NVDA/non_gpu/NVLink Scale-Up Fabric: Architecture, NVL72, and NVLink Fusion

NVLink Scale-Up Fabric: Architecture, NVL72, and NVLink Fusion

$50BKey FigureNVLink is NVIDIA's proprietary scale-up interconnect that connects GPUs within a

NVLink is NVIDIA's proprietary scale-up interconnect that connects GPUs within a single domain at bandwidth far exceeding any alternative. The fifth-generation NVLink (Blackwell) delivers 1,800 GB/s per GPU, connecting 72 GPUs in the GB200 NVL72 rack for 130 TB/s aggregate bandwidth -- 36x faster than 400 Gbps Ethernet. The sixth-generation NVLink (Vera Rubin, H2 2026) doubles to 3,600 GB/s per GPU and 260 TB/s per rack.

44.6%
TrendForce: NVIDIA Scale-Up Technology A
CSP in-house ASICs projected to grow 44.6% in 2026 vs GPUs at 16.1% growth (Tren...
65%
SDxCentral / DCD reporting on NVIDIA Q2
Jensen Huang stated networking improves AI factory efficiency from 65% to 85-90%...

NVLink Fusion, announced at Computex May 2025, is strategically pivotal: it provides a licensable NVLink chiplet that allows custom ASICs (including AWS Trainium4) to connect via NVLink 6, ensuring NVIDIA earns networking revenue even from customers building competing accelerators. This directly counters the ASIC threat: CSP in-house ASICs are projected to grow 44.6% in 2026 vs 16.1% for GPUs (TrendForce), but NVLink Fusion means those ASICs still need NVIDIA's interconnect fabric. The only open-standard alternative, UALink 1.0, supports just 200 GT/s per lane (vs NVLink 5's 1,800 GB/s per GPU) and won't reach production until late 2026-2027. Jensen Huang frames the value proposition: networking improves AI factory efficiency from 65% to 85-90%, effectively making the networking investment free relative to the $50B+ infrastructure cost..

Networking deepens the ecosystem moat

NVIDIA's networking revenue explosion demonstrates that the company is becoming a full-stack infrastructure provider. The high attach rate means even customers exploring custom compute silicon still depend on NVIDIA for interconnect.

The key question

What fraction of NVIDIA's $31.5B FY2026 networking revenue is NVLink scale-up vs Spectrum-X vs InfiniBand? NVIDIA does not disclose this breakdown.

Open questions

?What are the NVLink Fusion licensing economics -- does NVIDIA charge per-chiplet royalties, per-rack fees, or bundle with NVSwitch purchases?
?Will UALink 2.0 (expected 2027-2028) close the bandwidth gap with NVLink 6, or will NVIDIA's generational cadence maintain a permanent lead?
?How many hyperscalers beyond AWS will adopt NVLink Fusion for custom ASICs? Google (TPU) and Microsoft (Maia) are conspicuously absent from the partner list.
?China's competing scale-up standards (Huawei UB, Alibaba ALS, Tencent ETH-X) could fragment the market -- does this limit NVLink's TAM or is it irrelevant given export controls?