
Sign up to save your podcasts
Or


I packed The Private AI Lab into a suitcase and took it to NVIDIA GTC Europe in Berlin. Six conversations about what it takes to run AI on your own hardware.
You'll hear:
– GTC highlights with ITQ colleague Florian Casse (01:34)
– SUSE AI Factory, quantum-ready AI and secure agents with Sanjeet Singh (05:50)
– SLES for NVIDIA with Miguel Pérez Colino (16:44)
– AI security on the show floor with Luke Wignall (23:23)
– Building an agent with wooden blocks: Nemotron 3.5 Bolt and Switchyard on a DGX Spark, with Luke Wignall & Matthew Feinberg (33:52)
– Post-training Nemotron for your own domain, with Luke Wignall & Amelia (43:34)
– NemoClaw & OpenShell at Build-a-Claw with Mark Heaps (51:27)
Read more: https://johan.ml
Blog on OpenClaw, NemoClaw & OpenShell: https://johan.ml/from-yolo-ops-to-governed-agents-what-openclaw-nemoclaw-and-openshell-mean-for-the-rest-of-us/
The Private AI Lab: experimenting, exploring and debating the future of AI behind the firewall. Hosted by Johan van Amersfoort (ITQ).
NVIDIA just shipped a model with four times the context window of what's been running my home agent for months. So I swapped it in, broke my own setup within the first hour, and then threw 700,000 words at it to see if it would lie to me.
No guest this episode, just me and my home lab. Jarvis, my home agent, has been running on Qwen3.6-35B-A3B since the start, served through vLLM on the DGX Spark's 128GB of unified memory. Then NVIDIA released Nemotron 3.5 Lightning 30B-A3B in NVFP4, and one number stopped me mid-scroll: a validated 1 million token context window, four times what Qwen gives you at 262,000.I don't trust a spec sheet claim until I've broken it myself, so I swapped the model into Jarvis, found a memory reservation bug in NVIDIA's own deployment recipe within the first hour, fixed it, and then buried a fabricated secret code word 90% of the way through a 700,000 word document to see if the model could actually recall it. I also dug into DSpark, the speculative decoding checkpoint built specifically for this hardware, and measured what it actually does for throughput.
This episode is the honest version: what's proven on my own hardware, what's still sitting on NVIDIA's spec sheet, and what I still don't know about whether Nemotron makes Jarvis a better agent day to day.
Key topics covered:
Chapters
00:00 Introduction — what a 1M-token context window actually changes
00:37 Welcome back to The Private AI Lab (no guest this episode)
01:19 Recap: Jarvis running on Qwen3.6-35B-A3B via vLLM on the DGX Spark
01:58 NVIDIA Nemotron 3.5 Lightning lands with a validated 1M-token window
02:49 Deployment: why the Jarvis-LLM alias made the model swap painless
03:28 The memory reservation bug NVIDIA's own recipe doesn't warn about
04:53 The test: a 700,000-word document with one hidden code word
05:51 DSpark speculative decoding: free speed, fully verified
06:47 The reasoning quirk that quietly eats your token budget
07:31 Nemotron vs Qwen: what's proven versus what's still a spec sheet claim
08:01 The hybrid architecture behind why 1M tokens fits on a DGX Spark
09:24 NVIDIA's own SWE-bench Verified numbers, and their limits
09:56 Closing thoughts, and what's coming in the next episode
Links mentioned in this episode:
Most enterprise AI projects succeed as pilots. They fail on the way to production. In this episode, Rhys Oxenham, VP and General Manager of AI at SUSE, joins me to break down what SUSE announced at SUSECon 2026: SUSE AI Factory with NVIDIA.
We go deep on what it actually is, what problem it solves, and what happens when AI starts managing the infrastructure itself. We cover the assembly line concept, validated blueprints, why this isn't just for large enterprises, unified support across the full stack, digital sovereignty as resilience, and the very real governance challenge when agents start acting autonomously in your cluster, what I've been calling "YOLO ops." Release is July 2026. This is worth understanding before then.
🕐 CHAPTERS
00:00 — Intro
01:15 — Who is Rhys Oxenham and what happened at SUSECon 2026
03:19 — What is SUSE AI Factory with NVIDIA?
07:16 — The assembly line: what moves through it and where it starts and ends
10:24 — Blueprints explained: RAG, digital assistants, research agents
12:03 — ClickOps vs GitOps: who is each mode for?
15:10 — Scale: from single node to thousands of GPUs
17:29 — DGX Spark compatibility (teaser)
19:34 — Open source + enterprise support: one throat to choke
21:34 — Day-two operations: what does running it actually look like?
