Nvidia’s $3.5B MediaTek bet reveals its plan for tackling Big Tech’s AI chip buildout
Aug 31, 2026, 8:15 AM · TechCrunch
Nvidia is financing the custom-silicon market that could weaken its GPU dominance, provided those chips still speak Nvidia’s infrastructure language.
Why it matters
Nvidia is investing $3.5 billion in MediaTek, which will use NVLink Fusion to help customers build custom AI chips that can plug into Nvidia-based data centers. MediaTek said its custom data-center ASIC business could generate $2 billion in 2026.
The move comes as Amazon, Google, Microsoft, OpenAI and Anthropic pursue their own silicon. MediaTek will also continue work with Nvidia on DGX Spark, RTX Spark and software-defined vehicle platforms.
From the desk
Nvidia is not abandoning GPUs; it is trying to own the connective tissue around heterogeneous compute. If customers insist on custom silicon, NVLink, rack architecture and financing keep that choice inside Nvidia’s orbit. The source establishes the mechanics, but it does not settle the broader commercial outcome. The likelier read is that execution and adoption will matter more than the announcement itself.
Signal Desk's read: Nvidia is not abandoning GPUs; it is trying to own the connective tissue around heterogeneous compute. If customers insist on custom silicon, NVLink, rack architecture and financing keep that choice inside Nvidia’s orbit. The decision is strategically coherent, but its limits are visible. Readers should separate what is measured or committed from what the company hopes the system will enable.
The practical consequence is a shift in who bears the integration, governance or verification burden. That is where the story will become tangible for operators, developers and institutions.
Context
NVLink Fusion is Nvidia’s technology for connecting non-Nvidia chips to its scale-up and rack infrastructure. AWS announced a similar collaboration without the direct investment and plans two million additional Nvidia GPUs.
MediaTek’s customers for custom AI chips were not identified in the source, so the partnership’s eventual scale remains unproven.
Who feels it
- Hyperscalers
- They can pursue tailored accelerators without giving up Nvidia-based rack integration, while retaining a strategic dependency.
- Chip designers
- MediaTek gains a route into AI ASIC work but must prove it can deliver differentiated silicon at volume.
- Infrastructure teams
- Common networking may make mixed GPU and custom-chip fleets easier to operate.
What to watch
- Named MediaTek ASIC customers and 2026 revenue performance.
- A production custom chip using NVLink Fusion.
- Evidence that mixed racks beat GPU-only systems on cost or utilization.
Companies: NVIDIA