Nvidia’s $3.5B MediaTek Deal: Your ‘Custom AI Chip’ Still Pays Nvidia
Everyone says the AI chip market is opening up. Nvidia just spent $3.5 billion proving it can let competitors in without giving up the front door.
Nvidia has just put $3.5 billion into MediaTek, and the comforting story is that this means the AI chip market is finally becoming more competitive.
That is bollocks — or at least only half the story. Nvidia is not surrendering control. It is making sure that, even when customers build their own chips, they still need to rent the plumbing.
Nvidia didn’t buy MediaTek. It bought a seat in every future AI factory.
On August 31, Nvidia announced it had invested $3.5 billion in convertible bonds issued by Taiwan’s MediaTek. The investment sits inside MediaTek’s $3.9 billion overseas convertible-bond offering, which Reuters reported also drew participation from Alphabet, though Alphabet did not disclose its amount.
This was not a normal supplier partnership with a glossy press release and a few vague promises about “synergies.” Nvidia and MediaTek are deepening work across AI infrastructure, local AI computing and automotive platforms. Most importantly, MediaTek will adopt Nvidia’s NVLink Fusion platform, intended to help customers connect custom AI accelerators — often called XPUs — into Nvidia-connected, rack-scale AI systems.
That detail is the whole game.
The market likes to frame AI hardware as a cage match: Nvidia GPUs versus custom chips from Google, Amazon, Microsoft, Meta and anyone else with enough cash, engineers and electricity. But a data centre is not a shopping trolley where you throw in a few chips and head to checkout. The value increasingly sits in whether compute, networking, memory, software and power systems work together at scale without turning into an expensive bonfire.
Nvidia is betting that customers can build their own accelerator chips and still choose Nvidia’s rails to make the broader system work. That is a far stronger business than merely selling the biggest chip in the room.
The $3.5 billion is financing — but it is also strategy wearing a finance hat.
Convertible bonds are debt that can turn into equity under agreed terms. They let a company raise capital while giving investors potential upside if the share price rises. In MediaTek’s case, the structure matters because Nvidia is not simply buying a passive stake and hoping a Taiwanese chipmaker becomes more valuable.
It is funding a partner whose products will fit inside Nvidia’s ecosystem.
That is why people are using the phrase “circular financing” around these sorts of deals. The concern is straightforward: a dominant supplier invests in businesses that will, directly or indirectly, create more demand for its technology. The money can look like customer demand with a corporate logo slapped on it.
That concern should not be dismissed just because Jensen Huang is very good at selling a vision. If an AI company finances the builders, developers, cloud providers and infrastructure partners around its own products, investors need to ask a rude question: how much end-user demand is genuinely independent?
But calling every such deal circular and stopping there is lazy analysis.
Nvidia’s move has a real industrial logic. MediaTek has deep experience in system-on-chip design, connectivity, power efficiency and advanced packaging. It is a major smartphone-chip player that has been pushing harder into data-centre custom silicon. Nvidia has the accelerated-computing stack, network technology and software ecosystem. Put those together and you get an offer to hyperscalers that is much harder to ignore: build differentiated silicon if you want, but do it in a system that already speaks Nvidia.
That is not Nvidia defending a monopoly by refusing to cooperate. It is Nvidia turning interoperability into a tollbooth.
Why Wall Street shoved MediaTek up 10%
MediaTek shares jumped 10% after the announcement, adding roughly US$20 billion in market value in one session, according to Bloomberg reporting carried by the Taipei Times. That is a wild response to a financing deal until you understand what the market was really pricing.
Investors were not just cheering $3.5 billion of fresh capital. They were repricing MediaTek’s position in the AI hierarchy.
For years, much of the market treated chip companies as though there were only two categories: Nvidia at the top, and everyone else fighting for scraps. MediaTek’s deal suggests the next category may be more valuable: companies that can help giant customers design their own AI chips without asking those customers to rebuild an entire data-centre architecture from scratch.
