Nvidia's $3.5B MediaTek deal: the alternative plugs in

Nvidia announced this morning that it invested $3.5 billion in convertible bonds issued by MediaTek, and that MediaTek will adopt the NVLink Fusion platform so its customers can build custom AI accelerators that drop into Nvidia rack-scale systems. Read that sentence twice. The custom-silicon path everybody was treating as the escape route from Nvidia pricing now has an official Nvidia on-ramp.

What actually happened

Per Nvidia's announcement, the deal spans three layers: AI infrastructure, local AI computing, and automotive. MediaTek adopts NVLink Fusion so customers can develop custom AI infrastructure and XPUs that interoperate with Nvidia's rack-scale systems. The two companies also continue working on multiple generations of Nvidia RTX Spark and DGX Spark PC chips.

The financial instrument matters. This is convertible bonds, not an equity stake and not an acquisition — Nvidia gets debt that can become ownership, and MediaTek gets capital without an immediate control change. TechCrunch reports MediaTek expects roughly $2 billion in custom data center ASIC revenue during 2026, and frames the move as Nvidia's answer to Amazon, Google, Microsoft, OpenAI, and Anthropic all building their own accelerators.

Jensen Huang's framing in the release is the tell: AI is transforming every computing platform, "from the world's largest AI factories to the PC and the car." Nvidia is not defending a chip. It is defending an interconnect.

Why AI infrastructure consolidation matters for your business

You do not buy GPUs. You buy tokens, and the price of tokens is set several layers above you by how much genuine competition exists at the silicon layer. That is the only reason this is your problem.

The bear case for your inference bill goes like this: custom accelerators were supposed to be the pressure that keeps per-token prices falling. If the practical way to ship a custom accelerator is to make it speak NVLink and sit in an Nvidia rack, then "alternative silicon" and "Nvidia ecosystem" stop being opposites. Competition at the chip level becomes competition inside one fabric. Prices still fall — they have all year — but they fall on somebody else's schedule.

The good news is that none of your defenses change, and they are cheap:

Keep an abstraction between your app and any model API. One module, provider-agnostic, that every call goes through. When a vendor reprices, you change a config value, not forty call sites. We have watched three model repricings this year alone; the teams that shrugged all had this.

Keep AI contracts short. Twelve months maximum on anything that prices tokens. The market is repricing faster than annual procurement cycles, and a multi-year lock-in at today's rates is a bet you have no reason to make.

Benchmark on cost-to-complete, not price per million tokens. A model at twice the sticker price that finishes the job in a third of the steps is cheaper. Instrument your own workloads and compare finished tasks, because the headline rate tells you almost nothing.

Stay boring about where inference runs. If your stack can point at a different endpoint in an afternoon, silicon-layer consolidation is a headline you read, not an invoice you pay.

Sources: NVIDIA Newsroom, TechCrunch.