tech-ai

Nvidia’s $3.5B MediaTek Bet on Custom AI Chips

Nvidia isn’t scared of customers building their own AI chips. It just spent $3.5 billion making sure those chips still need Nvidia to work properly.

Nvidia’s $3.5B MediaTek Bet on Custom AI Chips

Nvidia is not defending its AI empire by selling more chips. It is buying a seat inside the chips meant to replace it.

On August 31, Nvidia put $3.5 billion into MediaTek convertible bonds and expanded a partnership that lets MediaTek build custom AI processors designed to plug into Nvidia’s NVLink-connected data-centre systems. Most people will read that as another giant AI investment. They are missing the actual play.

Jensen Huang is turning Nvidia from the bloke who sells the shovels into the owner of the road, the tollbooth and half the machinery on the mine site.

The $3.5 billion cheque is not the story

MediaTek is a serious chip business, not a logo Nvidia has slapped on a press release. It has deep experience in system-on-chip design, connectivity, power efficiency and advanced packaging. Nvidia has the accelerated-computing hardware, networking and software stack that dominates the AI data centre.

The announced deal joins those capabilities across three areas: AI infrastructure, local AI computing and automotive. The headline item is MediaTek’s adoption of Nvidia’s NVLink Fusion platform.

That name sounds like peak Silicon Valley perfume. The business logic is much simpler.

The biggest cloud companies and frontier AI labs increasingly want custom silicon. Amazon has its own chips. Google has TPUs. Microsoft, Meta and others have every financial reason to reduce dependence on Nvidia GPUs, which remain enormously expensive and in constant demand.

The old assumption was that custom chips were Nvidia’s problem. If hyperscalers designed their own accelerators, they would buy fewer Nvidia chips. Fairly obvious.

Nvidia’s response is cleverer than trying to stop them. It is offering a way for custom processors — called XPUs in the company’s announcement — to live inside Nvidia’s rack-scale AI systems. MediaTek can help customers develop those chips, while Nvidia provides the high-speed interconnect and surrounding infrastructure.

In plain English: build whatever chip you like, mate. But if you want it to operate smoothly at enormous AI scale, Nvidia would prefer you use its plumbing.

That is what the $3.5 billion is buying: not just financial exposure to MediaTek, but a stronger incentive for another major semiconductor company to build around Nvidia’s architecture.

Nvidia is conceding the chip battle to win the system battle

This is the part founders and investors should pay attention to. Winning every component is nice. Owning the system is better.

Think about Apple. It does not merely sell you a phone. It makes money because the phone, operating system, services, developer ecosystem and accessories all reinforce one another. Leaving is possible. It is just annoying enough that most people don’t.

Nvidia has been building that kind of position in AI infrastructure for years. CUDA software helped make Nvidia GPUs the default choice for AI development. Its networking gear and NVLink technology help those GPUs behave as giant coordinated systems rather than a pile of very expensive computers.

Now it is offering a version of that connected architecture to customers who want custom chips.

It sounds counterintuitive only if you still think Nvidia is principally a GPU company. It isn’t. Not anymore. It is an AI-factory company.

That distinction matters because a custom chip is not especially useful on its own. At serious scale, the workload depends on memory, networking, software, cooling, power delivery, servers and the ability to coordinate an absurd number of processors without turning your data centre into a very costly toaster.

Nvidia wants to own the layer where all those decisions meet.

Reuters reported that Nvidia’s $3.5 billion investment is part of MediaTek’s record $3.9 billion overseas convertible-bond offering. Alphabet also participated, though MediaTek did not disclose Alphabet’s investment size. That tells you where the pressure is coming from: Nvidia’s customers and partners are not waiting politely for cheaper GPUs. They are building alternatives.

Nvidia has decided that the best defence is to make those alternatives compatible with Nvidia.

The uncomfortable circular-financing question

There is a less flattering way to describe this deal: Nvidia is funding companies that may create demand for Nvidia’s broader ecosystem.

