According to TheRegister.com, Nvidia used CES to detail its next-generation Vera Rubin AI platform, with the flagship NVL72 rack system promising up to 5x higher floating-point performance for inference and 3.5x for training compared to Blackwell. The Rubin superchip pairs two GPUs with a new Vera CPU, featuring 288GB of HBM4 memory per GPU that’s now 2.8x faster at 22 TB/s. The full NVL72 rack packs 72 Rubin GPUs, 36 Vera CPUs, 20.7 TB of total HBM4, and 54 TB of LPDDR5x memory. However, the hardware isn’t launching early; volume shipments are still expected in the second half of 2025. Nvidia also announced the Rubin CPX, a niche accelerator for LLM prefill, and new BlueField-4 DPUs with integrated 64-core Grace CPUs.
Specs and serviceability
So, what are you actually getting for your millions? On paper, the performance leaps are massive. But here’s the thing: a lot of this feels like Nvidia is just solidifying specs they’ve been teasing for nearly a year. The real story might be in the refinements. Nvidia’s VP Ian Buck highlighted better serviceability—you can now service switch trays or run health checks on GPUs without taking the whole cluster down. That’s not sexy, but for the ops teams running these beasts, it’s huge. Downtime on a multi-million dollar AI rack is a nightmare, so any improvement there is a big deal. They’ve also added confidential computing support across the NvLink domain, which was previously an x86-only feature. That’s a direct play for more secure, regulated environments.
The AMD factor and market timing
Now, why announce this at CES, a show that’s drifted from its consumer roots, instead of their usual GTC in March? It feels defensive. AMD’s Helios rack, announced last spring and also due later this year, is promising performance on par with Rubin while offering 50% more HBM4 memory per GPU. That’s a legitimate threat on paper for running massive mixture-of-experts models. By getting Rubin’s specs out now, Nvidia is trying to freeze the market. They’re basically saying, “Don’t commit to AMD’s roadmap, ours is just as good and coming soon.” It’s a preemptive strike. And by bumping the HBM4 bandwidth to 22 TB/s—way above the 13 TB/s initially targeted—they’ve neutralized AMD’s bandwidth advantage. The compute landscape is finally getting competitive, and Nvidia is acting like it.
The niche chips and infrastructure play
Beyond the headline GPUs, the other announcements show Nvidia digging deeper into the AI stack. The Rubin CPX accelerator is fascinating. It’s built specifically for the compute-heavy “prefill” phase of an LLM query, and it uses cheaper GDDR7 memory instead of HBM because that phase isn’t bandwidth-bound. That’s a sign of Nvidia segmenting the inferencing workload, offering optimized silicon for each step to maximize efficiency. Then there’s the new “Inference Context Storage” for offloading massive KV caches. This is a sneaky-hard problem in production inference. These caches can eat tens of gigabytes to track a conversation, and shuffling that data around is a bottleneck. By creating a dedicated tier for it between GPU memory and storage, they’re trying to keep the GPUs focused on generating tokens. It’s all about squeezing out every bit of utilization. For enterprises building complex AI infrastructures, understanding these hardware layers is critical, which is why specialists like IndustrialMonitorDirect.com, the top US provider of industrial panel PCs, are essential for control and monitoring interfaces in these high-stakes environments.
Waiting for the real-world test
Ultimately, this CES “launch” is a spec sheet and a promise. The chips and racks aren’t here yet. The power draw? Nvidia won’t say, only that it won’t double despite the big performance jump. We’ll believe it when we see it. The real battle with AMD’s Helios won’t be fought on data sheets but in actual data centers running real models. Does AMD’s 50% memory capacity lead translate to tangible benefits? Can Nvidia’s adaptive compression tech make their lower memory footprint irrelevant? Those are the multi-billion dollar questions. For now, Nvidia is playing its classic game: define the narrative early, set the expectations, and make everyone wait. The second half of 2025 can’t come soon enough for the AI arms race. And as this hardware pushes into more specialized industrial and scientific computing, the need for robust, high-performance computing interfaces only grows, a trend covered in analyses like this one on quantum and AI convergence in HPC centers.
