Why speed is the AI feature no one asks for but everyone needs

Rob Ferguson is the VP of Technology and Strategy at Fireworks AI, a company that makes frontier AI infrastructure accessible to any organization that wants to own its own models rather than depend on someone else’s. He started studying machine learning before it was called AI, worked on Pro Tools and AI systems for a music streaming service, helped design Amazon Go’s grab-and-go technology, and built startup programs at AWS and Microsoft. Now he is at what he calls the front row of AI, sitting between applied research and go-to-market and translating what the technology can actually do into things companies can actually buy and build on. In this episode of Product Talk, iDonate VP of Product and Engineering Nacho Andrade sits down with Rob to talk about why speed is the AI feature no one asks for but everyone needs once they see it, how AI training went from lab-only to cheap as a cup of coffee, and his hot take that we are at the end of SaaS and entering an era he calls workwear.

Subscribe to the Product Talk podcast on Spotify and Apple Podcasts and catch every conversation with leading product executives. Episodes drop on Wednesdays.

Show Notes

1. Rob started studying machine learning at a time when it was considered a dead end in academia and the least cool thing a CS student could pursue. His interest came from a musician’s perspective: as a pianist, he was always thinking about tools that help other people create, and he wanted to understand what human-computer interaction would look like when computers could actually listen and respond.
2. His career has been a long experiment in where science meets product: building AI-driven recommendations and search for a music streaming service when no one was doing AI at scale, designing the computer vision systems behind Amazon Go, and building a car adapter that used AI on driving data to predict battery failure and qualify drivers for insurance discounts. Each project was a different answer to the same question of how this technology interfaces with the real world.
3. Rob’s role at Fireworks sits between the applied research organization and the go-to-market organization. His job is to translate what the research team has just figured out how to do into language and positioning that the market can understand and that customers can actually evaluate and buy. In a hard tech company, that translation work is itself a product function.
4. Fireworks AI has two core capabilities. The first is a curated network of the world’s best open-weight models, distributed across more than 20 clouds globally, packaged into a single seamless API that is drop-in compatible with Anthropic and OpenAI workloads and runs significantly faster and cheaper. The second is model training infrastructure that makes it possible for any company to fine-tune or train its own AI model on its own data and then run that model on the same network.
5. AI training used to require hundreds of millions of dollars and was thought of as pure research. That assumption has fundamentally changed. Fireworks can now help a company start training its own AI model for as little as the cost of a cup of coffee, with a realistic entry point around $150 for initial inference. The technology that was lab-only two years ago is now accessible to any company willing to learn how to use it.
6. Speed is the most underrated feature in AI and the one that most consistently produces an aha moment for customers. Nobody ever tells Rob that speed is their primary concern with their existing Anthropic or OpenAI workload. But when he puts a demo in front of them that responds as fast as they can type, the reaction is immediate: this would literally change how I work. Speed is not just a performance metric; it is a product unlock that makes entirely new use cases viable.
7. The shift from AI as talkers to AI as doers is the most important structural change in how companies need to think about this technology. A few years ago, AI was a one-to-one exchange: one message in, one response out. Now a single prompt can trigger 40 or 50 downstream actions. That changes what AI is, what it costs to run, and what it means to productize it.
8. Rob distinguishes between AI dependence and AI independence. AI-dependent companies use AI through other labs’ models without understanding how those models work or how to shape them. AI-independent companies, like Cursor, Harvey, and Notion, have done the work to understand the grain of the technology: where it fails, how to frame questions so it can give reliable answers, and how to train it on their own domain. That independence is what allows them to build products that others cannot easily copy.
9. The most common problem Rob sees with startups is agents that work but cost a fortune to run. A founder comes in having burned through every cloud credit from AWS, Google, Microsoft, and OpenAI, with something that functions but is not yet a business. The process of moving from that state to something economically viable is also the process of actually learning how AI works: understanding your data, asking the right questions, and getting the system into the right shape.
10. One founder told Rob that his company’s monthly AI bill was six figures. After switching to Fireworks and going through the process of understanding how to properly structure his data and workloads, the bill dropped to sub-four figures. His first reaction was to wonder if he had broken something. He had not; he had just finally learned how to build with AI reliably for the first time.
11. The key product discipline in AI right now is what Rob calls drawing the line in the sand: deciding which features actually need to ship based on real economic value, rather than promising everything AI could theoretically do. The companies that try to offer every possible feature at launch find that customers use a small number of them, then defect to a competitor who offers only that one feature at lower cost. Doing less, deeply, is the winning move.
12. Rob’s hot take is that we are at the end of SaaS. Software went through two eras: packaged software sold once on disc, and software as a service sold continuously with a direct customer relationship. He believes we are entering a third era he calls workwear, where the value is no longer at the application surface but in the relationship between the technology and the actual work being done, the domain knowledge, the data, the workflows, and the context specific to how a person or organization does their job.
13. The workwear era changes the competitive dynamics of software fundamentally. If someone ships a new feature on Fireworks, a competitor can clone the interface within 48 hours without understanding why it was built or who it was built for. The surface is no longer the moat. The moat is deep understanding of a specific work domain and the ability to make AI work reliably within it.
14. As the surface level of software commoditizes, domain expertise becomes significantly more valuable. People who deeply understand a specific field, whether through liberal arts training, years of professional experience, or serious study of a particular problem space, are better positioned to build the why of a product than people who are skilled only at the craft of assembling user interfaces. Rob expects to see more people with non-technical backgrounds building important things.
15. The biggest thing the AI industry gets wrong about itself is the assumption that the people who built the models have a grand design for how they should be used and governed. In his experience, the field is learning as it goes. The claim that any single lab should be trusted to steward the technology on behalf of society, without external scrutiny or competition from open alternatives, is one he pushes back on directly.
16. On the open versus closed debate, Rob thinks the framing is often wrong. The competitive advantage people assumed would come from keeping models closed turned out to be different than expected, because no one fully predicted what these models would be capable of or where they would fail. Humility about what the technology is and openness to outside scrutiny are more important than controlling the entire stack.
17. Rob is a strong advocate for people who ask hard questions about AI: ethicists, researchers studying how models work from the outside, people thinking about consent and data rights, and people who study the actual effects of these systems on real users. He sees a major unmet need for that kind of thoughtful scrutiny, as distinct from valley rhetoric.
18. The pandemic created a lasting dehumanization effect that the arrival of AI has accelerated. Rob is candid that the expectation of how much output a person should produce has ballooned in ways that feel unsustainable. His concern is not just about what AI can do but about whether we are thinking clearly about how we want to work with each other and with the technology, rather than simply racing to fill every available space with output.
19. Rob’s career operating principle comes from his father: always work yourself out of a job. The logic is that if you make every part of your work repeatable and teach others to do it, you will always have one, because you will accumulate a network of people who trust you and want to work with you again. He credits this mindset with the professional relationships that have sustained his career across multiple startups and companies.
20. The technology Rob is most obsessed with right now is voice. He uses Whisper Flow for transcription, Granola for meeting notes, and a hardware device called Implod to record his thoughts on his commute. He then has an AI agent structure those thoughts and do deep research overnight, and on the drive to work he listens to a synthesized two-person podcast via ElevenLabs summarizing what it found. For someone who thinks out loud, the ability to process ideas verbally and have them come back structured and researched has been, in his words, a total game changer.