NVIDIA's Vera Rubin NVL72 is in gigascale production at 350+ factory sites across 30 countries, and the benchmark from CoreWeave's first run is concrete: 10x more throughput per megawatt than Grace Blackwell NVL72 on DeepSeek-R1. That is the efficiency curve that makes AI factory build-vs-buy math favor newer hardware faster than any previous generation.
The production story is not the most important signal. NVLink Fusion is. For the first time, NVIDIA is opening its NVLink scale-up interconnect to third-party XPUs. Teams building custom AI accelerators (Tenstorrent, Cerebras, Qualcomm custom silicon, hyperscaler ASICs) can now connect their chips via NVLink rather than designing and validating their own scale-up fabric. The constraint being removed is custom interconnect as a prerequisite for competing in AI inference at rack scale. Previously, any team building an alternative XPU had to either accept Ethernet scale-out latency or build a proprietary interconnect, both of which are multi-year, high-capital efforts. NVLink Fusion changes the calculus.
Two other signals in the Vera Rubin release that are relevant to hardware development teams: first, rack assembly dropped from hours to one minute per compute tray, because three generations of rack-scale codesign produced a system with no cables, fans, or hoses in the tray. That is what happens when packaging, cooling, and interconnect are treated as co-design variables rather than integration afterthoughts. Second, the first co-packaged optics (CPO) switch is in volume manufacturing in the Spectrum-6 SPX, which eliminates the power and reliability tax of pluggable transceivers at scale.
The NVLink Fusion opening is worth watching closely over the next 6-12 months. If NVIDIA delivers real integration depth (not just a protocol bridge), the competitive map for custom AI silicon shifts. Teams that previously had to pick between NVIDIA-native and full independence now have a third option: custom compute tiled into NVIDIA's interconnect and software stack. That is a different product strategy and market position than anyone outside NVIDIA could offer independently.