NVIDIA’s NVLink Fusion Could Supercharge Custom AI Chips and Reshape the AI Factory Race

The race to build the next generation of artificial intelligence infrastructure is moving beyond faster chips. As AI workloads become increasingly massive and continuous, the real challenge is building entire AI factories capable of delivering enormous token throughput while controlling power consumption, cost, reliability and downtime.

NVIDIA says its new NVLink Fusion platform is designed to tackle that challenge by allowing hyperscalers and AI-native companies to integrate their custom XPUs with NVIDIA’s established AI infrastructure.

NVIDIA Targets a Major Bottleneck in Custom AI Silicon

Custom XPUs can be highly optimized for specific AI workloads, but turning an accelerator design into a production-ready data-center platform is far more complicated than designing the chip itself.

Companies must develop scale-up and scale-out networking, rack architectures, cooling and power systems, software, security, storage and manufacturing infrastructure. They must also coordinate a large network of suppliers and validate the entire platform before deployment.

NVIDIA argues that this complexity can significantly slow the path from silicon development to large-scale AI deployment.

NVLink Fusion is designed to remove much of that burden by connecting custom XPUs to NVIDIA’s mature AI infrastructure stack.

Faster Scale-Up for Massive AI Models

Modern AI systems increasingly depend on massive models, mixture-of-experts architectures and agentic workloads. When communication between accelerators becomes a bottleneck, utilization can fall while the cost of generating each token rises.

According to NVIDIA, NVLink Fusion brings custom XPUs into the company’s NVLink scale-up ecosystem.

Sixth-generation NVLink supports high-bandwidth, low-latency communication across a 72-XPU domain. NVIDIA says XPU-to-XPU transfer latency can be three times lower than comparable approaches based on off-the-shelf Ethernet, while packet rates can be up to 10 times higher.

The company also points to future NVLink configurations scaling to domains of up to 1,152 accelerators, alongside co-packaged optics.

NVLink Fusion also incorporates NVLink-C2C, allowing XPUs to connect with NVIDIA Vera CPUs or other ecosystem CPUs. NVIDIA says the technology can deliver up to six times the energy efficiency of PCIe for these connections.

A Complete Platform Instead of Just a Chip

For companies developing custom AI silicon, the biggest challenge may not be the XPU itself but everything surrounding it.

A production AI platform requires:

• High-speed CPU and scale-up interfaces
• Scale-up networking
• Compute and switch trays
• Rack architecture
• Advanced cooling and power delivery
• Security and storage
• Manufacturing and supply-chain integration
• Software for distributed AI workloads and cluster management

NVLink Fusion is intended to provide many of these building blocks through an established ecosystem.

NVIDIA says partners can leverage its MGX rack-scale architecture and supply chain, including infrastructure used for systems such as Vera Rubin NVL72.

That could allow custom-XPU developers to concentrate their engineering resources on differentiating silicon and software rather than rebuilding an entire data-center platform from scratch.

Standardized Infrastructure Could Reduce AI Factory Risk

Another major advantage is flexibility.

AI data-center construction often begins well before the final accelerator configuration is determined. Power procurement, cooling infrastructure, rack layouts and networking decisions must be made years or months ahead of silicon deployment.

A facility designed around a single accelerator architecture can therefore become a major scheduling and supply-chain risk.

NVIDIA says NVLink Fusion can help address this by allowing XPU-based and GPU-based systems to share infrastructure such as rack footprints, networking, cooling, power delivery and management systems.

That means operators could begin building their AI factories while delaying some decisions about the exact silicon mix.

Different systems could then be optimized for different workloads, including training, post-training, reasoning, retrieval and inference.

Software Completes the AI Factory

Hardware alone cannot operate an AI factory efficiently.

NVIDIA’s broader software stack includes NCCL for distributed workloads, Dynamo and NIXL for disaggregated AI infrastructure, and Mission Control for cluster management, telemetry and debugging.

Together, these components are intended to turn a collection of accelerators and racks into a coordinated computing platform.

NVIDIA is also positioning NVLink Fusion alongside its DSX AI Factory architecture, which is designed to help organizations model and optimize buildings, power, cooling, networking and computing infrastructure before physical deployment.

Why NVLink Fusion Matters

The bigger story behind NVLink Fusion is NVIDIA’s attempt to make its infrastructure relevant even when the accelerator inside the rack is not entirely NVIDIA-designed.

Instead of forcing every AI company to choose between proprietary custom infrastructure and NVIDIA’s GPU ecosystem, NVLink Fusion creates a middle ground: custom silicon combined with NVIDIA’s networking, rack architecture, software and manufacturing ecosystem.

For hyperscalers and AI-native companies, that could mean faster deployment, lower infrastructure risk and more flexibility in choosing accelerators for specific workloads.

If NVIDIA can successfully make NVLink Fusion the infrastructure layer connecting different classes of AI processors, the company could strengthen its position in the AI factory market even as custom silicon becomes increasingly important.

The future AI factory may not be built around one chip. It may be built around a common infrastructure platform capable of supporting many different kinds of AI processors — and NVIDIA wants NVLink Fusion to be that platform.