This interview is a guided tour of NVIDIA’s Blackwell-to-Vera Rubin data-center stack. Joe Delair explains how NVIDIA is co-designing six chips around AI inference and model reasoning, with the headline claim that Vera Rubin can deliver up to 10x better performance per watt at the rack level versus Blackwell by combining GPU, CPU, DPU, networking, and switch silicon into a more modular, liquid-cooled system.
Watch on YouTube ›Get the market thesis, key claims, assets, contradictions, and follow-up questions from any financial video — then unlock a version personalized to your portfolio, watchlist, and favorite speakers.
The core thesis is that NVIDIA’s next-generation Vera Rubin platform is less about one faster GPU and more about an end-to-end data-center redesign. Joe Delair says NVIDIA starts from data-center requirements and works backward, co-designing six chips together to maximize performance, energy efficiency, and cost. The key message is that AI workloads—especially mixture-of-experts and reasoning models that generate many more tokens—are driving an explosion in compute demand, and that demand requires a tightly integrated rack-scale system rather than isolated accelerators. He walks through the Blackwell Ultra stack first, identifying the major components and their roles: the GPU for core AI compute, the Grace CPU for management and CPU-friendly workloads, the BlueField DPU for offloading storage, compression, encryption, and north-south traffic, the ConnectX networking chips for east-west …
Near term, the stock/event setup is GTC-driven and headline-sensitive: any concrete Rubin details, demo footage, or performance claims could keep momentum around NVIDIA, but the same promotional framing leaves room for disappointment if specifics are thin.
Over the next several weeks to months, the base case is that investors focus on whether Rubin reinforces NVIDIA’s inference moat through faster, denser, easier-to-operate racks. That view weakens if customer uptake or real-world benchmarks fail to confirm the claimed step-change.
Structurally, the transcript argues NVIDIA is evolving from a chip vendor into a full AI-factory platform provider. If the full-stack rack strategy holds, the long-term regime is one where systems integration and networking matter as much as GPU silicon.
Nvidia's six-chip co-design approach produces only ~70% more transistors from Blackwell to Vera Rubin but yields 10x performance — far exceeding what Moore's Law alone would deliver.
Joe Delair argues that system-level co-design across GPU, CPU, DPU, ConnectX, NVLink switch, and Spectrum-X creates multiplicative gains beyond process-node transistor scaling.
Nvidia's Vera Rubin rack assembly is 20 times faster than Blackwell GB300 — from 2 hours to 5 minutes.
Joe Delair contrasts the modular, cable-free, 100% liquid-cooled design of Vera Rubin versus the hybrid-cooled GB300 with 43 hoses and fans.
What's the difference between Blackwell and Reuben?
Reuben features extreme co-design of six chips manufactured and designed together for best performance, energy efficiency, and lowest cost. They looked at data center requirements and worked backwards to determine what was needed across all six chips.
What is the difference between Blackwell and Reuben in terms of power and performance for the GPU specifically?
For inference workloads, there is up to 10x better performance on Reuben versus Blackwell in terms of performance per watt at the rack scale. At a given fixed latency, the performance is significantly better for users of the model.
Where are the other two chips — the NVLink switch and the sixth chip?
The NVLink switch lives in a separate switch tray, connecting all 72 GPUs at 1.8 terabytes per second with all-to-all connectivity. It also performs some compute functions like all-reduce collective operations. The sixth chip is the Spectrum X, which goes in separate east-west switch racks.
Unlock the full claims, asset map, scores, related transcripts, follow-up questions, and AI chat — shaped around your portfolio, watchlist, favorite speakers, and risks.
