20VC: Are OpenAI and Anthropic Overvalued? The Open-Source AI...
The Twenty Minute VC (20VC)Full Title
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
This episode features Lin Qiao, CEO of Fireworks AI, discussing the future of specialized AI models versus generalized AGI, the decreasing cost of AI inference, and the critical importance of owning proprietary intelligence for businesses. The conversation highlights the potential for 10x cost reduction and 100x usage growth in AI, driven by open-source models and specialized AI solutions.
Key Points
- The future of AI is not one dominant AGI, but millions of specialized models catering to unique use cases and individual business needs, contrasting with the AGI-centric view of companies like OpenAI and Anthropic.
- The cost of AI inference is predicted to decrease by 10x in the next three years, which will in turn drive a 100x increase in AI usage, indicating a massive growth phase for the AI economy.
- Open-source AI models are becoming increasingly capable, offering a more cost-effective and customizable alternative to proprietary models, and empowering users with greater control over weights and modifications.
- Companies need to own their proprietary intelligence to avoid being solely reliant on external providers and to ensure control and customization, especially in specialized domains like legal services.
- The development of AI is rapidly advancing, with a constant stream of new models and capabilities, leading to accelerated specialization and innovation at the application layer.
- Concerns about national security and data sovereignty are rising, particularly with the prominence of Chinese open-source models, emphasizing the need for robust guardrails and potentially sovereign AI solutions.
- The infrastructure layer for AI, including chips, energy, and data centers, is a critical bottleneck, and innovation in system design and deployment is necessary to support the massive scaling demands.
- Businesses are moving towards owning their AI intelligence rather than renting it, a trend driven by the need for optimization, customization, and control over their proprietary data and workflows.
- The rapid growth of AI companies is unparalleled, reflecting a fundamental disruption and an unlocking of creativity across industries, leading to significant revenue growth.
- The ideal candidate for the AI space possesses contradictory traits, such as deep experience combined with extreme curiosity, and demonstrates unwavering ownership and accountability.
Conclusion
The future of AI is decentralized, with specialized models outperforming generalized AGI for specific tasks, driving immense growth in usage and innovation.
Companies must prioritize owning their proprietary intelligence and leverage open-source solutions to ensure control, customization, and long-term viability in the rapidly evolving AI landscape.
Continued investment in AI infrastructure and a focus on ROI will be critical as the technology matures and becomes more deeply integrated into business operations.
Discussion Topics
- How will the increasing capability and decreasing cost of specialized AI models impact the dominance of large AGI providers?
- What are the most significant infrastructural bottlenecks for AI, and how can innovation in chips, data centers, and energy address them?
- As AI becomes more accessible and cost-effective, what ethical considerations and regulatory frameworks need to be established to ensure responsible development and deployment?
Key Terms
- AGI
- Artificial General Intelligence, a hypothetical type of AI that possesses the ability to understand, learn, and apply knowledge across a wide range of tasks at a human-like level.
- ARR
- Annual Recurring Revenue, a metric used by subscription-based businesses to track the predictable revenue a company expects to receive from its customers over a year.
- ASIC
- Application-Specific Integrated Circuit, a microchip designed for a particular application or purpose, unlike general-purpose processors.
- Infrence
- In the context of AI, inference refers to the process of using a trained machine learning model to make predictions or decisions on new, unseen data.
- KLD
- Kullback-Leibler Divergence, a measure used to quantify how one probability distribution diverges from a second, expected probability distribution. In AI, it can be used to measure the difference between a model's training and inference distributions, indicating quality.
- SaaS
- Software as a Service, a software distribution model where a third-party provider hosts applications and makes them available to customers over the Internet.
- Token
- In natural language processing, a token is a basic unit of text, such as a word or sub-word, that a language model processes.
Timeline
The core argument for specialized AI models is that a significant portion of world data is privately locked within enterprises and not used for training general intelligence models.
The speaker argues that while "power lines" (infrastructure providers like OpenAI) are crucial, they may not be good businesses in a world increasingly dominated by open-source alternatives.
The speaker believes that companies will need to own their own intelligence, analogous to how companies build their own software stacks, to maintain control and drive innovation.
The speaker's fundamental belief is that a future dominated by a single company owning all intelligence is undesirable and would lead to a loss of human creativity and diversity.
The high cost of AI features is a barrier for incumbents and startups alike, driving the need for more affordable alternatives like open-source models.
The benefits of open models include full control over weights, customizability, and the ability to tune them for specific use cases, which is crucial for specialized applications.
The prevalence of Chinese open-source models raises national security concerns, highlighting the need for transparency and control over AI development.
The bottleneck for AI is not just individual chips but the entire system, requiring smart engineering co-design from the model down to the chip layer.
The speaker shares a lesson from working with Jensen Huang, emphasizing the importance of deep context and precise judgment for effective leadership in a fast-paced environment.
The area of AI that is underinvested in is the monitoring of Return on Investment (ROI), attribution, and the cost-effectiveness of AI solutions, shifting focus from token maximization to business outcomes.
The speaker's clear vision for the next three years is that every company will own its own intelligence as a mandatory requirement for control and innovation.
