The going rate for a seat at the custom AI chip table is now three and a half billion dollars. NVIDIA has bought $3.5 billion of convertible bonds in MediaTek as part of an expanded partnership announced Monday, as reported by SiliconANGLE, planting its flag inside the Taiwanese company that designs processors for much of the world's phones and, increasingly, other people's AI accelerators.
A convertible bond, for anyone whose eyes glaze at the term, is a loan that can transform into shares later, which makes it a polite way to take a stake without quite taking a stake. NVIDIA has developed a habit of these arrangements, per the report, having bought $2 billion of shares in Marvell, a MediaTek competitor, earlier this year. When one company keeps buying into every firm that helps others design AI chips, the investments start to read less like a portfolio and more like a perimeter.
The substance of the deal runs through MediaTek's custom silicon business. Companies that hire MediaTek to design an AI accelerator pick interconnect options from a catalog, the technologies that let chips talk within a rack and across racks, and MediaTek will now offer NVIDIA's NVLink Fusion as one of those options, per the report. NVLink, the underlying interconnect, moves data more than ten times faster than a standard PCIe 6.0 implementation, according to NVIDIA's own comparison as carried in the coverage.
Readers who followed last week's news will recognize the pattern. NVIDIA recently extended NVLink Fusion with NVHBM, which relocates the memory controller into the memory stack itself, and Amazon's chip division signed on first. Add MediaTek's design catalog and the strategy becomes legible: whoever a hyperscaler hires to build its custom chip, the connective tissue on offer is increasingly NVIDIA's, which means the alternative to buying NVIDIA processors is buying processors stitched together with NVIDIA technology.
There is a consumer thread too. MediaTek helped design the chip inside NVIDIA's DGX Spark developer workstation, and the pair collaborated on the RTX Spark processor for laptops and desktops that arrived in June, per the report. The companies now plan to work together across multiple future generations of both lines, which quietly makes MediaTek NVIDIA's co-architect for putting AI silicon in personal machines.
What MediaTek gets is obvious: capital, an anchor relationship, and a differentiated catalog. What NVIDIA gets is subtler and better. The custom chip wave was supposed to be the industry's escape route from dependence on one vendor, and with each partnership, investment and interconnect license, that route runs a little more directly through the vendor everyone was escaping. The bet here is that if you cannot stop customers from designing their own chips, you can still own the roads between them.
Leave a Reply