That is a much better place to be than a generic chip supplier. It is also why Nvidia can afford to be generous. If MediaTek wins more custom-silicon projects and those projects connect through Nvidia’s technology, Nvidia has not lost a customer. It has expanded the size of its neighbourhood.
The overlooked point is that Nvidia is making a choice that many powerful incumbents fail to make. Rather than force customers into a binary decision — buy all of our hardware or become our enemy — it is offering a third path: customise what you want, but stay compatible with us.
That is how platforms survive when the market grows up.
The real boom is in the boring bit: the connective tissue.
Founders tend to obsess over the flashy bit of a technology stack. In AI, that is the model, the GPU or the chatbot demo that makes your aunt think you have invented Skynet.
The money, however, often pools around the boring bits that nobody can avoid: distribution, payments, standards, integration, workflow and infrastructure.
Nvidia understands this beautifully. A custom chip is only valuable if it can be deployed, networked, fed with memory, powered, cooled and programmed efficiently. If Nvidia’s interconnect and software become the default language those systems speak, it does not need to own every component to keep earning.
This is why the MediaTek deal matters beyond semiconductors. It is a reminder that a moat is not always the product. Sometimes the moat is the system that makes everyone else’s product useful.
Axios reported that companies had issued $186.8 billion globally in convertible bonds across 362 deals by September 1 — already more than in the whole prior year, based on Dealogic data. That tells you this is not a quirky one-off structure. Capital is being pulled toward AI businesses at speed, often on terms that reflect just how badly investors want exposure to the next infrastructure winner.
When money gets this easy, discipline gets harder. The winners will not be the companies that raise the biggest rounds or announce the most partnerships. They will be the ones that turn capital into an asset customers cannot easily replace.
The contrarian take: custom chips are not automatically bad news for Nvidia.
The standard bear case says custom chips will eventually gut Nvidia’s margins. There is truth in it. If the largest cloud companies successfully design, manufacture and deploy their own accelerators at scale, they will buy fewer premium Nvidia GPUs than they otherwise would.
But “fewer GPUs” is not the same as “less Nvidia.”
Nvidia’s answer is to become essential one layer higher up: the provider of the architecture, networking and software environment in which other chips operate. If it pulls that off, custom silicon may increase the total amount of AI infrastructure built while preserving Nvidia’s role at the centre of it.
That is the bet behind NVLink Fusion. Nvidia is effectively saying: go ahead, build your own engine. Just make sure the road, gearbox and service manual are ours.
There is risk in that strategy. Customers do not spend billions building custom hardware merely for the joy of being compatible. They want bargaining power. Over time, the biggest buyers will push to reduce dependence on Nvidia’s technology wherever they can.
Still, building a fully independent alternative is brutally difficult. It is not one chip. It is a full stack, a supply chain, an engineering culture and years of painful operational learning. I have built businesses, and I can tell you this much: most companies wildly underestimate the cost of owning more of the stack. Independence sounds sexy until you are the bloke paying for every outage, delay and design mistake.
What this means for you
If you are a founder, stop asking whether your product has a feature moat. Ask whether it can become part of the infrastructure other businesses need to operate. Features get copied. Systems that reduce complexity get embedded.
If you are an operator, look at your biggest supplier relationships. The dangerous question is not, “Can we get a better price?” It is, “Are we becoming strategically dependent without building any leverage?” Nvidia is showing what world-class leverage looks like. Your job is to avoid being on the wrong end of it.
If you are an investor, do not blindly celebrate every AI financing announcement. Follow the money one step further. Who gets paid? Who is funded? Who becomes more dependent on whose platform? The press release will call it an ecosystem. Sometimes it is. Sometimes it is a very expensive demand subsidy.
And if you run a company trying to build custom technology, remember this: owning the engine does not mean you own the road.
Nvidia’s $3.5 billion MediaTek deal is a warning shot. The next era of AI will not be won only by whoever makes the best chip. It will be won by whoever makes it hardest, slowest and most expensive for everyone else to leave.