NVIDIA and MediaTek signed a $3.5 billion investment agreement to accelerate the adoption of the NVLink Fusion interconnect architecture.
Building on a longstanding partnership, NVIDIA announced this week that it will deepen its ties with fellow fabless semiconductor firm MediaTek through a $3.5 billion investment. As part of the agreement, MediaTek will adopt NVIDIA’s NVLink Fusion platform to offer to its own customers, strengthening both financial and technological alignment between the two companies.
Under the terms of the deal, NVIDIA will purchase convertible bonds issued by MediaTek. On the technology front, MediaTek will license NVIDIA’s complete NVLink Fusion stack, including the Fusion chiplet, NVLink-C2C, and the newly introduced NVHBM memory protocol. This licensing arrangement addresses MediaTek’s rapidly expanding custom silicon business. By making its IP readily available to MediaTek’s clients—particularly those co-developing custom XPU accelerator chips with MediaTek—NVIDIA ensures broader ecosystem integration.
The partnership has grown increasingly significant within NVIDIA’s supply chain. Most notably, MediaTek served as NVIDIA’s primary partner in developing the GB10 system-on-chip, the core processor powering NVIDIA’s DGX and RTX Spark systems. While officially branded as an NVIDIA chip, MediaTek led the majority of the development, designing the CPU cores and memory controller, while NVIDIA supplied the integrated GPU and chip-to-chip interconnects. The additional capital injection underscores that this agreement extends well beyond standard IP licensing.
Given their close prior collaboration, some observers may find it notable that a broader technology agreement was not finalized sooner. MediaTek already utilizes NVLink-C2C for the GB10, for instance. However, this latest arrangement targets custom chip development rather than MediaTek’s internal product lines, representing a substantially more complex and strategic commitment.
The overarching implication is that NVIDIA is strategically positioning itself within MediaTek’s custom chip ecosystem. As hyperscalers and major AI operators—including Google with its TPUs, Meta with its MTIA, and OpenAI with its Jalapeno project—increasingly design proprietary accelerators to hedge against reliance on single vendors, NVIDIA is mirroring that strategy. By granting these developers streamlined access to the technologies required to interface with NVIDIA’s infrastructure, the company secures its relevance across a fragmented hardware landscape.
Technically, this means extending the full NVLink Fusion ecosystem to MediaTek’s client base. While NVIDIA has previously offered these tools directly to independent chip designers, MediaTek’s involvement effectively pushes the stack downward, enabling the Taiwanese firm to distribute the technology directly to its commercial partners.
Central to this offering is the NVLink Fusion chiplet, introduced by NVIDIA last year. Acting as a fabric bridge, the chiplet integrates NVLink support via a UCIe-attached module that delivers full-chip NVLink fabric capabilities. Rather than licensing the underlying IP, NVIDIA keeps the core NVLink components as a proprietary black box to be integrated, simplifying implementation for third-party designers. This architecture is essential for enabling custom XPUs to connect seamlessly to NVLink networks and NVIDIA’s rackscale hardware ecosystem.
The agreement also encompasses NVLink-C2C, NVIDIA’s short-range chip-to-chip interconnect technology. Primarily used to link custom CPUs with NVIDIA GPUs—as seen in the GB10—it can similarly interconnect multiple XPUs or disparate silicon dies. Additionally, MediaTek will distribute NVHBM, a custom memory protocol designed to interface processors with next-generation high-bandwidth memory. Unlike traditional implementations, NVHBM places the memory controller on the base die rather than integrating it into the processor itself.
Beyond NVLink Fusion, the partnership reaffirms ongoing collaborations on Spark computing devices and automotive initiatives. Crucially, the deal confirms that NVIDIA will not internalize the development of future Spark architectures. Instead, MediaTek will retain its developmental role across upcoming generations, including the Vera Rubin Spark and Rosa Feynman Spark platforms, which anchor NVIDIA’s public roadmap through 2030.
This transaction follows a broader pattern of strategic investments and partnerships initiated by NVIDIA. It signals that the company aims to transcend its identity as a mere hardware or intellectual property supplier, positioning itself instead as the foundational layer of the global AI infrastructure market. While the agreement introduces no previously undisclosed technologies, it represents a decisive push to standardize and widely distribute NVLink Fusion across third-party chip vendors.
Simultaneously, NVIDIA’s increasing allocation of profits toward equity investments in other companies underscores that the firm is not merely developing technologies and forging partnerships, but actively deploying capital to accelerate industry-wide developments. With nearly $60 billion in profit generated in its most recent fiscal quarter alone, the company possesses considerable financial leverage. This investment demonstrates a clear intent to put its resources behind its stated vision, actively bootstrapping advancements across the broader semiconductor and AI ecosystem.