NVIDIA Financed the Company That Could Replace It

Yesterday, August 31, NVIDIA and MediaTek announced an expansion of their long-standing partnership. The release talks about AI platforms from edge to cloud, software-defined cars, PC silicon. Buried in the middle is one line that changes what the announcement actually is: NVIDIA subscribed $3.5 billion in convertible bonds issued by MediaTek.

The full placement came to $3.9 billion, roughly NT$120 billion, and Alphabet took part as well. NVIDIA ended up with close to 90% of the issue. It is NVIDIA's first direct investment in a Taiwanese IC design house and, according to Focus Taiwan, the largest placement of its kind in MediaTek's history. The next day the stock jumped around 10%.

The rest of the deal is what you would expect: MediaTek adopts NVLink Fusion to design custom multi-die XPUs, adds the RTX Spark and DGX Spark chips for local-AI PCs, and extends the partnership into automotive. MediaTek CEO Rick Tsai spoke about making advanced AI computing pervasive. Jensen Huang said AI is transforming every computing platform, from the largest AI factories to the PC and the car. Press-release sentences. The $3.5 billion is not.

What NVIDIA actually bought

The biggest threat to NVIDIA is not AMD. It is the custom ASIC.

Amazon, Google, Microsoft, OpenAI and Anthropic have spent years funding their own silicon so they can stop buying GPUs at the price GPUs cost today. None of them designs that chip alone: they hire design houses. MediaTek is one of the few in the world that can do it at that scale.

A textbook incumbent would do one of two things: cut price, or try to buy the design house. NVIDIA did a third. It lent them money and put its own interconnect inside them.

That is the move. NVIDIA stopped being a GPU business a while ago; today it is the whole rack: the interconnect, the networking, the software, the integration. NVLink Fusion exists precisely so that an accelerator NVIDIA did not build can still sit inside an NVIDIA system. If the ASIC that displaces the GPU still has to speak that protocol to be useful, losing the chip sale stops being the same thing as losing the customer.

This is not an acquisition. It is a fence. And NVIDIA paid for it up front.

The uncomfortable part

Precision matters here, because the argument is already loud: NVIDIA did not buy MediaTek shares. It bought convertible debt. Right now it is a creditor holding an option, not an owner. That distinction matters the moment someone compresses this into NVIDIA bought 90% of MediaTek, which is not what happened.

That said, the pattern is the worrying one, and it is the same pattern already debated around NVIDIA's investments in its own compute customers: the supplier provides the capital that finances the purchase of its ecosystem. Nobody is inventing revenue. But when a growing share of the money moving through a sector comes off the balance sheet of its dominant supplier, measured demand stops being a clean signal. It is demand, yes. It is also paid distribution, in the shape of a bond.

What this means if you run a technology company

Three things, none of which requires $3.5 billion in the bank.

Defensibility is no longer exclusivity, it is the interface

What protects NVIDIA is not that nobody else can build an accelerator. It is that everyone else's accelerator has to speak its protocol to be worth anything. Scale that down to our size and the translation is literal: the most valuable part of your product is almost never the part that does the work, it is the part that defines how it connects to everything else. The data format, the API, the integration model. That is what survives when someone builds a cheaper alternative to your core module.

Funding the company that could displace you is usually cheaper than fighting it

It is counterintuitive, which is why almost nobody does it. A founder's natural reaction to an emerging competitor is to close doors on them; the expensive reaction is to compete on price. The smart reaction, when you can afford it, is to make them dependent on your standard and keep the layer where your margin does not erode.

Distrust demand your own vendor is financing

If your supplier is handing you the capital, the debt or the credit you use to buy from them, you have a commercial relationship and a financial one at the same time, and the second distorts how you read the first. It happens with the manufacturer financing its distributor, with the cloud provider giving away two years of credits, and with us when we accept a subsidized pilot and count it as market validation. It is not.

My read

What strikes me is how undefensive the move looks and how defensive it is. NVIDIA did not go out and argue that custom ASICs do not work. It assumed they will exist, that MediaTek will design them, and it bought the right to have them run through its cable.

That is incumbent maturity. Almost every large company I have watched lose its position did so by denying the substitution until it was too late. Very few sit down and calculate what it is worth to have the substitution happen inside their own house.

At Indrox we obviously do not move numbers like these. But the question scales down perfectly, and I ask it every time a client considers replacing a piece of what we built: am I defending the component, or am I defending the way everything connects? The value is almost always in the second one. And we almost always defend it worse.