Uploaded September 2026 | Updated November 2026, 1 day ago
### What Youβll Learn
Your GPUs are sitting idle, and the network is why. John Ousterhout, Stanford professor and creator of Tcl, explains why TCP, RDMA, and RoCE struggle with AI workloads and what the Homa transport protocol does differently.
On this episode of The Edge Case, recorded at the AI Engineer World's Fair, John walks through how agentic workloads are changing the way AI uses the network: smaller chunks of computation, more coordination, and short messages where tail latency stalls hundreds of GPUs at once. We get into whether Homa belongs at the edge (it doesn't, and he explains why), whether data center communication is bounded by the speed of light, and how much of that 50 to 100 microsecond delay is just software overhead. Then we shift to his other life's work: why AI tools make software design more important rather than less, the Stanford course he built to teach it, and the story of how Tcl went from a personal tool to something the whole industry picked up. He closes with one piece of advice for developers heading into the next few years.
### π Useful Links & Resources:
* Akamai Developers on GitHub: github.com/akamai-developers
* Akamai Developer Hub: developers.akamai.com
* $100 Akamai Cloud Credit: login.linode.com/signup?promo=akm-lp-akmtube-100-111325-T853
### π² Follow Us & Join Our Community
π X: https://x.com/akamaidevs?s=20
πΌ LinkedIn: linkedin.com/showcase/akamai-dev
πΎ Discord: discord.gg/5MJegzS8Y
### β±οΈ Timestamps:
0:00 John Ousterhout and the Homa Transport Protocol
1:18 Why Networks Matter When the Work Happens on the GPU
2:56 Agentic Workloads Are Changing How AI Uses the Network
4:04 Tail Latency: Why TCP, RDMA, and RoCE Struggle With Small Messages
5:00 Idle GPUs: How a 2 Millisecond Message Stalls the Whole Cluster
6:57 Training Workloads vs Agentic Workloads
7:40 Is Homa the Right Protocol for the Edge?
8:12 Are We Bounded by the Speed of Light?
8:58 AI and the Future of Software Developers
10:30 Teaching Software Design at Stanford
12:55 Do You Need to Understand Compute and Networking to Design Software?
14:08 The Origin Story of Tcl
16:14 One Piece of Advice: Stay Broad
#aiinfrastructure #networking #softwaredesign
### What Youβll Learn
Your GPUs are sitting idle, and the network is why. John Ousterhout, Stanford professor and creator of Tcl, explains why TCP, RDMA, and RoCE struggle with AI workloads and what the Homa transport protocol does differently.
On this episode of The Edge Case, recorded at the AI Engineer World's Fair, John walks through how agentic workloads are changing the way AI uses the network: smaller chunks of computation, more coordination, and short messages where tail latency stalls hundreds of GPUs at once. We get into whether Homa belongs at the edge (it doesn't, and he explains why), whether data center communication is bounded by the speed of light, and how much of that 50 to 100 microsecond delay is just software overhead. Then we shift to his other life's work: why AI tools make software design more important rather than less, the Stanford course he built to teach it, and the story of how Tcl went from a personal tool to something the whole industry picked up. He closes with one piece of advice for developers heading into the next few years.
### π Useful Links & Resources:
* Akamai Developers on GitHub: github.com/akamai-developers
* Akamai Developer Hub: developers.akamai.com
* $100 Akamai Cloud Credit: login.linode.com/signup?promo=akm-lp-akmtube-100-111325-T853
### π² Follow Us & Join Our Community
π X: https://x.com/akamaidevs?s=20
πΌ LinkedIn: linkedin.com/showcase/akamai-dev
πΎ Discord: discord.gg/5MJegzS8Y
### β±οΈ Timestamps:
0:00 John Ousterhout and the Homa Transport Protocol
1:18 Why Networks Matter When the Work Happens on the GPU
2:56 Agentic Workloads Are Changing How AI Uses the Network
4:04 Tail Latency: Why TCP, RDMA, and RoCE Struggle With Small Messages
5:00 Idle GPUs: How a 2 Millisecond Message Stalls the Whole Cluster
6:57 Training Workloads vs Agentic Workloads
7:40 Is Homa the Right Protocol for the Edge?
8:12 Are We Bounded by the Speed of Light?
8:58 AI and the Future of Software Developers
10:30 Teaching Software Design at Stanford
12:55 Do You Need to Understand Compute and Networking to Design Software?
14:08 The Origin Story of Tcl
16:14 One Piece of Advice: Stay Broad
#aiinfrastructure #networking #softwaredesign










