Nvidia is putting $3.5 billion behind Taiwanese chipmaker MediaTek in a deal that reveals how the world's most valuable semiconductor company plans to stay indispensable even as its biggest customers race to design their own AI silicon.

 

The partnership, announced Monday, centers on a $3.5 billion purchase of MediaTek convertible bonds — just under 90% of a roughly $3.9 billion bond offering — rather than a straightforward equity stake. The structure lets Nvidia potentially convert the debt into MediaTek shares later while sidestepping the regulatory friction that can accompany a direct acquisition of a foreign chipmaker. MediaTek shares jumped 10% on the news, extending a rally that has left the stock up roughly 200% for the year.

 

MediaTek, best known for smartphone processors and wireless chips, has been building a fast-growing sideline designing custom data-center ASICs — application-specific chips tailored to a single customer's workload. That business is projected to generate about $2 billion in revenue for MediaTek in 2026, and Nvidia's investment is a bet that it is about to get much bigger.

 

What's new

The core of the deal is technical rather than financial: MediaTek is being brought deeper into Nvidia's NVLink Fusion ecosystem, the interconnect platform Nvidia opened up earlier this year to let outside chipmakers build custom AI accelerators, or XPUs, that still plug directly into Nvidia's data-center systems. 

 

Jensen Huang, Nvidia's founder and chief executive, described the goal as letting partners "build semi-custom AI infrastructure, not just semi-custom chips" — meaning MediaTek's custom silicon would get access to Nvidia's rack-scale architecture, high-speed chip-to-chip links, and specialized memory, rather than functioning as a standalone rival platform.

 

In practice, that means hyperscalers and AI labs that want a cheaper, workload-specific chip than a general-purpose Nvidia GPU can now hire MediaTek to design one, safe in the knowledge it will still slot into their existing Nvidia-based infrastructure. 

 

The arrangement extends beyond the data center: the two companies are also deepening an existing collaboration on Nvidia's RTX Spark and DGX Spark desktop AI computers, and on automotive AI platforms built around MediaTek's Dimensity Auto chip line, with MediaTek chief executive Rick Tsai pointing to strengthened cooperation "across cloud, local AI, and automotive."

 

Why it matters

The deal is Nvidia's clearest move yet to hedge against the biggest threat to its dominance: Amazon, Google, Microsoft, OpenAI and Anthropic are all developing custom AI chips of their own, partly to cut costs and partly to reduce reliance on Nvidia.

 

 Rather than trying to stop that shift, Nvidia is positioning itself to profit from it by making sure any custom chip built by a partner like MediaTek still depends on Nvidia's interconnect standard and surrounding infrastructure.

 

Dion Harris, Nvidia's senior director of HPC and AI hyperscaler infrastructure solutions, framed the strategy plainly: "Nvidia is an AI infrastructure company. We expanded beyond pure computing chips years ago."

 

 Industry estimates cited alongside the announcement put the addressable market for custom AI chips at roughly $80 billion by 2027 — a pool of revenue Nvidia would otherwise watch flow entirely to rivals such as Broadcom and Marvell, which already design custom silicon for Google and Amazon.

 

The circular financing debate

The deal has also revived scrutiny of a pattern investors have flagged repeatedly in Nvidia's dealmaking this year. Critics note that Nvidia's convertible bonds appreciate in value largely if MediaTek's AI chip business succeeds — and that business succeeds largely by getting hyperscalers to adopt NVLink Fusion, the standard Nvidia itself controls. 

 

Bernstein Research analyst Stacy Rasgon said the structure "will clearly fuel 'circular' concerns," a reference to a broader debate over deals in which Nvidia's capital, directly or indirectly, ends up subsidizing demand for its own technology. Prominent short-seller Michael Burry was blunter, posting that the pattern amounted to "whistling past the graveyard" and pointing to Nvidia's own SEC filings as evidence of a self-reinforcing loop.

 

Skeptics point to a string of 2026 transactions as the pattern: a $30 billion equity investment in OpenAI in February, a $105 billion infrastructure guarantee tied to OpenAI's Ohio data center build-out in August, and now the MediaTek bond purchase at the end of August. 

 

Huang has dismissed the criticism, arguing the arrangement involves genuinely independent business operations and is standard practice in a capital-intensive industry; Nvidia and MediaTek also noted that Alphabet made an undisclosed co-investment in the same MediaTek bond offering, which they cited as external validation of the deal's merits. 

 

Critics counter that Alphabet, unlike Nvidia, has no direct revenue stake in whether MediaTek's chips run on Nvidia's infrastructure, making the two companies' incentives for investing very different.

 

What to watch next

The practical test of the deal will be whether hyperscalers actually commission MediaTek-designed XPUs at meaningful volume, and whether those chips ship with the promised full-bandwidth NVLink Fusion integration rather than a scaled-down version. 

 

Investors will also be watching how Nvidia accounts for the convertible bonds on its balance sheet and whether regulators in the U.S., Taiwan or elsewhere take a closer look at the arrangement given the circular-financing criticism already surrounding Nvidia's OpenAI commitments. MediaTek, for its part, will need to show that its custom-chip revenue can scale well beyond the $2 billion projected for this year without diluting its still-dominant smartphone and connectivity business.

 

For now, the deal underscores a broader shift in how the AI chip industry's biggest player is defending its position: not by out-designing every custom chip effort at hyperscale companies, but by making sure that whichever company wins that design contest still has to build on Nvidia's rails.