Token costs are predicted to decrease drastically due to increased competition and supply chain improvements, leading to a 10x cost reduction and a 100x increase in usage over three years.
Open-source models are becoming nearly as efficient as proprietary models but at a significantly lower cost, driving adoption and innovation.
The SaaS era's product-market fit and durable business were closely linked, but in AI, these are becoming separate concepts due to the high cost of inference.
The trend is for companies to move towards owning their intelligence rather than renting it, driven by the need for optimization and control.
The possibility of sovereign models and national ownership of AI infrastructure is emerging due to concerns about external control and potential disruptions.
Building custom hardware like chips or specialized data centers becomes economically sensible when usage passes a certain threshold and workloads stabilize.
Fireworks AI has a diversified customer base beyond coding, now encompassing various co-work spaces like legal, finance, and customer support, as well as consumer-facing companies.
A significant bottleneck in AI is the lack of systems designed for very large models (e.g., 10 trillion parameters), requiring co-design across the model, serving platform, and chip layers.
The key to scaling rapidly in Fireworks AI is hiring individuals with extreme ownership, a strong sense of responsibility, and the ability to drive projects end-to-end.
Marketing is not just about publicity but about education and clarity, helping customers understand trends and the value of specialized AI.
Jensen Huang stated that every company is unique and built on a special belief, justifying their existence and embedding this uniqueness into their product and system design.
The decision to hire George Hugh, former president of Salesforce, was timed when Fireworks AI was large enough to benefit from his expertise in scaling a business, not just a product.
Jensen Huang's observation about specialized general companies highlights that each company's unique approach justifies its existence.
Fireworks AI prioritizes model quality for specific applications and business use cases, differentiating them from cheaper, commoditized providers.
Fireworks AI processes over 40 trillion tokens daily, with the majority coming from customized models, not off-the-shelf solutions.
The core argument for specialized AI models is that a significant portion of world data is privately locked within enterprises and not used for training general intelligence models.
The speaker argues that while "power lines" (infrastructure providers like OpenAI) are crucial, they may not be good businesses in a world increasingly dominated by open-source alternatives.
The speaker believes that companies will need to own their own intelligence, analogous to how companies build their own software stacks, to maintain control and drive innovation.
The speaker's fundamental belief is that a future dominated by a single company owning all intelligence is undesirable and would lead to a loss of human creativity and diversity.
The high cost of AI features is a barrier for incumbents and startups alike, driving the need for more affordable alternatives like open-source models.
The benefits of open models include full control over weights, customizability, and the ability to tune them for specific use cases, which is crucial for specialized applications.
The prevalence of Chinese open-source models raises national security concerns, highlighting the need for transparency and control over AI development.
The bottleneck for AI is not just individual chips but the entire system, requiring smart engineering co-design from the model down to the chip layer.
The speaker shares a lesson from working with Jensen Huang, emphasizing the importance of deep context and precise judgment for effective leadership in a fast-paced environment.
The area of AI that is underinvested in is the monitoring of Return on Investment (ROI), attribution, and the cost-effectiveness of AI solutions, shifting focus from token maximization to business outcomes.
The speaker's clear vision for the next three years is that every company will own its own intelligence as a mandatory requirement for control and innovation.
Token costs are predicted to decrease drastically due to increased competition and supply chain improvements, leading to a 10x cost reduction and a 100x increase in usage over three years.
Open-source models are becoming nearly as efficient as proprietary models but at a significantly lower cost, driving adoption and innovation.
The SaaS era's product-market fit and durable business were closely linked, but in AI, these are becoming separate concepts due to the high cost of inference.
The trend is for companies to move towards owning their intelligence rather than renting it, driven by the need for optimization, customization, and control over their proprietary data and workflows.
The possibility of sovereign models and national ownership of AI infrastructure is emerging due to concerns about external control and potential disruptions.
Building custom hardware like chips or specialized data centers becomes economically sensible when usage passes a certain threshold and workloads stabilize.
Fireworks AI has a diversified customer base beyond coding, now encompassing various co-work spaces like legal, finance, and customer support, as well as consumer-facing companies.
A significant bottleneck in AI is the lack of systems designed for very large models (e.g., 10 trillion parameters), requiring co-design across the model, serving platform, and chip layers.
The key to scaling rapidly in Fireworks AI is hiring individuals with extreme ownership, a strong sense of responsibility, and the ability to drive projects end-to-end.
Marketing is not just about publicity but about education and clarity, helping customers understand trends and the value of specialized AI.
Jensen Huang stated that every company is unique and built on a special belief, justifying their existence and embedding this uniqueness into their product and system design.
The decision to hire George Hugh, former president of Salesforce, was timed when Fireworks AI was large enough to benefit from his expertise in scaling a business, not just a product.
Jensen Huang's observation about specialized general companies highlights that each company's unique approach justifies its existence.
Fireworks AI prioritizes model quality for specific applications and business use cases, differentiating them from cheaper, commoditized providers.
Fireworks AI processes over 40 trillion tokens daily, with the majority coming from customized models, not off-the-shelf solutions.
Episode Details
- Podcast
- The Twenty Minute VC (20VC)
- Episode
- 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
- Official Link
- https://www.thetwentyminutevc.com/
- Published
- July 20, 2026