NVIDIA details NVLink Fusion: plugging custom XPUs into its AI factory stack
NVIDIA Blog · rss · 2026-08-24
An NVIDIA blog post details NVLink Fusion, which lets hyperscalers and AI-native companies connect custom XPUs into NVIDIA's AI infrastructure to build semi-custom AI factories.
- Why: AI factory economics are defined by tokens/sec, tokens/watt, cost per token, and uptime; custom-XPU teams often underestimate the complexity of turning silicon innovation into data-center deployment.
- Performance: Sixth-gen NVLink across a 72-XPU domain delivers 3x lower end-to-end XPU-to-XPU latency and 10x higher packet rates than off-the-shelf Ethernet; NVLink-C2M offers up to 6x the energy efficiency of PCIe; the roadmap includes domains of up to 1,152 accelerators and co-packaged optics.
- Ecosystem: Endorsements from Intel, MediaTek, GUC, QCT/Quanta, and Amazon Annapurna Labs; builders can reuse the MGX rack architecture and supply chain, with NVIDIA DSX reference designs and Omniverse digital twins for pre-construction validation.
More from Infra
- Zai Releases GLM-5.3-Flash: 320B Open Source Model with 1M Context — markjeffrey · 2026-08-27
- Opinion: 'Prefill' Sounds Advanced But Is Simple Once You Understand LLMs — brandon_xyzw · 2026-08-27
- Nvidia's financials are insane; SpaceX may beat them in the future — mitchdeg · 2026-08-27
- Blue-Green Deployment Strategy for Zero-Downtime — _jaydeepkarale · 2026-08-27
- Chinese open models on Huawei chips said to crush US closed models on cost — chris_j_paxton · 2026-08-27
- Video generation speeds: 23.7s vs 11m shows Jevons Paradox in action — gorkem · 2026-08-27