25:54 — Digital sovereignty: open source + NVIDIA — where does the openness live?
30:14 — MCP servers, agentic AI, and integrations with N8N and others
34:00 — YOLO ops: how do you govern autonomous agents in your cluster?
38:32 — Model selection: which model for which use case?
44:16 — AI ops blueprints: what's coming after launch
46:03 — Release date and what to expect in July 2026
48:56 — Wrap up and where to follow Rhys
🔗 LINKS
SUSE AI Factory with NVIDIA → https://www.suse.com
Rhys Oxenham on LinkedIn → https://www.linkedin.com/in/rhys-oxenham/
Johan on LinkedIn → https://www.linkedin.com/in/hojanThe Private AI Lab newsletter → https://www.linkedin.com/newsletters/the-private-ai-lab-7381951883810111489
🎙️ ABOUT THE GUEST
Rhys Oxenham is VP and General Manager of AI at SUSE, where he leads all AI product strategy and execution. He came up through solution architecture and field engineering before running SUSE's Edge and Telco Engineering groups, deploying infrastructure in air-gapped, industrial, and tactical environments. He keynoted SUSECon 2026 in Prague to announce the SUSE AI Factory with NVIDIA partnership.
🔔 SUBSCRIBE for weekly episodes on private AI, self-hosted infrastructure, and what it actually takes to run AI inside your organisation.
📨 Follow The Private AI Lab newsletter on LinkedIn for episode breakdowns and experiment write-ups.
Johan breaks down NVIDIA’s ComputeEx 2026 announcement of Nemotron 3 Ultra 550B-A 55B, an open-weights mixture-of-experts model with 550B total parameters and 55B active, positioned as an orchestration “agent brain” for multi-step tasks behind the firewall. He reviews NVIDIA’s benchmarks versus GLM 5.1, Kimi K 2.6, and Qwen 3.5, highlighting best-in-class instruction following (82%), long-context performance (95%) with a 1M-token window, strong agent productivity (91%), and weaker coding results on TerminalBench versus Kimi. Johan emphasizes reported advantages in speed (~300 tokens/sec, ~5x faster), cost (up to ~30% cheaper on SWE-bench tests), and deployability via a unified NVFP4 checkpoint optimized for H100 and B200 GPUs, plus NemoClaw as the agent blueprint. He closes with an early-access demo comparing two agents researching Netherlands’ 2026 World Cup odds, showing Nemotron’s more granular path analysis and a 5.8% win estimate.00:00 Private AI Lab Intro01:19 Nemotron Ultra Explained02:22 Agent Brain Focus03:07 Benchmark Reality Check05:14 Speed And Cost Edge06:11 Training And Precision08:02 NeMo Claw Agents08:58 World Cup Agent Demo12:22 Why This Matters13:17 Wrap Up And Links
I've been trying to build my own Jarvis for years. Then I met Alexis Gallagher at GTC — and Sparky is the closest thing I've seen.
Alexis is an AI researcher and developer, formerly at Answer AI and Google, now building something most people in AI aren't: a robot designed not just to be useful, but to be *alive*. Sparky lives on his desk in San Francisco. He initiates conversations. He develops his own evolving interests — eels, catenary arches, abandoned infrastructure. He knows who's in the room, when to speak, and when to stay quiet. And he noticed when it was Alexis's first Friday after leaving his job.
In this episode we go deep on the two design goals behind Sparky (useful and alive), the OpenClaw orchestration layer, the social awareness architecture running five times per second, the shared workspace principle that unlocks genuinely useful AI at a desk, and the tradeoffs between cascading and voice-to-voice architectures. We also do a live model switch mid-episode — from Claude Sonnet 4.6 to Nemotron 3 Super 120B running locally on a DGX Spark. It goes impressively well. Until it doesn't. That's in there too.
Guest
Alexis Gallagher — AI researcher and creator of Sparky
🌐 myrobotSparky.com
🔗 https://www.linkedin.com/in/alexis-gallagher/
Key topics covered
- The two design goals: useful AND alive — and why "alive" is the one almost nobody builds for
- How Sparky develops and evolves
- The social awareness stack
- What OpenClaw enables
- The shared workspace principle
- Cascading architecture (STT → LLM → TTS) vs voice-to-voice — the intelligence tradeoff
- Hardware: Reachy Mini Lite, RTX 3090, DGX Spark, Raspberry Pi — the full spectrum
- Live model switch: Claude Sonnet 4.6 → Nemotron 3 Super 120B (the Flowers for Algernon moment)
- The future of personal AI — why embodied social presence is the natural human interface
Chapters
```
00:00 Introduction
00:39 Who is Alexis Gallagher?
