What if 90% of the world's data is sitting untouched? That's exactly where the real AI frontier lies.
In this episode, Cameron Adams sits down with Lin Qiao, co-founder and CEO of Fireworks AI, to explore why the future of intelligence won't come from a handful of foundation models, but from millions of customized ones built on private data. Lin co-created PyTorch at Meta, before founding Fireworks to tackle what she sees as the next great unlock: activating the private data locked inside every application and enterprise.
Lin walks through the rarely explained difference between training and inference, and why getting inference right is the invisible force behind every AI product people actually use. She makes the case that as applications become commoditized, the only durable moat left is your data and the model you build from it. She explains why Fireworks customers like Cursor, Shopify, and Uber aren't training once and moving on, but running a continuous loop of customization as models and hardware evolve at unprecedented speed.
The conversation turns candid on the real economics of AI at scale: the unit economics gap between GPU and CPU infrastructure, the tension between quality and efficiency in a supply-constrained world, and why passing costs to customers is the least of the problem compared to handing your competitive edge to a black-box API.
Key Quotes
“You can have a great product-market fit, but you cannot scale. You're holding a gate because you know once you scale, you're going to run out of money.”
“ I start to see not just software engineers adopting AI tools, but my head of accounting. She defines skills that her team can repeatedly use.”
“ I do not believe the world will be dominated by a few foundation models. I believe the future will be millions of customized models, one per application per use case.”
“ If you can power 10 times more tokens with the same hardware, that means you can acquire 10 times more customers. If they are not efficient, they are losing market share to their competitors.”
Timestamps
00:00 Introduction
01:41 The Prompt
03:42 Creative All Hands
04:59 Research is about the velocity of creativity. Production is about optimization.
07:59 Fireworks was founded in September 2022 — two months before ChatGPT existed.
12:03 Training is forward and backward passes. Inference is forward only.
17:14 If applications are easy to copy, what's left is the real moat: private data.
26:43 New AI hardware used to launch every 3 years. Now it's every 4-6 months.
32:56 In this dynamic environment - velocity trumps everything.
35:34 All Star Creative Team
37:25 Design Reveal
Links
Subscribe to the Prompted Newsletter: https://promptedwithcam.substack.com/
Cameron on LinkedIn: https://www.linkedin.com/in/themaninblue/
Lin on LinkedIn: https://www.linkedin.com/in/lin-qiao-22248b4/
Fireworks AI:https://fireworks.ai/
The State of Marketing and AI Report 2026:https://www.canva.com/marketing-ai-report/?utm_medium=organic_social&utm_source=google_youtube&utm_campaign=b2b_dg_global-gen_awareness_pro_wbr_prompted-podcast-2026&utm_content=description-link&utm_term=
Hosted by Simplecast, an AdsWizz company. See pcm.adswizz.com for information about our collection and use of personal data for advertising.
