AI EVENTS
AI engineering, from the conference floor.
I go to AI conferences like NVIDIA GTC, RAISE and MACHINA to hear from the people building the models, chips and infrastructure. These are my notes for engineers: the technical detail behind what was said, and what I’d try next.
RAISE Summit
Paris · July 2026
MACHINA Summit
Paris · July 2026
NVIDIA GTC
Berlin · October 2026
BERLIN · 20–22 OCTOBER 2026
NVIDIA’s GTC, in Berlin this year. I went with a list of questions people had sent me, and most of my notes ended up being about agents and the hardware they run on.
Agents are finally getting a real security boundary
This is the one I’d watch most closely. Most agents today run with whatever access the developer’s machine has, and we hope the prompt keeps them in line. OpenShell puts each agent in a sandbox, with a supervisor that checks every file access, network call and credential request against a policy, and it uses formal verification to show what a policy change would open up before you apply it. Sentry then watches from a BlueField DPU outside the host and can quarantine a misbehaving agent within milliseconds. An agent can’t switch off security it can’t reach.
Your agent is waiting on the CPU more than you think
We talk about agents as if they live on the GPU, but a long run spends a surprising amount of time off it, parsing, compiling, running tests and preparing data between model calls. NVIDIA built Vera for exactly that work: a CPU tuned for per-core speed and memory bandwidth, with LPDDR5X on SOCAMM modules you can replace in the field instead of having them soldered down. One honest caveat: some of what looks like slow tools is really network I/O, and no CPU fixes that, so profile before you buy.
Open models you can actually run at your desk
Nemotron 3.5 Bolt E4B is the one I’m most curious about. It’s post-trained from Gemma 4 E4B, runs in about 6GB of memory when quantized, and on NVIDIA’s own chart it scores above the larger Qwen 3.5 9B. A companion DFlash model drafts tokens ahead for speculative decoding, which is what makes on-device voice assistants and small robots feel responsive. At the other end, Super VL (120B total, 12B active) mixes Mamba-2, attention and MoE layers and reads charts, documents and video. NVIDIA is also releasing around 10 of the 37 datasets it used for RL, and for anyone fine-tuning, that matters as much as the weights.
Local AI is turning into an Arm story
RTX Spark puts a Blackwell GPU and a 20-core Arm CPU in a Windows laptop, with up to 128GB of memory shared between them and up to a petaflop of FP4. DGX Spark is the developer-kit version, and you can cluster a few over ConnectX-7 when one isn’t enough. That is a lot of model on a desk. Two things to check first: the OS and your other apps share that memory too, and CUDA on Arm doesn’t mean every x86 container in your stack will just run.
The AI factory is now something you order
The numbers here are wild. One Vera Rubin superchip pairs a Vera CPU with two Rubin GPUs, each with up to 288GB of HBM4 at 22TB/s, and an NVL72 rack holds 72 of those GPUs. DSX POD ships that kind of system pre-built in a container: on site, the crew sets it down, connects power and cooling, links it to its neighbours and runs a burn-in. NVIDIA quoted about 14 days for that, against 18 to 24 months to build a data-centre shell. As they put it, an AI factory creates value when it starts producing tokens, not when construction is complete.
FOR AI ENGINEERS
Five things to do after GTC
Sandbox your agents now
Write down what each one can read, write and call, then enforce it. OpenShell is on GitHub, so you don’t need new hardware to start.
Profile the whole agent loop
Parsing, tests and tool calls between model calls add up. Measure them before you blame the GPU.
Try a small model first
Bolt E4B runs in about 6GB. Test it on your own evals, not NVIDIA’s chart, before you pay for something bigger.
Check that your stack runs on Arm
RTX Spark and DGX Spark are Arm machines. Find your x86-only containers before your next dev box does.
Learn about the rack, not just the GPU
Memory, interconnect and power decide what you can serve, and the NVL72 rack is now the unit people plan around.
AI EVENTS
AI engineering, from the conference floor.
I go to AI conferences like NVIDIA GTC, RAISE and MACHINA to hear from the people building the models, chips and infrastructure. These are my notes for engineers: the technical detail behind what was said, and what I’d try next.
BERLIN · 20–22 OCTOBER 2026
NVIDIA’s GTC, in Berlin this year. I went with a list of questions people had sent me, and most of my notes ended up being about agents and the hardware they run on.
Agents are finally getting a real security boundary
This is the one I’d watch most closely. Most agents today run with whatever access the developer’s machine has, and we hope the prompt keeps them in line. OpenShell puts each agent in a sandbox, with a supervisor that checks every file access, network call and credential request against a policy, and it uses formal verification to show what a policy change would open up before you apply it. Sentry then watches from a BlueField DPU outside the host and can quarantine a misbehaving agent within milliseconds. An agent can’t switch off security it can’t reach.
Your agent is waiting on the CPU more than you think
We talk about agents as if they live on the GPU, but a long run spends a surprising amount of time off it, parsing, compiling, running tests and preparing data between model calls. NVIDIA built Vera for exactly that work: a CPU tuned for per-core speed and memory bandwidth, with LPDDR5X on SOCAMM modules you can replace in the field instead of having them soldered down. One honest caveat: some of what looks like slow tools is really network I/O, and no CPU fixes that, so profile before you buy.
Open models you can actually run at your desk
Nemotron 3.5 Bolt E4B is the one I’m most curious about. It’s post-trained from Gemma 4 E4B, runs in about 6GB of memory when quantized, and on NVIDIA’s own chart it scores above the larger Qwen 3.5 9B. A companion DFlash model drafts tokens ahead for speculative decoding, which is what makes on-device voice assistants and small robots feel responsive. At the other end, Super VL (120B total, 12B active) mixes Mamba-2, attention and MoE layers and reads charts, documents and video. NVIDIA is also releasing around 10 of the 37 datasets it used for RL, and for anyone fine-tuning, that matters as much as the weights.
Local AI is turning into an Arm story
RTX Spark puts a Blackwell GPU and a 20-core Arm CPU in a Windows laptop, with up to 128GB of memory shared between them and up to a petaflop of FP4. DGX Spark is the developer-kit version, and you can cluster a few over ConnectX-7 when one isn’t enough. That is a lot of model on a desk. Two things to check first: the OS and your other apps share that memory too, and CUDA on Arm doesn’t mean every x86 container in your stack will just run.
The AI factory is now something you order
The numbers here are wild. One Vera Rubin superchip pairs a Vera CPU with two Rubin GPUs, each with up to 288GB of HBM4 at 22TB/s, and an NVL72 rack holds 72 of those GPUs. DSX POD ships that kind of system pre-built in a container: on site, the crew sets it down, connects power and cooling, links it to its neighbours and runs a burn-in. NVIDIA quoted about 14 days for that, against 18 to 24 months to build a data-centre shell. As they put it, an AI factory creates value when it starts producing tokens, not when construction is complete.
FOR AI ENGINEERS
Five things to do after GTC
Sandbox your agents now
Write down what each one can read, write and call, then enforce it. OpenShell is on GitHub, so you don’t need new hardware to start.
Profile the whole agent loop
Parsing, tests and tool calls between model calls add up. Measure them before you blame the GPU.
Try a small model first
Bolt E4B runs in about 6GB. Test it on your own evals, not NVIDIA’s chart, before you pay for something bigger.
Check that your stack runs on Arm
RTX Spark and DGX Spark are Arm machines. Find your x86-only containers before your next dev box does.
Learn about the rack, not just the GPU
Memory, interconnect and power decide what you can serve, and the NVL72 rack is now the unit people plan around.