Nvidia is investing $3.5 billion in MediaTek and opening its NVLink Fusion ecosystem to the chip designer, which will keep building custom data-center chips for cloud companies and AI labs — now with Nvidia's interconnect and rack-level architecture inside them. The deal arrives as Amazon, Google, Microsoft, OpenAI and Anthropic all develop their own chips specifically to reduce their dependence on Nvidia GPUs. MediaTek said in June it expects $2 billion in revenue from custom server ASICs in 2026. Nvidia has just paid more than that entire expected year for a position in the business.
NVLink Fusion is the substance of the arrangement, and the money is the fee for entry. NVLink is Nvidia's technology for moving data quickly between chips, including chips Nvidia did not make. Extending it to MediaTek means that when a cloud provider commissions a custom accelerator, the thing around that accelerator — the fabric, the rack, the surrounding platform — can still be Nvidia's. Dion Harris, Nvidia's senior director of high-performance computing solutions and AI infrastructure for cloud providers, said the company has long operated not only as a maker of compute chips but as a company that builds infrastructure for AI. That claim usually reads as marketing. Here it is the business model, stated plainly.
Harris said cloud providers and model builders already run Nvidia's platform in a range of configurations, and that MediaTek will be able to offer its customers compatible rack-level infrastructure so they can standardize their AI data centers and place their own chips alongside Nvidia's on a single platform. The partnership, he said, should open NVLink Fusion to MediaTek's entire customer base.
The structure also follows a pattern Nvidia has used before: fund partners whose products then flow back into its own ecosystem. The investment is not charity and not diversification. It buys alignment.
Compare it with last week's announcement. Nvidia struck a similar partnership with Amazon Web Services — AWS will install 2 million more Nvidia GPUs and also integrate NVLink Fusion — and no direct investment was involved. The difference between the two deals is the more interesting part of this week. AWS is a customer; its incentive to standardize on Nvidia's fabric is already there, because most of the silicon in question is Nvidia's. MediaTek is not a customer. It is the company other people hire to build the chips that displace Nvidia's. Alignment there has to be bought, and $3.5 billion is what it cost.
Alongside the data-center work, the two companies will continue collaborating on Nvidia products that draw on MediaTek's experience in smartphone, smart home, automotive and wireless silicon: DGX Spark, Nvidia's compact desktop machine for AI developers; RTX Spark, its effort to push AI capability into consumer PCs; and automotive platforms for software-defined AI vehicles. MediaTek's car platforms already use Nvidia RTX graphics technology for intelligent cabins and work with Drive AGX, Nvidia's compute platform for autonomous driving. Jensen Huang, Nvidia's founder and chief executive, said AI is changing every computing platform, from the largest AI data centers to personal computers and cars, and that the two companies are building platforms to carry Nvidia's accelerated computing into new markets and to develop specialized AI systems at scale.
My read is that this is a concession dressed as an expansion. Nvidia's public position has been that general-purpose GPUs win on flexibility and software. Paying $3.5 billion to sit inside a merchant ASIC designer is an admission that a meaningful share of AI compute is going to be custom silicon regardless, and that the sensible response is to own the connective tissue rather than contest every socket. It is a good move. It is also a defensive one, and the list of in-house chip programs — Amazon, Google, Microsoft, OpenAI, Anthropic — is the reason it was necessary.
The announcement is quiet about two things. MediaTek has not said which customers want its custom AI chips, which makes the strategic value of the deal hard for anyone outside the two companies to price. And nobody has addressed the obvious tension in it: a buyer commissioning a custom chip to reduce its dependence on Nvidia now has the option of building that chip around Nvidia's interconnect, ecosystem and rack. Some will take it, because standardization is genuinely cheaper. Others will read it as trading one dependency for a quieter one. Nvidia's $3.5 billion is a bet that enough of them choose the first, and that being the fabric is durable in a way that being the accelerator may not be.