Why TCP and RoCE Are Slowing Down AI: John Ousterhout on the Homa Transport Protocol @AkamaiDevelopers
Why TCP and RoCE Are Slowing Down AI: John Ousterhout on the Homa Transport Protocol  @AkamaiDevelopers
Uploaded September 2026 | Updated November 2026, 1 day ago
### What You’ll 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.

### πŸ”— 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

### πŸ“² Follow Us & Join Our Community
πŸ†‡ X: https://x.com/akamaidevs?s=20
πŸ’Ό LinkedIn: linkedin.com/showcase/akamai-dev
πŸ‘Ύ Discord: discord.gg/5MJegzS8Y

### ⏱️ 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
Why TCP and RoCE Are Slowing Down AI: John Ousterhout on the Homa Transport ProtocolAutomating Terraform Workflows Using Variables | Terraform Tapas Ep 6Build a Full-Stack Customer Support Chat App with the MERN StackThe Power of $PATH in your Hands | DevOps Tools at your FingertipsWazuh is a Cybersecurity Powerhouse | Expert Open Source Security Monitoring & ResponseVirtual Private Cloud from Akamai | Secure Networking Within Your InfrastructureDistributed Database Computing | Apache Cassandra Running in Multiple Data CentersDeploy Docker Containers in Production with HashiCorps Nomad | Full Setup Part 1 of 2Build a Private Docker Registry with Quay | Cloud-Native Secure Docker Image StorageEnhancing Terraform Flexibility | Customizing Configurations Using Input VariablesFruits of Spec-Driven Development (SDD) Part 1: Codegen for Client SDKsProxy vs Reverse Proxy | Using Proxies with Cloud Infrastructure Explained
Akamai Developers |

Why TCP and RoCE Are Slowing Down AI: John Ousterhout on the Homa Transport Protocol

SHARE TO X SHARE TO REDDIT SHARE TO FACEBOOK WALLPAPER