A smartphone chip designer best known for powering budget Androids is now building the custom silicon that could sit inside the world's largest AI (人工智能) factories. The shift says as much about how AI compute is being reshaped as about the company itself.
On August 31, 2026, NVIDIA and MediaTek announced a deepened partnership. The headline: MediaTek will adopt NVIDIA's NVLink Fusion platform to help hyperscalers, cloud providers, and frontier model developers build custom XPUs — application-specific accelerators — and wire them into NVIDIA's rack-scale AI systems. NVIDIA also agreed to invest US$3.5 billion in MediaTek-issued convertible bonds.
Key takeaways
- MediaTek will use NVIDIA's NVLink Fusion as a design foundation so customers can build custom AI accelerators (XPUs) that plug into NVIDIA NVLink-connected racks.
- NVIDIA is investing US$3.5 billion in MediaTek convertible bonds, described as its largest direct investment outside the United States.
- The collaboration spans three areas: AI infrastructure (custom XPUs), local AI computing (RTX Spark and DGX Spark PC chips), and automotive platforms.
- MediaTek is positioning itself against Broadcom and Marvell in the fast-growing AI ASIC (application-specific integrated circuit) market.
- The company has told analysts it expects meaningful AI chip revenue in 2026 and is targeting a multibillion-dollar ASIC business by 2027.
What NVLink Fusion actually does
Custom AI accelerators are not new — Google's TPUs and Amazon's Trainium are examples — but building one that scales to a full rack has historically meant reinventing interconnect, memory, packaging, and qualification from scratch. NVLink Fusion packages those pieces as a prebuilt, prequalified foundation: a chiplet for scale-up fabric, a die-to-die link (NVLink-C2C), and a memory architecture (NVHBM) that leaves more silicon area for compute.
MediaTek's role is to take a customer's XPU design and handle the surrounding engineering — connectivity, memory, advanced packaging, and rack-scale integration — so the customer can focus on the differentiated compute. In effect, NVIDIA supplies the highway and MediaTek supplies the vehicle-building expertise.
Why MediaTek, and why now
MediaTek is a fabless designer with deep system-on-chip (SoC) experience, strong power-efficiency credentials, and existing NVIDIA collaboration on the GB10 Grace Blackwell superchip used in DGX Spark. Extending that to data-center custom silicon is a logical stretch of capabilities the company already sells at volume in phones, PCs, and cars.
The market pull is the agentic AI trend: workloads that plan, reason, and call tools generate far more compute demand than simple chat. Hyperscalers facing that load want silicon tuned to their exact models rather than one-size-fits-all GPUs, and they want supplier diversity. That is exactly the opening Broadcom and Marvell already occupy, and the one MediaTek is now contesting with NVIDIA's ecosystem behind it.
The money and the math
The US$3.5 billion NVIDIA investment comes as part of a larger US$3.9 billion convertible-bond offering. On its earnings calls, MediaTek has guided that its AI datacenter ASIC business could reach around US$2 billion in revenue in 2026 and a range of roughly US$7 billion to US$12 billion in 2027 — figures the company presented as targets, not guarantees, and which depend on customer tape-outs landing on schedule.
Those numbers matter because they reframe MediaTek from a consumer-chip supplier into a potential tier-one AI infrastructure vendor. Even partial success would diversify the custom-ASIC field beyond the current duopoly and, by extension, the pricing power concentrated there.
For NVIDIA, the deal is also defensive. By wrapping NVLink Fusion around custom XPUs, it keeps rivals' accelerators inside its own fabric rather than outside it. Every custom chip that adopts the platform extends NVIDIA's networking moat instead of bypassing it, which helps explain why the company put real capital — not just a press release — behind the relationship.
In practice, the arrangement also lowers the barrier for smaller cloud providers that cannot afford to design an entire rack from scratch. By reusing NVIDIA's interconnect and memory blueprint, they can field a differentiated accelerator without rebuilding the hard parts — a shift that could slowly widen who gets to sell AI silicon, even as NVIDIA supplies the rails underneath.
Honest limitations
Revenue and market-share figures for MediaTek's ASIC business come from the company's own guidance and secondary reporting; they are targets, not booked results, and were not independently verified. The strategic framing against Broadcom and Marvell reflects market commentary rather than disclosed customer wins. Specific customer names for custom XPU programs were not confirmed. This article does not assess the technical maturity of any MediaTek XPU design, which remains unannounced.
What readers can do now
- If you run AI infrastructure, track MediaTek's ASIC roadmap alongside Broadcom and Marvell as a signal of widening custom-silicon choice beyond general-purpose GPUs.
- Watch for disclosed hyperscaler design wins — that, not press announcements, will confirm whether NVLink Fusion custom XPUs reach production.
- For procurement planning, keep GPUs as the near-term default but ask vendors about ASIC roadmaps; supplier diversity is becoming a board-level risk question, not just a technical one.