What Nvidia’s $3.5 Billion MediaTek Investment Means for Custom AI Chips

Nvidia is investing $3.5 billion in convertible bonds issued by MediaTek, deepening a partnership that now reaches from cloud data centers to desktop AI systems and vehicles. The most consequential piece is MediaTek’s adoption of Nvidia’s NVLink Fusion platform: customers can pursue custom AI accelerators while connecting them to Nvidia’s rack-scale infrastructure. 1

That arrangement addresses a central tension in AI hardware. Cloud providers want specialized chips tailored to their workloads, but Nvidia wants its networking, interconnect and system architecture to remain part of the data-center stack even when another company designs the main accelerator.

What the $3.5 billion investment entails

Nvidia is purchasing MediaTek bonds that can convert into MediaTek shares under the bond terms. The transaction is therefore both a financing arrangement and a strategic alignment between two chip designers; it is not the same as an immediate outright purchase of MediaTek. 111

The companies announced the investment alongside a broader collaboration covering three areas: AI infrastructure, local AI computing and automotive technology.

How NVLink Fusion connects custom XPUs to Nvidia systems

MediaTek will use NVLink Fusion to give hyperscalers, cloud-service providers and frontier-model developers a prevalidated path for developing custom XPUs—specialized AI processors—and integrating them into Nvidia NVLink-connected, rack-scale AI factories. 2

The model is semi-custom rather than completely separate from Nvidia’s platform. Nvidia’s NVLink Fusion ecosystem is intended to let custom silicon participate in the high-speed scale-up fabric used to connect processors across AI infrastructure. Nvidia has described the platform as a way for industry partners to build custom silicon while remaining part of the broader NVLink ecosystem.

The expanded technology stack also includes:

• NVLink Fusion chiplets and integration: These provide a route for custom accelerator designs to connect into Nvidia-based systems rather than operating as isolated ASICs. The available announcements describe the platform and partner ecosystem, but do not establish that a finished MediaTek XPU is already deployed in Nvidia racks.
• NVLink-C2C: This is a chip-to-chip interconnect used for closely coupled components. Nvidia cites it in its RTX Spark and DGX Spark systems to link CPU and GPU resources with a coherent memory model.
• NVHBM: Nvidia’s custom high-bandwidth memory technology is being extended to partners using NVLink Fusion. Nvidia and Amazon’s Annapurna Labs are working to support NVHBM alongside future Trainium chips, illustrating how a company can use its own accelerator while adopting Nvidia interconnect and memory technologies.

In practical terms, the appeal for a cloud operator is flexibility: it can design silicon for a particular workload while retaining access to Nvidia-oriented rack-scale connectivity. The caveat is important—these announcements describe an architecture and partnership path, not confirmed production volumes or a guarantee that every custom XPU will achieve the promised performance.

Why Nvidia wants custom chips to stay connected

Hyperscalers are investing in proprietary accelerators to tailor hardware to their workloads and infrastructure. Nvidia’s response is to make its interconnect and scale-up architecture useful even when the compute silicon is designed by a partner or by the cloud provider itself.

That strategy can expand Nvidia’s role beyond selling GPUs. If custom accelerators connect through NVLink Fusion and use related Nvidia system technologies, Nvidia may remain involved in the networking, memory and rack-level architecture surrounding those chips. MediaTek’s adoption gives Nvidia another design partner as it builds that ecosystem; MediaTek was also listed among the early companies adopting NVLink Fusion when Nvidia introduced the platform.

The AWS relationship shows the same direction. AWS and Nvidia announced plans to deliver two million additional Nvidia GPUs and next-generation AI infrastructure, while Amazon’s Annapurna Labs is extending NVLink Fusion support to Nvidia’s NVHBM technology for future Trainium systems.

The partnership also reaches PCs and workstations

The companies will continue collaborating on RTX Spark and DGX Spark-class systems. Nvidia’s RTX Spark materials describe a system combining an RTX GPU with a Grace CPU through NVLink-C2C, with up to 128GB of unified memory and up to one petaflop of AI performance.

For MediaTek, the relationship extends its platform and low-power system-on-chip expertise into local AI computing. For Nvidia, it helps carry the RTX graphics and AI software stack into compact personal computers and developer workstations. Those product claims describe Nvidia systems and the companies’ collaboration; they do not mean that the MediaTek investment itself finances a particular consumer product launch.

Automotive collaboration adds another market

The expanded agreement also covers MediaTek’s Dimensity Auto platform. Nvidia says the companies will integrate MediaTek’s automotive systems with Nvidia AI capabilities, RTX graphics and DRIVE AGX technology. The intended result is a broader platform for in-vehicle computing, including cockpit and AI-enabled vehicle applications.

As with the data-center work, the announcement establishes collaboration and planned integration rather than proving the commercial success or shipment scale of future vehicle programs.

MediaTek’s ambitions—and the competitive pressure

MediaTek is using the partnership as part of a much larger push into custom AI data-center chips. Reuters reported that the company approved a discretionary $5 billion financing budget for long-term growth, including its data-center AI expansion. MediaTek also raised its estimate for the 2027 custom AI-chip market to $80 billion and increased its target share to 15%–20%.

Those are MediaTek targets, not independently verified outcomes. The company has also said it expects more than $2 billion in 2026 data-center revenue, while reporting on the market describes competition involving established custom-chip suppliers such as Broadcom and Marvell.

NVLink Fusion could give MediaTek a differentiator: it can pursue custom accelerator design while offering customers a path into Nvidia-connected infrastructure. Whether that is enough to win major hyperscaler programs will depend on performance, cost, software support, manufacturing and deployment timelines—none of which is settled by the bond investment alone.

The bottom line

Nvidia’s MediaTek deal is best understood as an ecosystem bet. Nvidia is supplying capital and extending its platform so customers can build more specialized AI chips without necessarily leaving Nvidia’s interconnect and rack-scale environment. MediaTek gains a strategic technology partner as it tries to become a larger custom-ASIC supplier, while the two companies broaden their existing work in AI PCs and automotive computing.

The immediate fact is the $3.5 billion convertible-bond investment and the announced NVLink Fusion collaboration. The larger payoff—shipping MediaTek XPUs inside Nvidia-connected AI factories, reaching MediaTek’s revenue targets or changing the balance among custom-chip suppliers—remains to be demonstrated.