click the image above to watch on youtube
Welcome, humans.
AI infrastructure used to be the part developers were supposed to forget about. Now the infrastructure is starting to determine what the AI can do.
If you hear words like data center and GPU, but have little understanding of what that means, this episode is for you.
In our latest podcast episode, Corey sits down with Chen Goldberg, Executive Vice President of Product & Engineering at CoreWeave, to unpack what has to change underneath the next wave of AI. Her simplest mental model: a modern AI cluster is no longer a pile of GPUs. It is one enormous supercomputer.
click the image above to watch on youtube
Here’s our favorite parts:
(07:01) The “oh, it’s a supercomputer” moment: Chen explains why compute, networking, storage, cooling, and software all have to behave like one machine once hundreds of GPUs are working on the same job.
(15:06) The engineering identity crisis: If you like writing code, Chen jokes, you may not be the one writing it anymore. One person on her team built something in three weeks that they estimated would once have taken a year.
(16:57) Agents break the old cloud assumptions: A chatbot request can end in seconds. An agent may run for hours, make hundreds of calls, degrade over time, and need the infrastructure to observe, heal, and improve it while it works.
(28:49) Kubernetes is not dead yet: Chen says more than 90% of CoreWeave’s AI workloads still run on Kubernetes, but the underlying resource model has to evolve for tightly coupled AI jobs.
(36:26) The reason to build all of this: Chen argues the payoff is making advanced AI infrastructure accessible to people with different professions, problems, and locations, not keeping it concentrated among a few labs.
The thread connecting all of it: the AI model is only one part of the system now. A slow GPU, a marginal network link, a cooling problem, weak security, or a bad orchestration decision can drag down the whole workload.
Why watch this? Because Chen turns “AI infrastructure” from a vague data-center phrase into a practical explanation of what agents, coding systems, and frontier models actually need underneath them.
Watch and/or Listen now: YouTube | Spotify | Apple Podcasts
P.S. The most human moment starts at 15:06, when the conversation jumps from rack-scale systems to the weird new career question underneath all of this: what does being a software engineer mean when most of the code is generated for you?
Keep scrolling for the Vera Rubin rabbit hole, a word from Outshift by Cisco, and four recent Neuron episodes worth catching up on.
THIS VIDEO WAS BROUGHT TO YOU BY…
Your agents can talk. Can they actually think together?
Today’s agents can call tools and hand work to one another, but shared intent, shared memory, permissions, and guardrails are still messy.
Outshift by Cisco is building the Internet of Cognition, an open approach to infrastructure that lets agents coordinate context, memory, and intent while keeping enterprise controls around the work.
Additional Resources: What “one enormous computer” actually means
CoreWeave’s new multi-rack NVIDIA Vera Rubin NVL72 deployment is a useful case study because it shows how quickly “more GPUs” turns into a systems problem.
CoreWeave’s multi-rack Vera Rubin announcement: the company connected hundreds of Rubin GPUs into one scale-out cluster for training, inference, and agentic workloads.
What it takes to bring the cluster up: a deeper look at networking, power, cooling, validation, and why one slow component can become a cluster-wide straggler.
CoreWeave Fully Connected 2026: the upcoming conference Chen mentions, focused on how customers and operators are actually building AI in production.
🔴 LIVE Thursday: GPT-6 Sol vs. Claude Opus 5.5, Round 2
We’re going back into the arena this Thursday, September 24. After today’s first head-to-head, Corey and Grant are running Round 2 with more prompts, more live testing, and more time to poke at where each model actually wins.
We’ll keep the matchup simple: same jobs, same conditions, and fewer launch-day claims. The point is to see what changes once we can push both models harder.
It’s like Rocky vs Apollo Creed all over again… who is Rocky, GPT? I could see Claude as Creed. Let’s get it!!
🎙️ In Case You Missed It…
Four recent conversations worth adding to the queue:
1. Want to see what frontier coding agents can already build?
click the image above to watch on youtube
TL;DW: Corey and Grant gave GPT-6 Astra six ridiculous one-shot build tests with almost no follow-up steering. It built a black hole simulator, a Blender scene, a physics game, a sci-fi world, a sound diagnostic prototype, and Cat Doom.
Why you should watch: It is a visual answer to the same infrastructure story Chen describes: models are taking on longer, messier jobs, which raises the bar for everything underneath them.
Watch / Listen: YouTube | Spotify | Apple Podcasts
2. Wondering what should stay on your PC instead of the cloud?
click the image above to watch on youtube
TL;DW: Intel’s Dr. Olena Zhu explains hybrid AI, where a local model, edge server, and frontier cloud model split work based on privacy, cost, capability, and available hardware.
Why you should watch: It is the user-device side of the same systems question: once AI becomes infrastructure, routing the work matters almost as much as choosing the model.
Watch / Listen: YouTube | Spotify | Apple Podcasts
3. Building agents? Start with the security boundaries.
click the image above to watch on youtube
TL;DW: Alice CEO Noam Schwartz explains why agent security becomes a different problem once AI can take actions, access tools, and influence other agents.
Why you should watch: Chen makes the same point from the infrastructure side: security can no longer sit in one layer. It has to follow the agent through the entire system.
Watch / Listen: YouTube | Spotify | Apple Podcasts
4. Can AI actually predict what happens next?
click the image above to watch on youtube
TL;DW: Neuralk CEO Alexandre Pasquiou explains why language models are great interfaces, but structured business prediction needs systems designed to learn from rows, columns, distributions, and numbers directly.
Why you should watch: It is another reminder that useful AI systems are stacks, not single models. The data, architecture, and workload shape what the model can actually deliver.
Watch / Listen: YouTube | Spotify | Apple Podcasts
One more before you go:
Chen’s best mental model is worth keeping: when a workload spans hundreds of GPUs, storage, networks, cooling, security, and orchestration, you are not renting a pile of parts. You are operating one giant computer. The more autonomous AI becomes, the more that system-level view matters.
Subscribe to our YouTube Channel for more!
Subscribe on YouTube to help us bring in more builders, researchers, and guests who can teach you something useful about AI every week.
Stay curious,
The Neuron Team
P.P.S: Love the newsletter, but don’t want these podcast and livestream announcement emails? Don’t unsubscribe.
