Rob Ferguson is VP of Technology and Strategy at Fireworks AI, the inference and model-customisation platform used by AI-native companies such as Cursor and Notion.
He previously served as Global Head of AI at AWS and CTO of Microsoft for Startups, and has been a startup CTO through three exits. Rob joins Randy Silver to ask what a moat looks like when features can be cloned in days. They cover why competition is a good signal, why data and customer proximity now matter more than shipping speed, and why “done” is the most dangerous word in AI product development.
Key takeaways
- Competition is evidence that a market is worth being in. Rob’s advice is to out-care and out-invest rivals rather than worry about copycats, as he saw at Automatic and Rdio.
- Features alone no longer protect you. A launch that took two years to build was matched to 95 to 97 percent parity within days. What kept the original ahead was its cookbooks, community, clear point of view and understanding of its audience.
- Rising above the noise is now the hard part. Rob’s example is a startup he advises whose cold email open rate has fallen from four percent to a fraction of one. You can no longer be just technically good. You need to be a tastemaker with a point of view.
- The most capable AI systems use a mixture of models, not a single one. Teams that win start by measuring what matters, splitting work into small, verifiable parts and separating “hunch” tasks from deterministic ones.
- Prototypes are now the start of the conversation. Rob’s best designers and PMs arrive with clickable prototypes, but the real work is the 50 or 60 iterations after that, and agreeing what “done” means before you show anything to a customer.
- The second 80 percent of product development is still hard. Replacing a function with AI overlooks the many small jobs inside it. That is why forward-deployed teams, embedded with customers and focused on the work itself, are becoming standard.
- Your data is worth more than the labs will tell you. Rob says Fireworks tried to calculate the discount it would need to offer customers to train on their data and couldn’t make the maths work, because the value was too high.
- Product fundamentals haven’t changed: build, measure, learn. Good data, a good experience, knowing your customer and asking great questions still decide who wins.
- Management teams should rebuild around customer proximity. Long-range plans matter less than teams that engage directly with customers, because a relationship is far harder to copy than a design.
Chapters
- 00:00 Introduction
- 01:55 What a moat actually is
- 03:33 Why customer relationships are the most defensible moat
- 05:40 When features can be copied in days
- 09:30 Distribution, drafting and staying power
- 12:57 Where compute costs are heading
- 17:56 Fireworks and the gold rush analogy
- 20:44 Why the big labs don’t offer model choice
- 26:00 Beyond the chat interface: mixtures of models
- 29:48 Prototypes, vibe coding and the seductive word “done”
- 33:17 The second 80 percent of product development
- 37:09 Is AI producing revolutionary software?
- 40:38 Curation as a core skill
- 42:14 What product teams should do about moats
- 47:48 What management teams should plan for
- 50:11 Wrap-up
We're refreshing The Product Experience and want your input. Take our two-minute survey and help shape where the show goes next!