That deserves scrutiny.

Reuters noted investor concerns about circular financing in AI. The concern is straightforward. If Nvidia puts capital into the companies building AI infrastructure or buying into its ecosystem, and those businesses then help sustain demand for Nvidia technology, how much of the demand is truly independent?

This is not an accusation that the MediaTek partnership is fake. MediaTek has real engineering capability, an existing business and clear reasons to work with Nvidia. But intelligent investors do not ignore incentives just because the press release says “long-standing partnership.”

Nvidia’s investment follows its earlier financial ties across the AI stack, including a $2 billion investment in Marvell connected to NVLink Fusion. Reuters also reported that Nvidia had provided a guarantee of up to $105 billion for an OpenAI data-centre lease in Ohio earlier in August.

Put all that together and the pattern is obvious: Nvidia is not just supplying the AI boom. It is helping finance the ecosystem that keeps the boom moving.

That can be brilliant. It can also become dangerous if capital starts chasing infrastructure because everyone assumes somebody else will eventually pay for the compute.

I have seen versions of this movie in business. When a supplier finances its customers, it can accelerate a category. It can also conceal whether the customer economics stand on their own two feet.

The question is never whether the technology is impressive. The question is whether the end customer produces enough value to justify the cost of the entire stack.

The overlooked winner may be the buyer, not Nvidia or MediaTek

Here is the contrarian angle: this deal may give the biggest AI buyers more bargaining power, even while it strengthens Nvidia.

If Nvidia had insisted on a closed world where customers could only buy its full stack, the hyperscalers would have fought harder to escape it. They have the money, talent and strategic need to do so.

By supporting a route for custom chips to connect into Nvidia’s systems, Nvidia offers them a middle ground. They can optimise a chip for their own workloads without having to rebuild every other part of the data centre from scratch.

That could lower switching costs at the chip layer while keeping Nvidia entrenched at the infrastructure layer.

For a hyperscaler, that is a better negotiating position. It can credibly say: “We have options.” For Nvidia, it is still a win if the alternative option drives demand for NVLink, networking, software and system integration.

This is what strong platform businesses do. They make room for a limited amount of competition where it is least damaging, so they can protect the layer with the real economics.

MediaTek also gets a meaningful strategic upgrade. TechCrunch reported that it expected its custom data-centre ASIC business to generate $2 billion in revenue in 2026. Plugging its designs into Nvidia’s ecosystem gives MediaTek a cleaner pitch to customers who want bespoke silicon but do not fancy gambling their entire AI roadmap on an untested standalone architecture.

The winners may be the companies that can offer customers speed, compatibility and a credible exit option all at once.

What this means for you

If you run a business, do not copy Nvidia by throwing $3.5 billion at a partner. Obviously. Copy the underlying strategy.

First: identify the layer of your market that customers cannot easily replace. It is rarely the shiny product feature. In AI, Nvidia is betting the durable layer is the system connection between chips, software and data centres. In your business, it might be distribution, workflow, trusted data, payments, compliance or customer relationships.

Second: stop treating every customer request for customisation as a threat. Sometimes customers building their own solution is inevitable. Your job is to make sure their version still runs through something you own. Build the integration, marketplace, standard or service layer that lets them customise without leaving you entirely.

Third: test whether growth is real before celebrating it. If a supplier incentive, investor cheque or discount is carrying a deal, ask whether the customer still wins when the sugar hit disappears. Revenue that needs permanent subsidies is not a moat. It is a bill arriving late.

Finally: own the bottleneck, not merely the product. Products get copied. Features get commoditised. The business that controls how everything connects, moves or gets approved has the better chance of making money for a very long time.

Nvidia’s $3.5 billion MediaTek deal is not a bet that custom AI chips will fail. It is a bet that they will happen — and that Nvidia can make itself impossible to avoid when they do.

That is a far more dangerous business model than simply selling the fastest chip in the room.

Sources