NVIDIA Vera Rubin、ワット当たり性能とトークンコストで世界のパートナーに優位性をもたらすNVIDIA Vera Rubin Driving Performance Per Watt, Lowest Token Cost for Partners Worldwide

NVIDIA has positioned its Vera Rubin platform as a step change in the economics of running artificial intelligence at scale, emphasizing two metrics that increasingly determine whether large deployments are viable: performance per watt and the cost of generating each token during inference. As demand for generative AI services grows, operators are less concerned with raw peak throughput and more focused on how much useful work a data center can produce within fixed power and budget constraints. Vera Rubin is aimed squarely at that shift.

The platform pairs a new NVIDIA-designed CPU, named Vera, with a next-generation GPU called Rubin, both named after the American astronomer Vera Rubin, whose work provided evidence for dark matter. This follows NVIDIA's established pattern of naming architectures after scientists, and it succeeds the Grace CPU and Blackwell GPU generation. The tight coupling of CPU and GPU through high-bandwidth NVLink interconnects is central to the design, allowing memory and compute resources to be treated more like a single large system than a collection of discrete chips. Rubin is expected to use next-generation high-bandwidth memory, reported to be HBM4, which increases the bandwidth available for feeding data to the processor, a common bottleneck in large language model workloads.

Performance per watt matters because power, not floor space or even chip supply, has become the limiting factor for many operators. Grid connections, cooling capacity, and electricity costs cap how many accelerators a facility can realistically run. If Vera Rubin can deliver more tokens or more training throughput for the same amount of energy, an operator can either serve more customers within an existing power envelope or lower the energy cost embedded in every AI query. NVIDIA frames this in terms of what it calls AI factories, where the relevant output is tokens produced and the relevant input is energy consumed, making efficiency a direct driver of margin.

The emphasis on per-token cost reflects how the industry's center of gravity has moved from training to inference. Training a model is a large but finite expense, whereas inference recurs every time a user interacts with a service, and at scale it can dominate total cost of ownership. Lowering the cost per token therefore has an outsized effect on the profitability of AI products, which is why cloud providers and enterprises evaluating deployments watch this figure closely. NVIDIA argues that architectural gains, combined with software optimizations in its inference stack such as TensorRT and the Dynamo serving framework, compound to reduce that cost. Actual savings will depend heavily on model size, workload mix, and how effectively customers tune their systems, so specific figures should be treated as vendor projections until independent benchmarks appear.

Context helps explain the stakes. Vera Rubin arrives as competition intensifies. AMD continues to push its Instinct MI300 and successor accelerators, while hyperscalers including Google, Amazon, and Microsoft are investing in custom silicon, such as TPUs, Trainium, and in-house designs, partly to reduce dependence on any single supplier. Efficiency claims are one way NVIDIA seeks to defend its dominant position, which rests not only on hardware but on the CUDA software ecosystem that many developers already rely on. The company's broader strategy also spans networking, through its Spectrum-X and NVLink technologies, and reference rack designs that integrate compute, memory, and cooling, reflecting a move toward selling systems rather than components.

For partners, the practical appeal is a more predictable path to scaling AI services without proportionally scaling energy bills or capital outlay. Rack-scale integration is likely to matter as much as the chips themselves, because dense, power-hungry accelerators increasingly require liquid cooling and carefully engineered power delivery that individual buyers struggle to assemble alone. Prospective customers will nonetheless want to weigh migration effort, supply availability, and total cost against alternatives before committing.

As with any newly introduced architecture, real-world results will become clearer once systems ship in volume and third parties publish comparative measurements. Until then, Vera Rubin represents NVIDIA's stated direction for the next phase of AI infrastructure, one in which efficiency and token economics, rather than headline performance alone, appear to be the decisive competitive terrain.