01:04 The pivotal AI moment: speech recognition in 2015
03:14 Science fiction to reality — where are the talking robots?
04:22 Sparky introduces himself (live on air)
05:33 The two design goals: useful and alive
07:02 How Sparky initiates conversations — and why that changes everything
08:10 Organic interests: how Sparky evolves what he cares about
09:48 OpenClaw as orchestration layer — soul.md and body control
12:55 Defining a custom robot node type in OpenClaw
15:26 Social awareness: face detection, diarization, presence sensing
16:15 Hardware options: Linux, RTX 3090, DGX Spark, Raspberry Pi
18:25 The Reachy Mini Lite kit — and why it's better than building a drone
19:40 Where to find Alexis and join the Discord
20:10 One eye, four ears — Sparky's hardware explained
24:25 What OpenClaw enables that other frameworks don't
28:13 "Do you have a body, or are you a body?" — a live philosophical exchange
31:17 Live model switch: Claude Sonnet 4.6 → Nemotron 3 Super
33:01 The shared workspace principle — implicit shared attention
38:04 Orchestration in practice: Emacs, sub-agents, cross-platform
40:11 Cascading vs voice-to-voice architecture — the real tradeoff
42:15 Designing Sparky's voice (and the 1930s experiment)
44:12 What's genuinely useful day-to-day — two real examples
48:47 Nemotron 3 Super live — impressive, then the context window
53:38 The model Sparky was running before (Claude Sonnet 4.6)
54:03 Five years out: the future of personal AI companions
58:14 The closest thing to Jarvis I've ever seen
01:00:22 What's coming next — how fast the pieces are moving
01:02:16 Where to find Alexis and join the community
```
Links
- Sparky project and Discord: https://myrobotSparky.com
- Reachy Mini Lite: https://huggingface.co/reachy-mini
The Private AI Lab is hosted by Johan van Amersfoort — Chief Evangelist and AI Lead at ITQ.
📬 Newsletter: https://www.linkedin.com/build-relation/newsletter-follow?entityUrn=7381951883810111489
📝 Blog: https://johan.ml
🔗 LinkedIn: https://www.linkedin.com/in/hojan
In this episode, Johan is joined by long-time colleague Sander Hardewijnen to pull back the curtain on Project Q9 — an ambitious internal project at ITQ that combines a Unitree Go 2 Pro robotic dog, private AI, computer vision, and modern cloud-native development practices.
From gesture recognition trained on 30,000 hand images to a Skynet-obsessed dog posting on LinkedIn, this episode is a deep dive into what happens when you give great engineers a suitcase full of robot and say, "see where it goes."
The conversation also covers the state of open-source AI coding assistants (OpenClaw vs NemoClaw), the realities of vibe coding in a production context, and what partner platforms like Red Hat OpenShift AI and SUSE AI actually enable beyond conversational AI.
Sander's blog: https://harre.dev
Q9's LinkedIn page: https://www.linkedin.com/in/q9-the-dog-2206863b1/
Chapters
00:00 Welcome & Introduction01:20 Icebreaker: Best AI Fail02:12 NemoClaw vs OpenClaw: Security & Sandboxing04:49 Running OpenClaw in an Isolated VLAN05:32 OpenClaw as a Personal Assistant: Home Assistant, News & Efteling API09:11 OpenClaw in the ITQ WhatsApp Group11:10 Introducing Project Q913:22 Why Robotics + Cloud-Native + AI?16:16 Technical Anatomy of Q918:30 Partner Platform Showcase: Broadcom, Red Hat & SUSE19:20 Debunking the GPU Myth23:05 Building the Gesture Recognition Model25:00 Training Progression: Epochs, Accuracy & Landmarks30:21 Hand Landmark Detection & the Gesture Pipeline32:34 Crowd Reactions at KubeCon33:57 Fine-Tuning vs Training From Scratch36:16 Use Case 2: Q9's LLM-Powered LinkedIn Persona40:41 Running LLMs on Partner Inference Platforms42:26 What's Next for Q9?43:44 Digital Twins in NVIDIA Omniverse + ROS245:10 Key Takeaways48:53 Responsible Vibe Coding49:58 Open-Sourcing Q9 — Coming Soon
In this episode of The Private AI Lab, Frank Denneman returns as the first recurring guest to go deeper into one of the most misunderstood challenges in AI:
👉 Resource management for GPU workloads
Building on our previous conversation, this episode shifts from why it matters to how to actually design it right.
