Nvidia Put $3.5 Billion Into MediaTek To Get Inside The Chips Built To Replace It

In this storyNVIDIAMediaTek

Nvidia has bought $3.5 billion of convertible bonds issued by MediaTek, the companies confirmed on 31 August. The purchase covers close to 90 per cent of a $3.9 billion offering that also drew Alphabet. MediaTek called it the largest overseas convertible-bond issue in Taiwan's capital-market history.

Alongside the capital, MediaTek will adopt NVLink Fusion as the design foundation for the custom AI accelerators it builds for other companies. The platform bundles three pieces: the NVLink Fusion chiplet, which connects an accelerator to Nvidia's scale-up fabric; NVLink-C2C, which links those accelerators to Nvidia Rosa CPUs and other compatible processors; and NVHBM, which moves the memory controller from the logic die onto the memory die to raise bandwidth and free silicon area for compute.

The practical effect is that a customer designing its own AI chip has to build only the compute core. Packaging, memory architecture, interconnect and rack integration arrive pre-qualified from Nvidia and MediaTek. Those layers involve high-speed SerDes, HBM integration, multi-die packaging and thermal qualification, which are among the hardest parts of the work.

The partnership also spans client PCs, where the two are collaborating on RTX-based systems, and automotive, where MediaTek's Dimensity Auto platform already uses Nvidia technology and can integrate with DRIVE.

Nvidia announced a comparable NVLink Fusion arrangement with Amazon Web Services last week, without an investment attached. AWS is separately deploying two million additional Nvidia GPUs.

Neither company has published a shipping timeline for MediaTek-designed NVLink Fusion chips, named the hyperscalers lined up as design partners, or disclosed the bonds' conversion price and maturity.

Read this against what Nvidia did a week ago and the strategy comes into focus.

Nvidia paused parts of its AI Compute Partnership in late August, the financing programme under which it guaranteed revenue floors for smaller cloud operators in exchange for a share of their income and, reportedly, a say in who could lease the guaranteed capacity. Its own employees raised antitrust concerns. That programme tried to secure Nvidia's position through contractual control over what customers did downstream, and that is exactly the kind of arrangement a regulator can read off a page.

The MediaTek deal achieves something comparable through architecture instead. No clause is required. If the custom accelerators hyperscalers commission are built on NVLink Fusion, NVLink-C2C and NVHBM, then the chips designed to reduce dependence on Nvidia GPUs ship with Nvidia's fabric, interconnect and memory design inside them. Ryan Shrout of Shrout Research put the economics plainly: Nvidia collects rack-scale adoption even on the designs where it loses the compute socket.

Architecture is a better moat than contract because it does not need enforcing. A company that has standardised on an interconnect does not switch it the following year, and nobody has to be told what they may or may not sell.

MediaTek is the specific choice that makes this work.

It is a Google TPU design partner. Its first hyperscaler accelerator enters mass production this quarter. In late July it raised its 2027 target to 15 to 20 per cent of a custom AI chip market it sizes at roughly $80 billion, backed by a $5 billion financing plan approved that month. The companies buying MediaTek's design services are, almost by definition, the hyperscalers working to reduce their reliance on Nvidia GPUs.

Nvidia has therefore funded the firm arming its own defectors, and ensured the weapons use its ammunition. That is the whole trade, and it costs less than two weeks of Nvidia's revenue at its current run rate.

The circularity deserves naming. Nvidia has spent several years investing in companies whose spending flows back into its ecosystem, and the return here is not the bond coupon. Alphabet took part of the same offering while being both a MediaTek accelerator customer and an Nvidia customer. Everybody is on every side of this, which is what a mature platform economy looks like and also what makes it difficult for anyone to leave.

For India there is a specific lesson in where the value settled. The compute die, the thing the design-linked incentive scheme is aimed at, is the part being commoditised, because hyperscalers are designing their own. What proved defensible was the layer beneath it: interconnect fabric, memory architecture, packaging and rack integration. Those are systems problems demanding long-horizon capital and physical qualification infrastructure, and they are precisely what a country with $162 million of total chip-design funding since 2023 is least equipped to attempt.

The honest caveat is that none of this has shipped. There is no timeline, no named design partner and no disclosed conversion terms. What has been announced is an intention, backed by real money, to make the alternative to Nvidia contain Nvidia.

For daily, sharp analysis of the biggest moves in the Indian business and startup ecosystem, follow StartupFox.