Check out our latest blog article: From component to enterprise – modular robotics done right.
Why 85% of Robot Infrastructure Costs Aren't the Robot: Dan Cummins on Physical AI
Dan Cummins is a Dell Technologies Fellow whose career spans embedded defense systems, enterprise storage architecture, and edge AI. He began his career building sonar, radar, and surveillance systems for the U.S. Navy, including a fleet-wide submarine noise-leak detection system. At EMC and later Dell, he led major storage transformations including VNX2 and the ground-up reinvention of PowerStore. He now leads Dell's enterprise edge and physical AI strategy, having helped build NativeEdge. Dan also serves on the board of the FIDO Alliance and as an advisory board member for the University of New Hampshire's computer science program. In this episode, he shares how his systems-engineering background shapes his view on where physical AI is headed next.
● https://www.linkedin.com/in/danielcummins603/
● https://www.linkedin.com/in/techris/
● https://www.linkedin.com/showcase/forging-the-future-with-chris-howard/
● https://www.softeq.com/forging-the-future-podcast
Stay inspired and ahead of the curve by subscribing to Forging the Future. Share your thoughts on this episode with the hashtag #ForgingTheFuture or tag us online!
🎧 Episode Highlights
- [00:02:00] Dan's origin story: building embedded warfare, sonar, and radar systems for the Department of Defense before college.
- [00:03:00] The submarine noise-leak detection system Dan built and deployed across the entire fleet; an early lesson in end-to-end systems thinking.
- [00:06:00] Joining EMC to modernize the aging, single-threaded Clarion storage stack for the flash era.
- [00:07:00] The VNX2 project: taking a 16-core architecture from 200,000 IOPS to over a million IOPS in under a millisecond.
- [00:13:40] The pivot to Dell and Edge AI — solving for the fragmentation of solutions.
- [00:24:00] Where edge deployments actually fail: Deploying physical AI at enterprise scale.
- [00:28:00] The physical AI data problem: why 90% of robot trajectory and experience data is mishandled today, and the opportunity for a purpose-built data platform.
- [00:38:50] The generational gap: why systems-design experience, not just AI fluency, is becoming a critical and increasingly rare skill.
- [00:44:00] Dan's five-year outlook: physical AI, sovereign AI, and customer-owned training loops.
🔑 Key Takeaways:
- Takeaway 1: The hardest part of physical AI isn't the model or the pilot, it's deploying and managing it at enterprise scale. Most physical AI today remains siloed and piloted because organizations lack the orchestration platforms needed to provision and manage fleets of endpoints consistently.
- Takeaway 2: Experience data generated by physical AI (robot trajectories, sensor data, etc.) is quickly becoming valuable intellectual property — and today, roughly 90% of it is mishandled, siloed, or lost, making replay and root-cause analysis nearly impossible when something goes wrong.
- Takeaway 3: Systems-level thinking has to come first. AI coding tools are only as good as the design requirements you give them. Non-functional requirements (like data reduction or latency targets) that aren't explicitly written into a spec simply won't be honored by an AI agent.
You may be interested in:
