Premium: Vera Rubin decoder ring
NVIDIA's Vera Rubin platform represents a strategic shift toward Agentic AI at scale, disaggregating workloads across specialized chips and rack systems. Key announcements include a Groq-powered LPX rack for disaggregated inference, a standalone Vera CPU rack for agentic orchestration, and a redesigned MGX modular architecture that dramatically reduces assembly time. NVIDIA is also scaling up supply chain capacity and expanding networking capabilities to support clusters exceeding 500,000 GPUs.
Summary
The transcript analyzes NVIDIA's Vera Rubin platform announcements from GTC March and GTC Taipei, positioning it as a strategic evolution from Grace Blackwell's inference-at-scale focus toward Agentic AI at scale. Vera Rubin achieves this by disaggregating AI workloads across multiple specialized components including GPUs, CPUs, DPUs, and networking hardware, organized into distinct rack systems: Vera, SPX, Spectrum-6, and Groq LPX.
A major operational improvement is the redesign of NVIDIA's MGX modular architecture, which has reduced tray assembly time from over two hours to just five minutes, while also improving ongoing maintainability. NVIDIA is simultaneously doubling its supply chain capacity for Vera Rubin and accelerating assembly timelines across the board.
One of the most significant developments is the introduction of a Groq-powered LPX rack designed to work in tandem with the Vera Rubin NVL72 rack for disaggregated inference. This pairing is intended to drastically improve latency and per-user interactivity by leveraging the strengths of LPU chips while avoiding their limitations. This enables GPUaaS and AI providers to offer premium AI workloads at higher pricing tiers. Notably, NVIDIA has quietly shelved the previously announced Rubin CPX rack, likely due to rising DRAM prices and TSMC bottlenecks, though the author suggests it may return once memory prices stabilize.
NVIDIA is also elevating the role of its Arm-based Vera CPU, which will now be sold as a standalone CPU and rack system specifically designed for agentic orchestration workloads. New AI storage (STX) and shared AI memory (CMX) systems leverage the Vera and BlueField-4 chips to improve data and cache access speeds from Vera Rubin clusters.
On the networking front, Spectrum-X Ethernet has advanced to 800GbE with co-packaged optical ports, enabling scale-out and scale-across cluster sizes expected to exceed 500,000 GPUs. This was announced as in production at GTC Taipei. Finally, NVIDIA introduced the DSX suite of reference designs and tooling to help AI data center operators maximize GPU density per unit of power, addressing ongoing power demand challenges.
About this episode
A detailed look at Vera Rubin's modular racks across disaggregated compute, networking, and storage.
Key Insights
- The author argues that Vera Rubin represents a deliberate strategic shift from inference at scale (Grace Blackwell's domain) to Agentic AI at scale, achieved by disaggregating workloads across specialized chips and rack types rather than consolidating them.
- The author claims NVIDIA quietly dropped its previously announced Rubin CPX rack in favor of the Groq-powered LPX rack, attributing the pivot to rising DRAM prices and TSMC bottlenecks, and suggesting the CPX may return once memory market conditions normalize.
- The author argues that the Groq LPX rack partnership enables a new business model for GPUaaS and AI providers, allowing them to offer premium, latency-sensitive AI workloads at higher pricing tiers than was previously possible.
- The author highlights that NVIDIA's MGX modular architecture redesign is not merely incremental — cutting assembly time from over two hours to five minutes represents a structural supply chain acceleration that supports NVIDIA's goal of doubling Vera Rubin supply capacity.
- The author notes that NVIDIA is repositioning its Arm-based Vera CPU as a serious standalone product for agentic orchestration, signaling a more aggressive push into CPU territory beyond just GPU-adjacent compute.
Topics
Transcript
Now that we covered the major Vera Rubin announcements , let's dive into the finer details on those new modular rack systems across GTC in March and GTC Taipei this week. Grace Blackwell was their push into inference at scale. Vera Rubin is their push into Agentic AI at scale, by disaggregating its needs across different chips (GPU, CPU, DPU, networking) and rack systems (Vera, SPX, Spectrum-6, Groq LPX). NVIDIA is drastically ramping up its capacity , by doubling its supply chain for Vera Rubin and massively speeding up assembly time across it. They redesigned & simplified their MGX modular architectures, taking tray assembly from over 2 hours to now 5 minutes and greatly improving the ongoing maintainability. They are…
Full transcript available for MurmurCast members
Sign Up to AccessMore from HHHYPERGROWTH
Premium: Farther out waves
NVIDIA is expanding its AI ecosystem beyond data centers into physical AI, edge AI, and industrial automation through autonomous vehicles, robotics, and smart devices. The company is anchoring itself as the end-to-end stack across multiple compute tiers, while NVLink Fusion partnerships and a future roadmap including Vera Rubin Ultra and Feynman extend its dominance further.
Premium: Wave after wave of demand
NVIDIA is aggressively driving demand for AI compute through agentic AI software investments, ecosystem partnerships, and supply chain positioning. Major AI buildouts are scaling dramatically, with individual facilities projected to grow from 400-600MW this year to over 2GW by 2028-2029. Frontier AI labs like OpenAI and Anthropic are diversifying compute sources, while Google is pushing its TPU ecosystem into neoclouds.
Premium: Modular inference
NVIDIA's Vera Rubin platform represents a major expansion into a modular AI factory system, incorporating 7 chips including the acquired Groq architecture. The platform is expected to begin generating revenue in Q3 2027, with management projecting significant TAM expansion and revenue uplift across multiple dimensions including inference, storage, and CPU workloads.