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

Summary

Lin Qiao makes the case that AI’s future is a specialized, multi-model ecosystem rather than a world dominated by one or two AGI providers. She argues most valuable intelligence comes from private enterprise data, so companies will increasingly want to own and control their own models, workflows, and routing layers. Fireworks’ strategy is to stay focused on the inference/specialized-intelligence layer, optimize each deployment for workload-specific quality and cost, and leverage open-source models because they offer weight-level control and easier customization. Qiao says token costs should fall about 10x over three years, which she believes would unlock roughly 100x usage growth, while also warning that many AI startups can “scale into bankruptcy” if inference economics are ignored. She also dives deep into infrastructure constraints—data centers, chips, cooling, power, and manufacturing throughput—arguing those lower-stack bottlenecks, not just models, will shape the next phase of the AI economy.

Notes

• Fireworks is positioned as a specialized intelligence and inference company, not an application-layer vendor.
• Lin Qiao forecasts a 10x token-cost decline over three years and 100x usage growth.
• She argues intelligence is a derivative of data, and most important data lives in private enterprise systems.
• The company’s thesis is that specialized intelligence should activate private data, not rely only on frontier pretraining.
• Open-source models are attractive because released weights can be fully controlled and modified.
• She says open and closed models have both crossed a quality threshold that makes them broadly useful.
• Fireworks argues that AI startups can have product-market fit and still fail if inference costs swamp growth.
• Cost reduction matters most at scale, where even small improvements can be huge.
• The company sees the future as millions of specialized models, not a few dominant AGI systems.
• Fireworks expects routing, sub-agents, and multi-model orchestration to become core AI infrastructure.
• She describes a distributed RL setup built with Cursor that spans five or six regions and scattered GPUs.
• Fireworks says it processes more than 40 trillion tokens per day, mostly from customized models.
• The lower AI stack is constrained by energy, chips, and physical manufacturing throughput.
• Fireworks wants to focus where it adds the most value and run on every possible AI chip.
• Open models are framed as a supply-chain issue for the U.S. if there is no native open ecosystem.
• Price competition should compress token costs as supply constraints ease.
• She gives a concrete forecast of 10x lower cost in three years and 100x usage growth.
• Fireworks prioritizes quality above all else in custom model inference.
• The firm maintains an Applied Machine Learning Engineering team that acts like deployment FDE support.
• Lin says hypergrowth makes gross-margin optimization a tradeoff against innovation.
• Fireworks will not move into application layers, though data centers remain possible.
• Heterogeneous data centers can be powerful, but homogeneous deployments are easier to operate.
• GPU FLOPS are better for prefill, while SRAM-heavy accelerators are better for generation.
• Data centers are hard because they require construction, power, fiber, cooling, and liquid cooling.
• She notes that global supply-chain constraints still affect electrical and construction capacity.
• Hardware depreciation assumptions are changing because model launches and vendor SKUs are moving much faster.
• Fast model and hardware cadence changes the build-versus-buy calculus for infrastructure.
• Fireworks says it grew from about 50 people to 200 in a year.
• The company had around 800 million in ARR and expected to at least double by year-end.
• Qiao believes every company will own its own intelligence as a necessity, not an option.