← Accueil actu | Salle du marché
2026-09-10 · Source：Nvidia Blog
Nvidia argues AI factories run continuously and must be measured on tokens per second, tokens per watt, cost per token, utilization and uptime.
Nvidia has laid out its case that AI compute should be engineered the way industrial plants are: as an integrated facility whose value is measured at the point of output rather than component by component. In a blog post examining how custom XPUs — the accelerators that hyperscalers and AI-native companies increasingly design in-house — connect to a world-class AI factory, the company frames every infrastructure decision around one goal: generating intelligence at scale.
According to the post, AI factories run continuously, and their economics are defined by delivered output rather than raw hardware specifications. Nvidia lists the governing metrics as tokens per second, tokens per watt, cost per token, utilization and uptime. On that reading, a data center packed with high-performance accelerators still falls short if the facility as a whole cannot convert silicon into usable tokens efficiently, hour after hour.
The article's central claim is that AI infrastructure must be designed and built as a full factory, not a collection of individual accelerators. That argument lands squarely on the custom-silicon programs now underway at hyperscalers and AI-native firms. Nvidia says these builders must weigh additional considerations beyond the chips themselves; the post, filed under the company's NVLink Fusion banner in its web address, points to how custom XPUs are integrated into factory-grade systems as a decisive piece of that equation.
For infrastructure buyers and silicon teams, the significance lies in the shift of the competitive frame from chip-versus-chip benchmarks to facility-level economics. By defining success through tokens per second, tokens per watt, cost per token, utilization and uptime, Nvidia is arguing that custom XPU projects should be judged — and designed — against whole-factory performance rather than accelerator specifications alone, a standard that favors tightly integrated system design.
Mots-clés: Nvidia · NVLink Fusion · XPU · AI Factory · Custom Silicon
萬安算交所 AIXX · AIXX · Bourse compute vérifiée · Accéder à la salle