Nvidia Bought 90% of a Rival's Bond Sale to Get NVLink Inside

Nvidia disclosed on August 31 that it had purchased $3.5 billion of convertible bonds issued by MediaTek — roughly 90% of the Taiwanese chip designer's record $3.9 billion offshore offering. Alphabet also participated, without disclosing its size. MediaTek shares rose about 10% on the news. It is the largest direct investment Nvidia has made outside the United States.

The bonds are zero-coupon. MediaTek pays no interest. Nvidia's return comes only if MediaTek's stock clears a preset strike and the bonds convert to equity — upside without an acquisition, and without a regulatory review of one.

The centerpiece of the expanded partnership is MediaTek adopting Nvidia's NVLink Fusion interconnect.

What NVLink Fusion is for

This is the part that explains the price.

NVLink is Nvidia's high-bandwidth interconnect — the fabric that lets many accelerators in a rack behave as one large accelerator. It has historically been closed, and it is a substantial part of why Nvidia's rack-scale systems outperform assemblies of comparable individual chips. NVLink Fusion opens that fabric to third-party silicon: it lets other companies design custom AI accelerators — XPUs — that plug directly into Nvidia's rack architecture.

MediaTek adopting it means MediaTek's customers can build custom accelerators that live inside Nvidia AI factories rather than replacing them.

That is a precise response to the largest structural threat to Nvidia's position. Broadcom has built a business co-designing custom XPUs for hyperscalers and frontier labs who concluded that buying merchant GPUs at Nvidia's margins does not scale — Google's TPUs, OpenAI's Broadcom program, and a guidance curve that runs to $230 billion of AI revenue by fiscal 2028. Every one of those chips is a chip Nvidia did not sell.

Nvidia cannot stop custom silicon. What it can do is make sure custom silicon still racks into Nvidia's system, runs on Nvidia's fabric, and leaves Nvidia collecting on the networking, the software stack, and the platform even when the accelerator says someone else's name.

MediaTek is a good vector for that. It has deep relationships across the Asian design ecosystem, serious ASIC capability, and a customer base that overlaps with exactly the mid-tier and regional buyers most likely to be shopping for a Broadcom alternative. Buying 90% of its bond offering is a very direct way to make sure MediaTek's XPU roadmap points at NVLink rather than away from it.

The circular financing objection

Analysts flagged this immediately, and the objection is not frivolous.

The pattern across the AI buildout has become recognizable: Nvidia invests in, guarantees, or takes equity in the companies that buy or enable its products, which strengthens their ability to buy or enable its products, which shows up as Nvidia revenue. Nscale takes a stake in Figure while Figure commits $3.5 billion to Vera Rubin compute. Nvidia backstops large infrastructure buildouts. Nvidia buys MediaTek paper while MediaTek adopts Nvidia's interconnect.

The critique is that this makes Nvidia's financial results partly a function of an ecosystem Nvidia is capitalizing — that demand which looks organic is at least partly vendor-financed, and that the whole structure is more reflexive and more fragile than a straightforward supply-demand read suggests.

The defense is that this is normal for platform businesses at inflection points, that the underlying compute demand is unambiguously real and unmet, and that a zero-coupon convertible in a profitable public company with a 10% price pop is a genuinely different risk than vendor financing an unprofitable customer's purchase order. MediaTek is not a startup being propped up. It is a large, established designer that would have raised this money regardless; Nvidia bought the paper to buy a seat.

Both readings are defensible. What is not defensible is treating the arrangement as arm's-length. It is not, and nobody is pretending it is.

The second half of the deal

Beyond data center silicon, the two say they will work on "local AI computing" — chips for PCs, and systems for cars.

That is a smaller line in the announcement and a large strategic surface. MediaTek is a leader in mobile and edge SoCs, an area where Nvidia has repeatedly tried and repeatedly failed to establish durable position. Automotive compute in particular is a market Nvidia has invested in for a decade with real but bounded results, and MediaTek already sits inside a substantial share of the world's shipping devices.

If NVLink Fusion is the defensive half of this deal, edge and automotive is the offensive half — access to volume markets Nvidia has not been able to reach with its own silicon.

What to watch

Whether other XPU designers adopt NVLink Fusion. MediaTek is one. The strategy only works if it becomes the default plug for custom accelerators, and that requires designers who are not being paid $3.5 billion to consider it.

Whether Broadcom's customers change behavior. Broadcom's guidance assumes custom silicon displaces merchant GPUs. Nvidia's bet is that it complements them instead. The next several quarters of Broadcom disclosures against Nvidia's data center revenue are where that resolves.

Whether the convertibles convert. They are zero-coupon, so if MediaTek's stock does not clear the strike, Nvidia has made a $3.5 billion interest-free loan to secure an interconnect partnership. That would still arguably be worth it — which tells you what Nvidia thinks NVLink Fusion adoption is actually worth.

Nvidia did not buy a rival. It bought a rival's willingness to build inside Nvidia's rack, and paid in a currency that costs nothing unless the bet works.