Nvidia is investing $3.5B in Taiwanese chipmaker MediaTek in a deal that gives MediaTek access to Nvidia's NVLink Fusion ecosystem and lets it design custom AI chips that plug directly into Nvidia-based data centers. The arrangement is Nvidia's clearest answer yet to the custom-silicon push from Amazon Web Services, Google, Microsoft, OpenAI, and Anthropic, all of which are building in-house chips to reduce reliance on Nvidia GPUs. Rather than fight the trend, Nvidia is trying to host it.
MediaTek has been quietly scaling a custom data center ASIC business, and said in June it expects that unit to generate $2B in revenue in 2026. The Nvidia investment is designed to accelerate that ramp while keeping any resulting chips inside Nvidia's rack-scale architecture. NVLink, the high-speed interconnect that lets even non-Nvidia chips communicate at data center speeds, is the technical hook.
The strategy is a variant of Nvidia's now-familiar pattern: invest in the companies that most threaten to route around it, then re-anchor them to its stack. Last week, Nvidia announced that AWS will deploy an additional 2M Nvidia GPUs and integrate NVLink Fusion — a partnership without direct investment but with the same architectural end state. Between AWS and MediaTek, the message is that even hyperscaler custom silicon can live inside an Nvidia-defined AI factory.
Key facts
- 01Nvidia is investing $3.5B in Taiwan's MediaTek and granting access to its NVLink Fusion ecosystem for custom chip design.
- 02MediaTek expects its custom data center ASIC business to generate $2B in revenue in 2026.
- 03The deal follows last week's AWS agreement to deploy an additional 2M Nvidia GPUs and integrate NVLink Fusion.
- 04MediaTek and Nvidia will extend collaboration on DGX Spark, the new RTX Spark consumer AI PC, and Nvidia Drive AGX for vehicles.
For MediaTek, the appeal is customer access. The company hasn't disclosed which cloud providers or AI labs it's building custom ASICs for, but the NVLink Fusion tie-in lets any customer standardize on Nvidia's rack platform while running their own accelerators alongside Nvidia GPUs. That collapses one of the main headaches of custom-chip programs: designing a full system around a bespoke die.
Dion Harris, Nvidia's senior director of HPC and AI hyperscaler infrastructure solutions, framed the deal on a Monday call with reporters as an ecosystem extension rather than a defensive move. He described MediaTek's role as offering a path for its customers to "standardize on the rack-scale infrastructure across their AI factories" while deploying custom chips on the same standard platform. "This is really about opening up this ecosystem to the entire MediaTek customer base," he added.
The partnership goes beyond data center silicon. MediaTek and Nvidia will continue to collaborate on DGX Spark, Nvidia's small developer-focused desktop AI computer, and are extending the arrangement to RTX Spark, Nvidia's push to embed its AI hardware in consumer AI PCs. MediaTek's expertise in low-power designs for smartphones and smart homes is a natural fit for that form factor.
The two companies are also continuing joint work on automotive platforms, with MediaTek's cockpit systems using Nvidia RTX graphics and Nvidia Drive AGX handling autonomous driving workloads. The auto tie-in matters strategically because it locks in a second high-volume market — cars — where Nvidia wants its stack to become the default.
The circularity of Nvidia's financing has drawn attention across the industry. The company has repeatedly invested in firms whose spending flows back into Nvidia's own ecosystem, whether through GPU purchases, platform licensing, or, as with MediaTek, adoption of Nvidia interconnect and rack designs. The $3.5B commitment fits that pattern cleanly.
The counterweight is straightforward: custom ASIC programs at AWS, Google, Microsoft, and the model labs exist precisely because those customers want to escape Nvidia's margins on training and inference. NVLink Fusion softens the escape, but it does not eliminate the underlying economics. If MediaTek's custom-chip revenue grows well beyond the projected $2B in 2026 and hyperscalers deploy those chips at scale, Nvidia will have traded some GPU volume for platform tax — a real trade, not a costless one.
The MediaTek deal signals that Nvidia has decided the fight is no longer chip-versus-chip but stack-versus-stack. As long as NVLink, DGX rack designs, and the CUDA-adjacent software layer remain the default connective tissue of an AI factory, custom accelerators built by MediaTek — or by anyone else — end up reinforcing the position Nvidia already holds. That is a more durable moat than GPU exclusivity, and it explains why Jensen Huang was willing to write a $3.5B check to a company that will now help competitors build alternatives to his own silicon.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
