NVIDIA and MediaTek Open NVLink Fusion for Custom AI Chip Design

At COMPUTEX 2025, NVIDIA CEO Jensen Huang unveiled NVLink Fusion, a landmark initiative that opens NVIDIA's proprietary high-bandwidth interconnect to third-party chip designers. For the first time, partners can build custom XPUs—ASICs, domain-specific accelerators, and custom silicon—that coherently connect to NVIDIA GPUs using the same NVLink fabric that powers DGX systems and hyperscale AI factories.

The program has three core pillars:

1. NVLink Fusion Chiplet: A licensable IP block that lets partners integrate NVLink endpoints directly into their custom silicon, enabling coherent memory sharing and low-latency communication with NVIDIA GPUs.
2. NVLink-C2C (Chip-to-Chip): A die-to-die interconnect protocol that supports both NVIDIA-manufactured and partner-manufactured chiplets within the same package, enabling heterogeneous multi-chiplet designs.
3. NVHBM (NVIDIA High Bandwidth Memory): A next-generation memory technology that delivers approximately 30% more bandwidth and 15% lower power consumption compared to standard HBM3e, available to partners through the Fusion ecosystem.

MediaTek: From Smartphone Chips to AI Infrastructure

MediaTek, the world's largest mobile SoC supplier, has emerged as the flagship NVLink Fusion partner. The Taiwanese chipmaker is developing custom AI accelerators that leverage NVLink Fusion to connect coherently with NVIDIA GPUs in data center environments—a strategic pivot that bridges its mobile chip empire with the AI infrastructure boom.

On August 31, 2026, NVIDIA announced a $3.5 billion strategic investment in MediaTek, taking a meaningful equity stake. The investment signals NVIDIA's commitment to building a broad custom-silicon ecosystem rather than maintaining a single-vendor walled garden. For MediaTek, the capital and technology access accelerate its entry into the AI accelerator market, potentially disrupting the current GPU-centric order.

Why Custom Silicon Matters for AI

The AI compute landscape is fragmenting. Hyperscalers—Google with TPU, Amazon with Trainium, Meta with MTIA—have already demonstrated that custom silicon can outperform general-purpose GPUs for specific workloads. NVLink Fusion extends this option to a broader set of companies that lack the engineering scale to build entire AI systems from scratch but can now design accelerators that plug into NVIDIA's proven interconnect and software stack.

Key technical advantages of the NVLink Fusion approach:

• Coherent memory: Custom XPUs can directly access GPU memory through NVLink, eliminating the PCIe bottleneck that typically limits accelerator-to-GPU communication to roughly 32 GB/s. NVLink 5 delivers up to 1.8 TB/s per GPU.
• Unified software stack: Partners' custom silicon works with CUDA, NCCL, and NVIDIA's full AI software ecosystem, avoiding the software fragmentation that has plagued alternative accelerator platforms.
• Heterogeneous packaging: The NVLink-C2C protocol enables multi-chiplet designs mixing NVIDIA and partner die on the same substrate, opening new architectural possibilities for domain-specific AI systems.

Competitive Implications

The NVLink Fusion announcement reshapes the competitive dynamics of AI silicon in several ways:

For AMD: NVIDIA's ecosystem opening undercuts AMD's differentiation around open interconnect standards (Infinity Fabric, UA Link). If partners can get NVIDIA-class interconnect performance while designing their own compute, AMD's value proposition weakens.

For Intel: Intel's foundry strategy gains a potential customer base among NVLink Fusion partners designing custom chiplets, but NVIDIA's packaging technology and TSMC partnership may limit Intel's foundry opportunity.

For the broader ecosystem: The initiative validates the trend toward heterogeneous AI compute—multiple specialized accelerators working in concert rather than homogeneous GPU clusters. This could accelerate AI hardware innovation while simultaneously deepening dependence on NVIDIA's interconnect and software ecosystem.

The Strategic Calculus

NVIDIA's decision to open NVLink reflects a sophisticated strategic calculation. By making its interconnect the de facto standard for custom AI silicon, NVIDIA positions itself as the foundation layer of AI infrastructure regardless of who builds the compute units. Even if custom XPUs capture significant workload share, NVIDIA earns licensing revenue, sells companion GPUs, and maintains software ecosystem control.

For MediaTek, the partnership represents a transformational opportunity. The company's mobile chip business faces secular pressure from smartphone market saturation. AI infrastructure offers a growth market where MediaTek's system-on-chip expertise, cost engineering capabilities, and manufacturing relationships can create genuine competitive advantage.

The $3.5B investment also has geopolitical dimensions. It strengthens the U.S.-Taiwan semiconductor alliance at a time when AI chip export controls and supply chain security dominate policy discussions. MediaTek's role as a NVIDIA ecosystem partner could make it a more strategically important company for both U.S. and Taiwanese interests.

Looking Ahead

The first NVLink Fusion-compatible custom silicon from MediaTek is expected to sample in 2027, targeting AI inference workloads in data center environments. If successful, the program could attract additional partners—particularly cloud providers and enterprise AI platforms seeking custom acceleration without abandoning the CUDA software ecosystem.

The initiative also raises questions about the future of NVIDIA's GPU roadmap. If partners can build custom compute that connects via NVLink, will NVIDIA focus more on interconnect, memory, and system-level innovation rather than raw GPU compute performance? The answer likely depends on how the custom silicon market evolves—and whether NVIDIA's partners can deliver accelerators that genuinely compete with NVIDIA's own GPUs for the most demanding AI training workloads.

What is clear is that NVLink Fusion represents a significant shift in NVIDIA's platform strategy: from building the entire AI system to becoming the interconnect and software foundation upon which a diverse ecosystem of custom AI silicon is built. Whether this openness ultimately strengthens or dilutes NVIDIA's market position will be one of the defining strategic questions of the AI hardware era.