We dive into real-world challenges like GPU fragmentation, siloed capacity, and why traditional infrastructure thinking breaks down when AI enters the data center. Frank shares practical insights from his latest research, blog series, and tools—helping architects and platform engineers understand how to design efficient, scalable AI environments.
🔍 What you’ll learn in this episode
Why GPU workloads behave fundamentally differently from CPU/memory workloads
What GPU fragmentation really is (and why it kills utilization)
The difference between same-size vs mixed-mode placement
How placement IDs turn GPU scheduling into “Tetris”
Why “right-sizing” beats “perfect fitting” in AI environments
How to design a GPU profile catalog that actually scales
The role of state, agents, and storage in next-gen AI platforms
🔧 Tools & Resources mentioned
Frank created practical tools to help you design and validate your GPU environments:
👉 vGPU Silo Capacity Calculator
https://frankdenneman.ai/tools/vgpu-silo-capacity-calculator/
👉 Same-size vs Mixed-mode Placement Tool
https://frankdenneman.ai/tools/same-size-vs-mixed-mode/
👉 Deep dive on unified memory & modern AI workloads
https://frankdenneman.ai/posts/2026-03-23-understanding-unified-memory-dgx-spark-nemoclaw-nemotron/
Chapters:
00:00 Intro — Frank Denneman returns
01:30 AI hype vs real engineering
03:00 DGX Spark, NemoClaw & local AI agents
10:30 From LLMs to agents & stateful systems
12:00 Why AI infrastructure is different
15:00 What is GPU fragmentation?
19:30 Same-size vs mixed-mode placement
23:00 GPU “Tetris” and placement IDs explained
27:00 Right-sizing vs perfect fitting
32:00 The tools: capacity & placement simulation
36:00 GPU silos vs stranded capacity
41:00 Model sizing, KV cache & dynamic usage
48:00 Future of AI: smaller models & orchestration
55:00 AI-assisted coding & real-world impact
59:00 Key lessons learned
01:02:00 Closing thoughts
AI in healthcare didn’t start with ChatGPT.
Long before generative AI, hospitals were using machine learning for sepsis detection, imaging diagnostics, and predictive analytics. In this episode of The Private AI Lab, Johan sits down with Vincent Tsugranes, Chief Architect at Red Hat, to explore what’s real, what’s hype, and why platform matters more than ever.
They discuss:
Why 95% of AI projects fail
The evolution from OpenShift Data Science to OpenShift AI
Models-as-a-Service inside hospitals
vLLM vs LLMD for large-scale inference
Guardrails, hallucinations, and enterprise risk
Sovereign cloud and why healthcare is moving on-prem again
What “ambient AI” might mean in the next 12 months
This episode is for architects, platform engineers, healthcare IT leaders, and anyone building private AI in regulated environments.
00:00 – Red lights & farming with AI
02:10 – The first AI spark moment
04:00 – When “AI” became AI (ChatGPT moment)
07:20 – Why 95% of AI projects fail
11:00 – Machine learning vs modern AI
13:30 – Platform vs point solutions
16:00 – The history of OpenShift AI
19:00 – What is OpenShift AI under the hood?
22:00 – Hardware enablement & NVIDIA
25:00 – vLLM explained
27:30 – LLMD and distributed inference
30:00 – Healthcare use cases (sepsis, imaging, insurance)
33:00 – Models-as-a-Service inside hospitals
36:00 – Guardrails & hallucination risks
39:00 – Observability & FinOps explosion
42:00 – OpenShift 5 and platform intelligence
44:30 – Sovereign cloud in healthcare
48:00 – The future: ambient AI & rising power bills
Recorded live at NVIDIA GTC 2026, this episode dives into one of the biggest announcements in open AI: Nemotron Super.
Together with Joey Conway, we explore how NVIDIA is pushing open source AI forward — with models that are not only powerful, but also efficient and enterprise-ready.
We discuss:
The evolution from Llama-based models to Nemotron
Why reasoning + agentic capabilities matter
How NVIDIA balances performance and efficiency
What NVFP4 means for running AI locally
And why this could be a turning point for AI behind the firewall
Chapters
00:00 Intro
01:56 Welcome
02:37 GTC insights
03:31 Nemotron buzz
04:53 Model evolution
07:14 Core design principles
09:05 Reasoning capabilities
10:52 Scaling challenges
12:00 Architecture deep dive
13:12 Performance improvements
14:14 Quantization strategy
15:39 NVFP4 explained
16:16 DGX Spark use case
18:23 Broader adoption
19:37 Agentic AI impact
21:25 Try it yourself
22:03 Outro
Links
Try Nemotron: https://build.nvidia.com
More episodes: https://johan.ml