20VC: Are OpenAI and Anthropic Overvalued? The Open-Source AI Reality | How Token Costs Will Fall 10x And Usage Will Explode 100x | The Future Is Not One AGI; It's Millions of Specialised Models with Lin Qiao, Founder and CEO @ Fireworks

Lin Qiao founded Fireworks at age 48 after spending seven years at Facebook learning how to build companies at scale. The company has scaled to $800 million in ARR in four years by focusing on inference optimization and specialized model deployment rather than training general models.

Qiao's central thesis challenges the AGI narrative promoted by OpenAI and Anthropic. He argues that the future belongs to millions of specialized models tailored to specific use cases, not one dominant general intelligence. This belief stems from his observation that 90% of world data is private and locked within enterprises—data that cannot be used to train public models but could drive tremendous value if activated through specialized intelligence.

On the open source question, Qiao explains why Fireworks bet on open models early despite their infancy. Open models provide users full control over weights, enabling customization with proprietary data. While frontier models and open models have both crossed quality thresholds, open models are easier to tune with small amounts of unique company data. He notes that Fireworks processes 40 trillion tokens daily, mostly from customized models, not off-the-shelf ones.

Qiao predicts a fundamental shift in enterprise AI adoption. During the SaaS era, product-market fit and durable business were equivalent. Now they're separate—many startups achieve PMF but cannot scale profitably because inference costs would bankrupt them. This creates demand for alternatives: either companies must use cheaper open models they can customize, or frontier labs must drastically reduce prices. The cost reduction Qiao predicts (10x in three years) will drive 100x usage expansion.

On whether frontier model companies are overvalued, Qiao likens them to power lines distributing intelligence infrastructure rather than replacements for everything companies do. He views Anthropic's enterprise strategy as validation that specialized intelligence matters, but notes Anthropic's AGI belief means they see no need for specialization—a fundamental philosophical difference.

Regarding Fireworks' positioning between chips and applications, Qiao emphasizes the company specializes in what it adds most value to: the customization and optimization layer. The company achieves 30-40% gross margins during hypergrowth because it prioritizes innovation velocity over margin optimization. They deliberately avoid the application layer but could eventually move into data centers if timing makes sense.

On the data center and chip questions, Qiao explains that chip building makes sense only after workloads stabilize. The AI industry's workloads remain too dynamic to justify custom silicon—hardware launches every year with three SKUs, models launch weekly, and depreciation cycles are mismatched. This cascading funnel of stability works downward: once application workloads mature, then infrastructure patterns, then data center design, then chips.

Qiao hired George Hu, former Salesforce president, after initially telling him Fireworks was too small. Only after the company grew to 200 people and achieved very high growth rates did the timing feel right. Qiao emphasizes hiring people with extreme ownership mentality who claim end-to-end problems rather than waiting for delegation.

On national security and sovereign models, Qiao believes every country should own their own power line of intelligence, similar to electricity infrastructure. He sees this not as protectionism but as essential independence—companies and nations shouldn't risk being cut off by a single provider.

Qiao's prediction for the next three years: every company will own their own intelligence as a must-have, not optional. This mirrors how every company owns its software stack, picking which parts to build versus which to use from common infrastructure. The difference will be that while some use frontier models, most will customize open models for competitive advantage and cost efficiency.