NVLink Fusion Integrates Custom Chips into NVIDIA's AI Infrastructure

Summary

NVIDIA's NVLink Fusion solution allows hyperscale and AI-focused companies to integrate their custom XP's into NVIDIA's 72-accelerator scaling domain, shared rack architecture, and software stack. This approach aims to shorten the time and reduce the cost of building a full AI factory platform around custom silicon.

Why it matters

Custom ASICs from companies like Amazon, MediaTek, and Intel are becoming an increasingly central part of hyperscale AI strategies. However, the infrastructure surrounding these chips—scaling fabric, rack design, cooling, and supply chain—remains a significant bottleneck. By offering a 'bring your own XP' path into the NVLink ecosystem, NVIDIA aims to become the default connectivity and rack layer even for chips it does not design. This move could fundamentally alter competitive dynamics in the AI data center market.

Highlights

• 6th-gen NVLink claims to reduce XP-to-XP latency by 3x and increase packet speed by 10x compared to standard Ethernet in a 72-XP domain. The roadmap progresses toward 1,152 accelerators and co-packaged optics.
• NVLink-C2C delivers up to 6x energy efficiency compared to PCIe for XP-to-CPU connections. This feature is designed for agentic workloads that blur the lines between control and computation.
• Adopters share the same MGX rack platform, 100% liquid-cooled hot-swappable trays, and supply chain as Vera Rubin NVL72. Partners such as Intel, MediaTek, Amazon (Annapurna Labs), and Quanta are part of this ecosystem.

Critical take

All performance metrics (3x, 10x, 6x) are self-reported by NVIDIA without independent benchmarking. The phrase 'horizontally open' attempts to mask deep lock-in. Adopters remain dependent on NVIDIA-owned NVLink domains, MGX racks, and software stacks. This marketing copy leaves critical questions unanswered regarding applicable licensing or royalty costs and how much architectural freedom third-party XP's actually retain.