NVIDIA has put $3.5bn into MediaTek convertible bonds and MediaTek is adopting NVLink Fusion as a design blueprint for custom accelerators. It matters less for what you buy this year than for what a rack looks like in three.
What was announced
- NVIDIA has invested $3.5 billion in MediaTek-issued convertible bonds as part of an expanded strategic collaboration.
- The partnership spans enterprise AI infrastructure, edge computing, client PCs and automotive platforms.
- MediaTek is adopting NVIDIA NVLink Fusion as a modular design blueprint so customers can build custom XPUs that drop into NVIDIA rack-scale environments.
- The NVLink Fusion framework covers NVLink Fusion chiplets, NVLink-C2C die-to-die connectivity, and NVHBM custom memory interfaces.
- MediaTek will handle packaging, physical implementation and supply-chain logistics for the resulting multi-die packages.
- The two companies previously collaborated on the GB10 Grace Blackwell Superchip used in NVIDIA DGX Spark, and are extending that work into client systems via NVIDIA RTX Spark.
The ETON view
The headline number is the least interesting part. What matters is that building a custom AI accelerator is being turned into a procurement exercise: pre-qualified interconnect, pre-qualified memory, a partner who owns the packaging, and a rack architecture that already exists. That lowers the barrier for large operators to build their own silicon, and it keeps all of it inside NVIDIA's rack and networking standards.
For hosting providers and enterprises the read-across is straightforward. Rack-scale standardisation on MGX and NVLink means the mechanical, power and network envelope of an AI rack is settling down, which is genuinely good news if you are planning a hall or a colo footprint: you can design to a known power and cooling profile rather than guessing. It also means the accelerator inside that envelope will get more varied over the next few years, and a buying decision made purely on brand will age badly. Our advice is unchanged. Design the rack, the fabric and the power to the standard, then keep the compute decision open and buy it on availability, lead time and price when you actually need it.
Related infrastructure
Category: AI & GPU · Vendor: NVIDIA, MediaTek · Technology: NVLink Fusion, NVLink-C2C, NVHBM, MGX, GB10 · Last verified: 02 Sep 2026 · ~2 min read
Sourcing this kind of infrastructure? Talk to ETON about availability, lead time and pricing.
