization of models for each application. He explains how Fireworks abstracts model, hardware, and deployment complexity, enables open-model adoption, and supports a small-big-small tuning workflow to produce fast, cost-effective specialized systems.

Main Topics: Autonomous intelligence vs. one-model AGI (Priority: 5/5): Lin frames Fireworks’ mission as activating private enterprise data and continuously customizing models, arguing the future is millions of specialized models rather than one universal foundation model. Open-source models and democratized AI infrastructure (Priority: 5/5): Fireworks’ bet on open models is tied to PyTorch roots, lower economics, and greater control for customers; Lin says open-model performance is converging with closed models. Model, hardware, and fleet optimization (Priority: 5/5): The platform helps customers navigate rapid model and hardware depreciation by abstracting selection across quality, speed, cost, and deployment choices. Small-big-small training workflow (Priority: 4/5): For real-time use cases, Fireworks recommends starting with a small model to validate data quality, scaling to a large model for best performance, then distilling back down for low-latency inference. Agentic AI and changing work (Priority: 4/5): Lin connects autonomous intelligence to the rise of agents in hiring, marketing, customer service, and coding, noting that coding agents already resemble junior engineers. Evaluation and standardization (Priority: 4/5): Because evals and tuning are tightly coupled yet fragmented across platforms, Fireworks open-sourced Eval Protocol to standardize interoperability between evaluation and tuning systems. Hiring and company growth (Priority: 3/5): Fireworks is hiring broadly after its Series C, reflecting strong demand for its platform across engineering, GTM, finance, and operations.

Key Arguments: Most of the world’s intelligence sits in private enterprise data that foundation models have never seen, so enterprise AI must learn continuously from proprietary data. AGI is valuable, but the bigger practical opportunity is autonomous intelligence: continuously customizing models and inference stacks for each application. Model selection is necessary but exhausting because frontier models change rapidly and specialize differently across chat, coding, multimodal, and long-context tasks. Hardware choice is also becoming difficult because GPU/ASIC generations are evolving quickly, making infrastructure abstraction critical. Open models are increasingly viable because they are cheaper, more controllable, and now approaching closed-model performance. For real-time systems, low latency matters more than extended reasoning; for offline research/agent workflows, slower, deeper thinking is often appropriate. Evaluation is currently too fragmented and subjective, so interoperable standards are needed to connect eval systems with tuning systems. Enterprise AI adoption will accelerate as organizations learn the bounds of open-model licensing, privacy, and deployment options.

Data Points: Private data share: Over 90% - Lin’s claim that the majority of the world’s intelligence is locked in private enterprise data unavailable to foundation models. Fireworks funding: Over $300 million - The company has raised this amount in venture capital to build its AI infrastructure platform. Recent financing round: $250 million Series C - John references Fireworks’ recent raise as part of the company’s rapid growth. Engineering leadership at Meta: 300+ engineers - Lin previously led a team of more than 300 engineers at Meta. Latency target: Sub-2 second / sub-500 millisecond - Referenced as examples of Fireworks’ fast real-world inference performance. Optimization search space: 100,000+ options - Lin describes 3D Fire Optimizer as searching across many combinations of model quality, speed, and cost settings. Hardware refresh cadence: Every couple of weeks / three years previously - Lin contrasts the rapid pace of model and hardware changes with the older, slower hardware cycle. 2026 model trend: Convergence expected to become more prominent - Lin predicts open and closed model performance will continue to converge in 2026.

Pivotal Quotes: "Over 90% of the world's intelligence is locked inside private enterprise data that no foundation model has ever seen." — John Crohn (intro framing the episode): Sets up the central thesis for why enterprise data and customization matter. "The future is not one model rules all. It's going to be millions of models, one per application, per use case." — Lin Chow: Defines the autonomous intelligence vision and the move toward specialization. "We're like, hey, this is actually not a linear problem. It's a complex problem. It's with exponential search space... there are more than 100,000 options in this search space." — Lin Chow: Explains why model optimization requires automated search across many dimensions.

Implications: Enterprise AI will likely shift from generic copilots to continuously learning, highly specialized systems. Companies that master private data, model tuning, and deployment optimization will gain durable moats and stronger performance.