5YF Episode #30: Fireworks AI CEO Lin Qiao

Future AI: An Open Source Tsunami

subscribe and listen today

Today’s episode comes perfectly timed with the news on DeepSeek and open source AI models. Our guest couldn’t be better placed to comment on AI’s current events and, importantly for our purpose, what it all means for the future — Lin Qiao built open source infrastructure at Meta, leading a team of 300 AI engineers. Her Sequoia-backed startup Fireworks AI, is barely 3 years old and already valued at $550M thanks to having built infrastructure to deploy over 100 AI models. Fireworks AI was one of the first to make DeepSeek deployable for US enterprises.

This current AI wave is unprecedented. It’s a shockwave across the whole entire industry.

Join us as we dive into the future of AI models and infrastructure, fast forwarding this battle for AI supremacy.

My 5 Year Outlook:

• Open Source Tsunami Comes For AI: DeepSeek has (briefly) put Open Source ahead of proprietary AI performance. This is just the beginning of the community movement that will see it win the model wars.‍
• Shift To Small Expert Models As LLMs Commoditize: Low cost, expert small models unlock the important last mile of value in enterprises and consumer.‍
• Enterprises Transform Their Private Data In AI Assets: Private data eclipses public internet data, and it will steadily processed to be high value AI assets.

Curious? Read on as I unpack each below 👇🏼

Open Source Tsunami Comes For AI

The collective effort we put in together can really beat proprietary (AI models).

Over the next five years, open source AI will outcompete proprietary models, reshaping how AI is built and deployed. The rise of DeepSeek and LLaMA proves open models are no longer mere alternatives—they’re now outperforming closed systems in logical reasoning, adaptability, and efficiency. For instance, DeepSeek already beats GPT-4 on reasoning benchmarks. While OpenAI’s latest reasoning model, released this week, set new records for reasoning, it’s an endurance race: given enough time, the collective innovation of open source will overrun elite labs.

The open source community is global, moves fast, shares best practices, and iterates in public. GitHub reported over 10,000 new AI code commits in January alone—a pace that proprietary labs can’t match. The industry is shifting toward an open, modular ecosystem much like Linux did for operating systems or Kubernetes for cloud computing. In five years, the biggest breakthroughs won’t come from closed labs but from the collective intelligence of the open-source movement.

Enterprise adoption is set to drive this trend. Facing high API fees and restrictive models from providers like OpenAI and Anthropic, companies are increasingly turning to self-hosted AI solutions for full customization and control. A recent survey of Fortune 500 companies revealed a 20% surge in self-hosted AI deployments in the past month, signaling that as open models improve, enterprise demand will only accelerate the shift.

Lin Qiao, CEO of Fireworks AI

Fireworks AI, a company pioneering scalable AI infrastructure to help businesses build, customize, and deploy AI models and applications with speed and efficiency. Fireworks provides seamless access to over 100 AI models, including those from OpenAI, DeepSeek, and other leading providers, making enterprise AI adoption faster and more flexible. Founded in 2022, Fireworks AI has already gained high-profile customers such as DoorDash, Verizon, and Upwork. Backed by Sequoia, the company has raised $77 million to date and is currently valued at $550 million.

CEO Lin Qiao brings a wealth of experience in AI infrastructure and engineering leadership. She has held key technical roles at IBM and LinkedIn, but is best known for her tenure at Meta, where she led a team of 300+ world-class engineers in AI frameworks and platforms. She played a pivotal role in scaling PyTorch, deploying it across Facebook’s global data centers and billions of devices, cementing it as one of the most widely used open-source AI frameworks today. Lin holds a Ph.D. in Computer Science from UC Santa Barbara and a Master’s in Computer Science from Fudan University in China.

Shift To Small Expert Models As LLMs Commoditize

The next major shift in AI is evolving from monolithic, general-purpose models toward small, specialized expert models. Enterprises don’t need a single massive AI that performs everything at a mediocre level—they require highly optimized models that excel in specific domains like healthcare, law, and finance. Large general-purpose models and small expert models are intertwined but to date the focus has been on developing large scale LLMs, going forward there will be shift towards the proliferation of light weight specialized models.

My prediction is simple: the future is hundreds of small expert models… fine-tuned for specific tasks, and outperforming general-purpose models.

This evolution mirrors the software shift from monolithic architectures to microservices, where agility, speed, and efficiency became paramount. Open-weight models like DeepSeek and fine-tuned variations of LLaMA are already demonstrating superior performance in niche applications while significantly lowering inference costs.

Ultimately, it all comes down to the unit economics of individual AI runs—critical for interactive applications where even a slight increase in latency can have a huge impact. For example, a mobile banking app with an AI fraud detection system that adds 100ms per call can quickly slow down during peak times, frustrating users and driving up server costs. A streamlined, small expert model cutting latency by 20% can dramatically enhance performance and reduce expenses.

This trend is also evident in the race for on-device AI. About a year ago, there was immense excitement over on-device models—those designed to run directly on devices rather than in the cloud. Apple led the charge by launching an on-device model for iPhones in June, shortly after Google’s Gemma and Microsoft’s Phi-2 small models. While the hype around on-device AI subsided as large language models like GPT-4 began dominating cloud-based reasoning tasks, DeepSeek’s emergence has reinvigorated interest in smaller, cost-effective models that can operate efficiently on limited hardware.

This is particularly relevant for Apple, which has long championed on-device processing for privacy and performance. By leveraging small expert models, Apple could offer more powerful, tailored AI experiences on its devices without relying solely on cloud infrastructure—further differentiating its ecosystem from traditional LLM-dependent platforms.

Zyphra, an open-source small model developer that is in talks to raise $100 million at a $1 billion valuation, say on-device models are built to work with limited storage and processing power compared to models running in the cloud. They are trained on fewer parameters and are cheaper to run than larger models. As Steve Jurvetson, founder of Future Ventures and investor in Zyphra, explains, “On-edge developers’ constraints are very different. Your entire development pipeline looks different, and the products are different.”

Together, these shifts point to a future where hundreds of small expert models power a diverse array of applications, fundamentally transforming how AI is integrated into everyday technology and business processes.

‍

Enterprises Transform Their Private Data In AI Assets

(Enterprises) need to convert data assets into AI assets. And we have seen the biggest gap of doing that is a lack of proper customization process…DeepSeek models makes that transformation much easier.

The biggest untapped opportunity in AI is converting vast, underutilized enterprise data into actionable intelligence. Every major company sits on enormous datasets—from customer interactions to supply chain logistics—but much of this data remains unstructured and siloed. Over the next five years, businesses will shift from simply storing data to activating it through AI. Recent surveys show nearly 70% of enterprise executives view leveraging proprietary data as a key differentiator by 2025, echoing the transformative impact of the cloud computing revolution.

The transformation is already underway. For example:

• Walmart has integrated AI-driven forecasting to optimize inventory management, boosting efficiency by around 20%.

• UPS uses machine learning for route optimization, achieving significant fuel savings and reduced delivery times.

• GE leverages sensor data for predictive maintenance, cutting maintenance costs by up to 30%.

Automation and seamless integration are critical to this change, enabling companies to fine-tune AI models without extensive in-house expertise. Emerging platforms like DeepSeek are lowering barriers for data infusion, bridging the gap between general-purpose models and company-specific applications.

As Lin emphasizes, enterprises must adapt their models to internal data to fully capitalize on this opportunity. The winners will be those who can automate the conversion of raw data into dynamic, self-improving AI systems, fundamentally redefining competitive advantage.

Time to turn data into intelligence. Happy building!

‍