XPUはいかにして世界クラスのAIファクトリーと融合するかHow XPUs Meet a World-Class AI Factory

NVIDIA has published a technical overview describing how custom accelerators, which it refers to broadly as XPUs, can be integrated into large-scale AI data centers through its NVLink Fusion interconnect technology. The topic matters because the economics of modern AI infrastructure increasingly depend not on the raw specifications of any single chip but on how efficiently an entire facility converts power and silicon into usable model output.

At the center of NVIDIA's framing is the concept of an "AI factory," a term the company uses to describe data centers purpose-built to generate intelligence at scale rather than to run conventional enterprise workloads. According to the company, these facilities operate continuously, and their economics are defined by delivered output measured in metrics such as tokens per second, tokens per watt, cost per token, utilization and uptime. In this view, effective AI infrastructure must be designed and built as a full factory, not a collection of independently optimized parts. The argument is that isolated component improvements do not necessarily translate into better system-level performance if the surrounding architecture, interconnect and software are not aligned.

NVLink is NVIDIA's high-bandwidth, low-latency interconnect, originally developed to link GPUs together so that many processors can behave more like a single, larger computing resource. This matters for large AI models, where the work of training or serving is spread across many accelerators that must exchange data constantly. Bottlenecks in that communication can leave expensive silicon idle, reducing utilization and raising the effective cost per token. NVLink Fusion extends this ecosystem beyond NVIDIA's own GPUs, allowing partners to connect their own custom silicon, whether XPUs or CPUs, into NVIDIA's rack-scale systems and the broader NVLink fabric.

The strategic implication is notable. Many large cloud providers and chip designers have been developing bespoke accelerators tailored to their specific workloads, seeking greater efficiency or reduced dependence on a single supplier. By offering NVLink Fusion, NVIDIA appears to be positioning its interconnect and system architecture as the connective tissue for these heterogeneous designs, rather than requiring every component to be a NVIDIA GPU. In practice this means a customer's custom XPU could sit alongside NVIDIA hardware within the same coherent, rack-scale system, drawing on shared networking, memory access patterns and software tooling.

This fits into a wider industry pattern in which the unit of design has shifted from the chip to the rack and, increasingly, to the entire data center. NVIDIA has previously described rack-scale platforms built around its Grace CPUs and Blackwell GPUs, connected by NVLink switching, as examples of this integrated approach. Supporting technologies in the same ecosystem include high-speed networking such as InfiniBand and Ethernet-based fabrics, along with software layers for orchestration, scheduling and monitoring that aim to keep accelerators busy and power draw predictable. The emphasis on tokens per watt also reflects the growing importance of energy efficiency and power availability, which have become practical constraints on how quickly new AI capacity can be deployed.

For organizations building or renting AI capacity, the framing carries a clear message: raw compute is only one input, and the returns on a facility are shaped by how well all elements work together over sustained periods of operation. Metrics like uptime and utilization underline that reliability and consistent throughput are commercial concerns, not just engineering ones, because idle or underused hardware directly erodes the value of a costly installation.

It is worth noting that this is a vendor overview from NVIDIA's own blog, so the perspective is naturally aligned with the company's product strategy. The extent to which custom XPUs from third parties will be adopted through NVLink Fusion, and how they will perform against alternative interconnect and integration approaches, is likely to depend on real-world deployments, pricing and the willingness of large operators to standardize on NVIDIA's fabric. Even so, the underlying trend the piece describes, treating AI infrastructure as a unified system engineered for delivered output rather than a set of discrete components, reflects a broadly recognized direction in how large-scale AI data centers are now being planned and evaluated.