There is a specific kind of capital allocation move that separates great businesses from merely profitable ones: spending money not to buy a company, but to make competition with you structurally harder. Nvidia just executed one at scale.
On August 31, Nvidia confirmed it bought $3.5 billion of convertible bonds issued by MediaTek, deepening a partnership centered on custom artificial-intelligence accelerators. The offering totaled $3.9 billion, the largest overseas convertible bond offering by a Taiwan-based company. Alphabet also participated. Nvidia, writing a check worth roughly 90% of the deal, is not a passive bystander here.
The structure deserves close reading. The investment is structured as convertible bonds, which can convert into MediaTek shares in the future rather than functioning as an immediate direct equity purchase. Zero coupon. No acquisition premium, no regulatory merger review, no hostile board dynamics. Nvidia gets equity upside tied directly to MediaTek’s success in AI chips, while keeping its balance sheet flexibility and avoiding the complexity of an outright buyout. Charlie Munger called the highest form of business intelligence knowing what you don’t need to own. Nvidia apparently agrees.
The technology layer underneath is where the strategic intent becomes clear. NVLink Fusion gives customers a foundation for building custom AI accelerators, combining NVLink connectivity, high-bandwidth memory, advanced packaging, and rack-scale integration, so engineering resources can focus on differentiated compute rather than surrounding infrastructure. In plain terms: hyperscalers that commission a custom chip through MediaTek will build it on Nvidia’s plumbing. MediaTek will offer NVLink Fusion to customers designing their own accelerators, putting Nvidia technology inside chips meant to replace its GPUs. The apparent threat becomes a tributary.
The platform is part of Nvidia’s broader strategy to remain central to AI infrastructure even as large customers move toward designing their own chips. The partnership will evolve from “Nvidia GPUs and MediaTek SoCs” to “Nvidia AI platforms and MediaTek customized XPUs,” leveraging Nvidia’s interconnect and networking architecture so customers can develop their own accelerators without building the infrastructure from scratch. That is a meaningful shift: the moat expands rather than contracts as custom silicon proliferates.
MediaTek is a major rival to Qualcomm and is not a minor partner. MediaTek received board approval on July 31, 2026 for a $5 billion financing plan and said it expects its data-center AI chip business to generate more than $2 billion in revenue in 2026. It has also talked about targeting roughly 10% to 15% of an approximately $70 to $80 billion AI data-center solutions market in the coming years. Shares of MediaTek closed 10% higher on September 1, after Nvidia’s announcement.
The risks are real. Critics argue the structure creates a self-reinforcing demand loop: Nvidia funds partners, partners build on Nvidia infrastructure, Nvidia books the resulting product revenue, and the cycle repeats. That circular concern has already surfaced around Nvidia’s other 2026 commitments. Neither company identified a customer developing an accelerator through the relationship or announced a deployment, which means the revenue thesis for this specific partnership has not been tested in silicon yet. And Broadcom remains outside the NVLink Fusion orbit for now.
Still, the logic holds. MediaTek’s adoption of NVLink Fusion enables it to design accelerators for Nvidia’s fully developed rack-scale platforms, while Nvidia gets a slice of the growing market of custom AI accelerators. The convertible structure means Nvidia profits if MediaTek’s AI chip business grows, which it will only do by deepening its dependence on Nvidia’s own interconnect standard. That is not circular financing. That is a well-designed flywheel, and history’s best capital allocators have always known the difference.