tech-ai September 1, 2026 By Blake Colman

Nvidia’s $3.5B MediaTek Deal: Jensen Huang’s AI Chip Tollbooth

Nvidia is spending $3.5 billion to help customers buy fewer Nvidia chips. That sounds mad until you realise it is buying the road every rival chip will need to travel.

Nvidia is spending $3.5 billion to help customers buy fewer Nvidia chips. That sounds mad until you realise it is buying the road every rival chip will need to travel.

On August 31, Nvidia put $3.5 billion into convertible bonds issued by MediaTek and expanded a partnership built around NVLink Fusion, its technology for connecting custom AI chips into Nvidia-based, rack-scale data centres. This is not Jensen Huang waving goodbye to GPU sales. It is a very expensive way of making sure Nvidia still gets paid when the biggest customers decide they want to design their own silicon.

That is the real AI story today. Not another chatbot feature. Not another bloke promising artificial general intelligence by Christmas. The fight has moved underneath the software, into the plumbing.

Nvidia is selling the toll road, not just the trucks

For years, Nvidia’s business has looked beautifully simple: everyone wants AI, AI needs vast computing power, and Nvidia sells the GPUs and software stack that make it work.

But no customer enjoys being strategically dependent on one supplier forever. Amazon, Google, Microsoft and the major AI labs have every reason to pursue custom chips: lower cost for specific workloads, tighter control of supply, and a bit more leverage when negotiating with Nvidia.

The usual lazy read is that custom silicon threatens Nvidia. It does, at the level of the individual processor. If a cloud giant can replace some expensive Nvidia GPUs with its own chip, Nvidia loses some unit sales.

Huang’s answer is not to pretend the threat does not exist. It is to make the threat compatible with Nvidia’s system.

MediaTek will use NVLink Fusion to help hyperscalers, cloud providers and frontier model developers build custom XPUs that connect into Nvidia’s rack-scale AI factories. An XPU is simply the industry’s catch-all label for specialised processing hardware. The important bit is not the acronym. The important bit is the connection.

Modern AI infrastructure is not a box with one clever chip in it. It is a factory floor of chips, memory, networking, cooling, software and orchestration. A fast chip that cannot communicate efficiently with the rest of the system is like owning a Formula 1 engine with a shopping-trolley gearbox. Impressive on a brochure; useless where it counts.

Nvidia wants to own the gearbox.

Why MediaTek matters more than most people think

MediaTek is hardly a household name outside technology circles, which is precisely why this deal will be misunderstood. It is known largely for system-on-chip design in phones and connected devices, but it also brings experience in custom silicon, power efficiency, packaging, connectivity and large-scale chip integration.

That skill set matters because the AI market is splitting in two.

At one end are broadly useful, high-performance GPUs for training models and running a wide variety of workloads. Nvidia remains extraordinarily strong there.

At the other are purpose-built chips tailored to a customer’s own models, data centres and inference workloads. That is where the cloud giants and AI companies see an opportunity to cut costs and control their destiny.

Nvidia and MediaTek are now explicitly working across three fronts: AI infrastructure, local AI computing through RTX Spark and DGX Spark systems, and AI-powered vehicles. In plain English, they are trying to extend Nvidia’s relevance from the giant data centre to the desk and the car.

The $3.5 billion itself is also worth reading properly. Nvidia bought convertible bonds, not ordinary MediaTek shares. That gives it a financial claim with the potential to convert into equity later, while helping fund a partner that will build more products around Nvidia’s platform.

It is a partnership, a financing arrangement and a strategic defensive move wearing the same jacket.

The custom-chip rebellion is real

Let’s be clear: Nvidia did not do this because it is bored and has spare change in the couch.

The custom-chip push is real. The largest buyers of AI compute have the capital, engineering talent and workload scale to justify designing their own hardware. If you run enough inference every day, shaving a little power use or improving performance for one specific task can be worth a mountain of money.

