Fusion Worldwide News & Insights - Read Online Today

Network Interface Cards (NIC): How to Scale GPUs in AI Infrastructure.

Written by Ashley Papa | 09.9.2025

In AI Infrastructure, graphics processing units (GPUs) dominate the headlines, but they do not tell the full story. In fact, networking is often the true bottleneck in AI infrastructure, and scaling success depends as much on NIC performance as on the GPUs themselves. The real determinant of scale is networking. Without the right network interface card (NIC), like NVIDIA’s ConnectX‑7 adapters (e.g MCX715105AS-WEAT) and the upcoming ConnectX‑8 generation (e.g 900-9X81Q-00CN-ST0), even the fastest GPUs can sit idle, waiting for data to move.  

These cards are not simple add‑ons. They move data at speeds of up to 400Gbps today and are quickly moving toward 800Gbps. They offload workloads, reduce congestion, and enable scaling across thousands of GPUs, making them as critical as the processors they support.  

What is a NIC and Why It Matters in AI Infrastructure

Simply put, a network interface card (NIC) is the hardware that allows servers and processors to communicate at high speeds. In the context of AI and high-performance computing, NICs are critical for moving massive amounts of data between GPUs and the rest of the system. Without high-performance NICs, GPUs cannot operate at their full potential, leaving expensive infrastructure underutilized.

How NICs Impact GPU Performance in AI Infrastructure

In AI environments, the performance of GPUs is directly tied to the quality and capacity of the NICs supporting them. High-speed NICs reduce latency, manage congestion, and keep massive volumes of data moving seamlessly across thousands of GPUs. This is why demand for advanced solutions like ConnectX-7 and ConnectX-8 continues to rise as AI workloads become more complex and data-heavy.

How ConnectX-7 Improves GPU Performance

The ConnectX-7 family delivers 400Gbps speeds and advanced congestion management capabilities that allow GPUs to operate at maximum efficiency. By offloading workloads and handling traffic with precision, ConnectX-7 enables organizations to scale AI workloads across increasingly large and complex infrastructures.

How ConnectX-8 Future-Proofs GPU Performance

The ConnectX-8 lineup builds on the foundation of the ConnectX-7, offering 800Gbps speeds and PCIe Gen6 support. This next-generation platform not only unlocks higher throughput but also provides the flexibility to future-proof data center performance as AI models and processing demands grow exponentially.

Why Supply of ConnectX-7 Is Tight 

The supply chain pressures around ConnectX‑7 mirror what we saw with GPUs a few years ago. Several key factors drive the shortage: 

  • Advanced IC dependence: Controllers rely on 5nm and 7nm SerDes with limited foundry slots available for production 
  • Optics bottleneck: Each adapter depends on compatible 400G or 800G optical modules, which are also constrained because of Digital Signal Processor (DSP) and laser shortages 
  • Allocation rules: NVIDIA  prioritizes DGX and HGX system builds, leaving smaller buyers with extended lead times 

Delivery windows are stretching to 20-30 weeks, depending on the region and configuration. As noted by Fusion Worldwide’s  Global Commodity Manager, demand for high‑speed Ethernet switches above 200Gbps is growing at double‑digit rates through 2026, signaling that networking demand will continue to outpace stable supply.  

ConnectX-8 Raises the Stakes 

At Hot Chips 2025, NVIDIA revealed the ConnectX‑8 roadmap, signalling a lead to 800Gbps performance with PCIe Gen6 support. This unlocks higher throughput and efficiency but also increases complexity for procurement teams. Early adopters will face similar allocation challenges seen during the ConnectX-7 launch, making forecasting and multi-vendor strategies essential.  

Why Networking Infrastructure is Bigger Than GPUs 

AI scaling is not just about processing power. It is about moving data faster, cleaner, and smarter. 

  • Without ConnectX-7 or ConnectX-8, GPUs wait for data 
  • Networking determines whether you can efficiently scale 1,000 GPUs or 100,000 
  • Supply fragility makes AI infrastructure timelines unpredictable 

The GPU is still the star, but the supporting cast of NICs, DPUs, switches, optics may drive the next critical bottleneck in the AI supply chain.  

What Procurement Teams Should Focus On 

Procurement and engineering leaders can mitigate risk and secure capacity by: 

  1. Treating NICs as priority components in build planning 
  2. Anticipating longer lead times for high-bandwidth networking, especially advanced SKUs. 
  3. Building a global sourcing strategy to navigate region-specific bottlenecks and diversify risk 

By taking a proactive approach, teams can better position themselves to secure critical components, avoid costly delays, and maintain the agility needed to scale efficiently as market demands evolve.

Build Long-Term Supply Chain Confidence 

As the network interface card (NIC) continues to shape AI infrastructure, procurement teams need partners that can deliver agility and insight at scale. This is especially critical for AI and ML teams, where supply constraints on NICs and GPUs can derail deployment timelines if not managed proactively.

At Fusion Worldwide, we match urgent sourcing needs across markets, provide visibility into lead times, and reduce dependence on a single channel. By planning ahead and leveraging tools that enable faster searches, real-time inventory visibility, and streamlined RFQ submissions, procurement teams can respond to market shifts with speed and precision. With Fusion Worldwide on your side, you can turn today’s volatility into tomorrow’s competitive advantage.  

Access the August Greensheet for the latest market data and insights to guide your strategy. Create your free account today to get started.