Jensen Huang spent August writing checks to the companies that are supposed to be killing his business.
On Monday, Nvidia said it would put $3.5 billion into convertible bonds issued by MediaTek, the Taiwanese chipmaker best known for the silicon inside half the world's smartphones. The sum represents nearly the whole of a record $3.9 billion offering, the largest convertible MediaTek has ever brought to market. MediaTek shares, already up close to 200% this year, jumped another 10% on the news, according to CNBC.
The money is not the interesting part. The mechanism is. As part of the deal, MediaTek will let customers use Nvidia's technology to design their own AI chips that connect directly to Nvidia's larger computing systems — a platform called NVLink Fusion that gives chipmakers ready-made connectors and specialized memory so custom AI chips can plug into Nvidia's larger data-center systems. In plain terms: Amazon, Google, Microsoft, OpenAI, and Anthropic are all racing to build their own AI accelerators so they need fewer Nvidia GPUs. Nvidia just paid the company most of them will hire to do that engineering — and made sure the resulting chips still have to plug into Nvidia's plumbing to work.
The circularity question, again
Huang has heard the word "circular" a lot this year, and he headed it off directly. Asked about the MediaTek stake on Bloomberg TV, he said "This is not circular because obviously they do their own business and we do our own business." The distinction matters because Nvidia has been on something of a spending spree across its own supply chain — the MediaTek deal follows Nvidia's up to $105 billion guarantee earlier this month for OpenAI's Ohio data-center lease, and could intensify scrutiny of arrangements where the chipmaker funds companies that help drive demand for its AI products.
One portfolio manager put a finer point on the distinction than Huang did. Joe Tigay of the Rational Equity Armor Fund noted that financing a company building tools that extend your own architecture is different from financing a customer directly — but, he added, "it is still Nvidia using its balance sheet to accelerate ecosystem growth." That is a generous read of a maneuver that looks, from a distance, like paying a rival contractor to build a bridge that only leads back to your own house.
Nvidia's own executives are unbothered by the framing. Dion Harris, the company's senior director of HPC and AI hyperscaler infrastructure solutions, told reporters on a call Monday that "Nvidia is an AI infrastructure company." Not a GPU vendor — an infrastructure company, full stop, regardless of whose logo is stamped on the silicon doing the computing.
MediaTek's second act
For MediaTek, the deal is validation of a pivot that started years before anyone outside Taipei was paying attention. The company has been slowly building its custom data center ASIC operations, saying in June that it expects the business to generate $2 billion in revenue in 2026, and hopes to target more of the market in the coming years. Specifically, MediaTek is targeting up to 15% of an $80 billion data-center market segment next year — which works out to as much as $12 billion, a sixfold jump from this year's guidance. CEO Rick Tsai was careful not to overpromise on timing, telling reporters he expects the AI chip business to accelerate but conceding "it will take a little bit of time."
The company is not doing this alone in the room. MediaTek's investment is part of a record $3.9 billion overseas convertible bond offering, and the Taiwanese chipmaker said Alphabet, a key AI infrastructure partner, also took part, though it did not disclose the size of the search giant's investment. That puts two of the largest forces in AI computing — one that sells chips, one that increasingly wants to stop buying them — inside the same financing round for the company designing the alternative.
The move also has a target on its back that has nothing to do with Nvidia. Hyperscalers and other tech companies are increasingly building custom chips to run AI workloads, and MediaTek is emerging as a player that can build these, posing a potential challenge to Broadcom, a market leader. Broadcom has spent years as the default partner for hyperscaler silicon programs; MediaTek, with Nvidia's balance sheet behind it, just became a credible second option.
None of this changes what the hyperscalers are actually trying to do, which is spend less on Nvidia GPUs over time. What it changes is who profits along the way. If a cloud giant swaps Nvidia chips for a MediaTek-designed accelerator, Nvidia still collects on the interconnect, the memory architecture, and now, thanks to the convertible bond, a stake in MediaTek's upside. It is a hedge disguised as a partnership — and if custom silicon really does eat into GPU sales the way Big Tech hopes, Nvidia will have made sure it still gets a toll on every chip that drives past.