Nvidia’s own announcement makes the logic plain. MediaTek’s customers can develop custom AI infrastructure that integrates with Nvidia rack-scale systems. TechCrunch reported that Nvidia made a similar technical partnership with Amazon Web Services the previous week, with AWS set to deploy an additional 2 million Nvidia GPUs while integrating NVLink Fusion.

That is the pattern: customers are allowed to build their own engines, provided the engine fits Nvidia’s chassis, networking and factory layout.

That is much smarter than trying to stop them. You cannot stop a customer with a $2 trillion market capitalisation, an army of chip engineers and a board that hates dependency. You can, however, make independence expensive, inconvenient and technically inferior unless it still works through your ecosystem.

This is how great platform businesses defend themselves. They stop insisting that every dollar must be spent on their original product. Instead, they make sure every alternative product creates demand for their standard.

Microsoft understood this with Windows. Apple understands it with the App Store and its hardware ecosystem. Visa does not need to own every shop; it needs the transaction to pass over its rails.

Nvidia wants NVLink Fusion to become that rail for serious AI infrastructure.

The overlooked risk: circular financing can hide a weak customer

There is a catch, and investors should not clap like trained seals just because the number has three commas in it.

Reuters noted that the MediaTek investment could add to concerns about circular financing in AI. That concern is fair. When a company funds partners, customers or infrastructure providers that then help expand demand for its technology, revenue can look more self-sustaining than it really is.

To be fair, not every strategic investment is dodgy. Nvidia is not writing a cheque to a random startup and calling it demand. MediaTek is a major chip designer, the bonds are a defined financial instrument, and the technology partnership has a clear industrial purpose.

But the question operators and investors should ask is brutally simple: would this project still exist if Nvidia did not help finance the ecosystem around it?

If the answer is yes, great. Nvidia is attaching itself to genuine demand and improving its position.

If the answer is no, then part of the apparent AI boom is a vendor using its balance sheet to manufacture tomorrow’s customer spending. That can work for a while. It is not the same thing as durable end-user economics.

There is another risk for Nvidia. Open systems can become genuinely open. Once Nvidia helps prove that custom chips can plug into a common rack-scale environment, some customers may gain more negotiating power than Nvidia intended. Standards create ecosystems, but ecosystems also create escape routes.

Still, I would rather own the standard than be the supplier pretending standards do not matter.

The contrarian angle: this is a concession, and that is why it is strong

The chest-thumping version of business says a market leader must crush every rival and defend every inch of margin. That is movie nonsense.

The stronger move is often a controlled concession.

Nvidia is effectively saying to its biggest customers: fine, build custom chips. We will even help a world-class designer build them. Just make them work best inside the infrastructure we define.

That is not surrender. It is the confidence to move one layer up the stack.

Founders get this wrong all the time. They fight to retain a low-value piece of a customer workflow because it produces revenue today, while ignoring the chance to own the system that governs ten times as much value tomorrow.

If your customer insists on doing part of the job themselves, you have two choices. You can sulk, try to lock them in and eventually get replaced. Or you can build the tools, integrations and standards that make their in-house solution dependent on your platform.

The second option is harder. It requires swallowing your ego and thinking in decades rather than quarters.

What this means for you

If you are a founder, do not copy the dollar figure. Unless you have $3.5 billion lying about, that would be ambitious even by Australian pub-chat standards. Copy the strategic move.

First, identify the thing customers are determined to bring in-house. Do not fight reality. Ask how you can become the layer they still need when they do it.

Second, obsess over interoperability. The winning product is not always the one that does every job itself. Often it is the product that lets everyone else plug in faster, switch less painfully and create more value around it.

Third, separate revenue from real demand. If your growth depends on rebates, financing, discounted implementation or heroic founder effort, label it honestly. A sale that needs you to bankroll the buyer is not worthless, but it is not ordinary product-market fit either.

And if you are an investor, look past the chip-count headlines. Ask who owns the bottleneck after customers customise the hardware. The value may shift from the processor to the network, the software, the deployment layer or the financing.

Nvidia’s $3.5 billion MediaTek deal is a reminder that the best businesses do not merely sell the gold rush gear. They own the narrow bridge every miner has to cross, including the ones trying to build their own shovel.