Source: SPEEDA Edge research

Background

Questions

1. What does Outerport do?
2. Could you provide more detail about the cold start problem?
3. Are there specific types of architecture, companies, or use cases where the cold start problem is particularly acute?
4. What does it look like to solve the cold start problem both before and after Outerport?
5. Can you talk about your early customers and where your message has resonated the most?
6. Are customers excited about Outerport because it saves them money, or is it more about managing complexity and improving the developer experience?
7. From a developer experience standpoint, when building these tools and testing them on a local machine, the processing time can be extremely long. Is the developer experience an important consideration in this context?
8. Why do people use custom models?
9. Can you discuss custom models in terms of security, privacy, and on-premises deployment? Is Outerport part of that ecosystem?
10. One of the things you mentioned is that this kind of problem might not be as acute for SaaS companies using AI. At Ramp, they use multiple models - some smaller local models, custom models, and GPT-4 for specific problems. They optimize based on latency, cost, quality of output, and the type of output they want to generate. Is this kind of setup well suited for Outerport, or is it just that Ramp is more sophisticated than your average SaaS company using AI? Is this a different class of problem altogether?
11. To the extent that this is a problem that Outerport is solving, why did you choose to address this particular challenge among many others? What does solving this problem enable you to accomplish for your customers?
12. Why is this a good wedge problem to solve in terms of helping solve other MLOps-type problems?
13. Could you map out how you see the MLOps landscape?
14. For companies that aren't doing any training and just need to deploy and fine-tune models, there are tools like Vellum. Do you see this market segmenting based on how extensively companies apply AI? Is this distinction meaningful, or will these different approaches eventually diverge or converge?
15. How have you seen the tooling shift as part of that change?
16. When we talk about deployment and this DevOps-MLOps analogy, do you think the comparison is particularly accurate? Will we see the same categories and setup, such as version control, continuous deployment, containerization, orchestration, iPaaS, and PaaS? Or is it more of a superficial analogy?
17. Is Outerport's expertise primarily focused on optimizing CPU/GPU performance for large files? And does that optimization help solve a horizontal range of deployment problems?
18. When considering your target customer, are you aiming to sell to DevOps professionals or individual engineers? In terms of go-to-market strategy, are you pursuing a self-serve model where engineers might start using it on their personal machines before bringing it into their companies? Or are you focusing on top-down enterprise sales, since larger companies currently show the most interest in this technology?
19. Are the major AI companies like OpenAI and Anthropic, along with cloud providers such as AWS and Azure, potential customers for Outerport? Are they already implementing similar systems internally to serve their APIs?
20. Is there a competing vision? There’s one perspective that suggests there will be a single model to rule them all, where the trajectory is that this current solution is nice but temporary—more of a bridge solution rather than something of long-term importance. The alternative view is that multiple models will continue to serve different purposes.
21. Looking in the opposite direction, is there a scenario where microservices-style orchestration becomes extremely granular, with countless small custom models? In this case, would the models become so small that the size problem essentially disappears?
22. Are new chips like those from Cerebras—which aren't GPUs at all—potentially an existential threat to what you're doing? How do you think about that?
23. Can you discuss the current state of open source models? Is this development fundamentally enabling the open source model ecosystem? What does the development stack look like for open source models, and are you optimistic or pessimistic about the ecosystem's future?
24. Could you talk about what a gold standard ML ops pipeline looks like? Since you work with many companies, what does it look like when you see a company implementing ML ops correctly?
25. Is making Outerport an open source project important for becoming a standard and winning hearts and minds?
26. Can you discuss the developer experience and deployment layers, particularly the MLOps infrastructure that's emerging from cloud GPU companies? For example, companies like Vercel partnering with CoreWeave, or similar initiatives from AWS and other major cloud providers?
27. If everything goes right for Outerport over the next 5 years, what does it become, and how has the world changed as a result?