NVLink Fusion

NVLink Fusion is a product brand from NVIDIA Corporation, introduced as an extension of the company's NVLink high-speed chip-to-chip interconnect. NVLink itself is NVIDIA's proprietary link technology used to connect GPUs and CPUs inside its data-center systems at far higher bandwidth than conventional PCIe. NVLink Fusion opens that fabric to non-NVIDIA silicon: partners can build semi-custom systems in which their own CPUs, AI accelerators or custom ASICs communicate over NVLink with NVIDIA GPUs, while the rack-level design, switch chips and reference architecture remain anchored in NVIDIA's platform. The brand sits at the center of NVIDIA's data-center strategy rather than in its consumer graphics business. NVIDIA sells GPUs, complete rack-scale systems and networking, and its competitive moat has increasingly rested on the software and interconnect layer — NVLink and the NVSwitch fabric — as much as on the chips themselves. By licensing NVLink connectivity to third parties, NVLink Fusion lets hyperscalers and chip designers integrate custom accelerators without abandoning NVIDIA's ecosystem, and it gives NVIDIA a way to keep custom-silicon projects inside its orbit instead of losing them to open interconnect standards. Typical adopters are cloud service providers, system OEMs and semiconductor companies building custom AI silicon. The offering is positioned as a semi-custom platform: partners supply their own processor or accelerator, while NVIDIA supplies the NVLink interface, switch silicon and system design guidance. It is a business-to-business offering, sold as part of NVIDIA's data-center product family rather than as a standalone retail product. NVLink Fusion is not a separate company. It is a branded platform within NVIDIA, which is publicly listed on Nasdaq under the ticker NVDA and headquartered in Santa Clara, California. Its commercial significance lies in ecosystem control: as large customers pursue in-house AI chips, NVLink Fusion gives them a supported path to stay compatible with NVIDIA infrastructure, reinforcing demand for NVIDIA's GPUs, switches and networking.

NVLink Fusion emerged from the evolution of NVIDIA's NVLink interconnect. NVLink was developed as a high-bandwidth, low-latency link for connecting GPUs and CPUs within a server, and successive generations raised per-link and aggregate bandwidth as NVIDIA moved from single GPUs to multi-GPU boards, then to rack-scale systems in which dozens of GPUs operate as one unit. NVSwitch switch chips extended the fabric so that every GPU in a rack could communicate with every other at full bandwidth. As AI workloads scaled, NVIDIA's data-center business shifted from selling individual GPUs toward selling complete systems and networking, with the interconnect and software stack becoming a defining part of the value proposition. At the same time, large cloud customers and chip companies began developing their own AI accelerators to reduce dependence on NVIDIA hardware. Those custom chips faced a practical obstacle: the surrounding ecosystem — system design, switching, software and deployment know-how — was built around NVIDIA's fabric. NVLink Fusion was NVIDIA's answer. Rather than treating custom silicon purely as a competitive threat, the platform offers partners a semi-custom route: their own CPU or accelerator can be attached to the NVLink domain, with NVIDIA supplying the link technology, switch silicon and reference designs. The brand was announced as part of NVIDIA's data-center portfolio and positioned alongside its GPU and networking lines. It reflects a broader pattern in NVIDIA's history of opening parts of its stack — as with CUDA libraries and partner programs — in order to widen adoption of its core platform. The strategic logic is defensive as well as commercial. Open interconnect initiatives and custom-accelerator programs threaten to fragment the data-center market; NVLink Fusion keeps participating silicon compatible with NVIDIA racks, preserving demand for NVIDIA GPUs, NVSwitch and networking. For partners, it lowers the engineering burden of building a system around custom silicon while retaining NVIDIA's software ecosystem. The brand therefore functions less as a discrete product than as a licensing and co-design program embedded in NVIDIA's data-center business.

• NVLink introduced as NVIDIA's high-bandwidth GPU-to-GPU interconnect

• NVSwitch fabric extends NVLink to rack-scale all-to-all GPU communication

• NVLink Fusion launched as a semi-custom platform opening NVLink to third-party CPUs and accelerators

Semi-custom, ecosystem-anchored interconnect platform that keeps custom AI silicon inside NVIDIA's NVLink data-center architecture.

• NVLink Fusion

  Semi-custom interconnect platform allowing third-party CPUs, AI accelerators and custom ASICs to connect over NVLink within NVIDIA-designed systems.

• NVLink

  NVIDIA's proprietary high-bandwidth, low-latency chip-to-chip interconnect.

• NVSwitch

  Switch silicon that links GPUs into a single high-bandwidth fabric at rack scale.

• 2026-08-26

  NVIDIA launches NVHBM custom high-bandwidth memory technology[2]英伟达发布NVHBM定制高带宽内存技术163.com

  Reported: 2026-08-28

  NVIDIA announced NVHBM, a new custom high-bandwidth memory technology that moves the memory controller from the XPU compute die to the base die of the HBM stack, and integrated it into the NVLink Fusion ecosystem. NVIDIA claims up to 30% higher memory bandwidth, 15% lower HBM power consumption and up to 25% more XPU die area for compute versus standard HBM4E.

1. 1斥资867亿元，80倍估值！英伟达收购开源AI平台抱抱脸-观察君-金融界finance.jrj.com.cn
2. 2英伟达发布NVHBM定制高带宽内存技术163.com
3. 3NVIDIA Unveils NVHBM Memory To Turbocharge AI Chip Speeds By 30%hothardware.com