Deep Dive — Nvidia pays $3.5B for MediaTek's custom-silicon handshake

Nvidia announced on Monday that it is investing $3.5 billion in MediaTek via convertible bonds and putting MediaTek inside the NVLink Fusion ecosystem — the technical foundation behind Nvidia's push to let hyperscalers and frontier-model developers build custom AI accelerators that still plug into Nvidia's rack-scale fabric. The headline reads like a routine partnership upgrade. The substance is the opposite: Jensen Huang is paying real money to make sure the custom-silicon wave does not leave Nvidia behind.

The deal was framed by both companies as a deepening of an existing relationship. It is much more than that. As we covered last week in Nvidia's $6 billion plan to beat China at open weights, Nvidia is increasingly using its balance sheet to keep the frontier inside its own orbit — buying people, pipelines, and now chip partners. The MediaTek investment closes a strategic gap that the open-weights deal left open: even if Nvidia controls the software and the model layer, the actual silicon under a custom workload could be someone else's. The convertible-bond structure is the part of the story that does the most quiet work.

What the deal actually contains

Two distinct things happened on Monday, and they should not be confused. First, MediaTek will adopt Nvidia's NVLink Fusion platform — the open-standards approach Nvidia announced earlier this year to let third parties build custom XPUs that slot into NVLink-connected rack-scale "AI factories." Second, Nvidia is investing $3.5 billion in MediaTek-issued convertible bonds, which means the chipmaker can convert that debt into equity under defined conditions. Neither side disclosed coupon, maturity, or conversion terms in the announcement.

The NVLink Fusion piece is the technology story. NVLink Fusion is Nvidia's answer to a problem the hyperscalers have been solving on their own for years: how to build accelerators tuned to specific workloads — inference at extreme batch sizes, recommendation systems, mixture-of-experts routing — that still cooperate with a rack-scale fabric instead of fighting it. Google did this with TPUs behind its proprietary ICI fabric. Amazon did it with Trainium behind NeuronDevice. Meta is doing it with MTIA. Until now, those custom silicon programs all lived inside one company's walls; anyone building a custom accelerator for sale to multiple buyers had no good answer for scale-up networking, high-bandwidth memory integration, or multi-die packaging.

NVLink Fusion packages the surrounding scaffolding — the scale-up fabric, the chiplet interconnect, the memory integration, the chip-to-chip links, the CPU companion — and offers it as a prevalidated design. Nvidia announced the platform's general availability in May, alongside its own NVLink Fusion silicon and the next-generation NVHBM memory. The pitch to custom-XPU builders is that they no longer have to build the entire rack by themselves; they can focus on the accelerator die and let Nvidia handle the connective tissue. MediaTek is the first major customer to publicly commit to that path.

Why MediaTek, and why convertible bonds

MediaTek is an unusual choice. The Taiwanese company is best known as the world's largest supplier of smartphone application processors — the chips that run Android phones from the Chinese and global mid-tier — and as a major player in TV SoCs, Wi-Fi, and automotive infotainment. It does not have a long history in data-center accelerators. What it has is world-class capability in custom-silicon design, advanced packaging, high-speed SerDes, and power-efficient SoCs — exactly the building blocks needed to bring a custom XPU from tape-out to volume manufacturing.

Rick Tsai, MediaTek's vice chairman and CEO, made the strategic logic explicit in the announcement. "MediaTek and NVIDIA share a vision for making advanced AI computing pervasive across the technology landscape," he said. "By combining NVIDIA's leadership in accelerated computing and AI software ecosystem with MediaTek's expertise in a diverse AI technology portfolio from edge to cloud, and our leadership position in custom silicon, we can accelerate innovation for our customers." In plain terms, MediaTek is offering Nvidia a design house that can take a hyperscaler's custom-XPU specification and turn it into shippable silicon — including the packaging and the I/O, which are where most custom-accelerator programs stall.

The convertible-bond structure is more interesting than the headline price suggests. As we noted when we tracked Nvidia's broader balance-sheet strategy in Nvidia backs $500B data center deal with GPU value guarantee, the company has been using its roughly $48 billion in free cash flow to extend its reach without diluting its own equity or paying full acquisition premia. A convertible bond gives Nvidia two things cash does not. It locks in a senior claim on MediaTek's cash flow in the downside case, and it gives Nvidia an option — not an obligation — to convert that claim into equity if the partnership produces the kind of growth both sides are forecasting. The $3.5 billion figure is the size of the bet, not the size of the takeover.

Huang framed the deal in his own way: "MediaTek is one of the world's great semiconductor companies, with exceptional expertise in system-on-chip design, connectivity, leading performance and power efficiency. Together, we're building platforms that bring NVIDIA accelerated computing to new markets and give customers the freedom to create differentiated AI systems at enormous scale." Read between the lines: "differentiated AI systems at enormous scale" is a direct pitch to hyperscalers who have been told until now that buying Nvidia is the same as buying Nvidia. With NVLink Fusion and a MediaTek-built custom XPU inside the same rack, that story changes.

Three buckets, one strategy

The announcement spells out the partnership in three buckets: AI infrastructure (the NVLink Fusion work), local AI computing (RTX Spark and DGX Spark PC chips that pair Nvidia GPUs with MediaTek SoCs), and automotive (the software-defined vehicle platforms the two companies have been working on since 2023). Only the first is genuinely new. The Spark products are existing designs that Nvidia has been shipping for over a year; the automotive work predates this announcement. The novelty lives entirely in the data-center bucket, and the data-center bucket lives or dies on the custom-XPU thesis.

That thesis is structurally important to Nvidia, because the custom-silicon threat is real and growing. Hyperscalers have spent the last three years proving that carefully tuned inference silicon can deliver 30 to 50 percent better price-performance than general-purpose GPUs on specific workloads — Google's TPUs for training and inference, Amazon's Trainium for its own shopping and Alexa workloads, Microsoft's Maia chips, Meta's MTIA family. None of those programs has yet displaced Nvidia at the frontier-training layer, but each one represents a portion of hyperscaler capex that flows to a vendor other than Nvidia. In a world where the largest AI customers each spend tens of billions a year on accelerators, even modest displacement has material revenue consequences.

NVLink Fusion is Nvidia's response. Instead of fighting the custom-silicon wave, Nvidia is positioning to be its connective tissue — the rack fabric, the memory subsystem, the scale-up topology, the CPU companion. The strategic logic is the same one we analyzed in Deep Dive — Nvidia's quiet takeover of the AI stack: if Nvidia cannot be every chip, it should at least be every rack's nervous system. By bringing MediaTek in as the neutral design house, Nvidia adds a third option that hyperscalers have not had until now: "build a custom XPU that lives inside Nvidia's fabric, designed and packaged by MediaTek, certified for NVLink scale-up, and sold through a partner rather than directly competing with Nvidia on the open market."

What the skeptics will say

The skeptical case has two parts. The first is structural. Custom XPUs have always failed for one of two reasons: the volume is too small to amortize the design cost, or the software ecosystem is too thin to make the chip useful. Nvidia's pitch is that NVLink Fusion solves the first problem by providing the surrounding scaffolding as a prevalidated building block. It does not solve the second. A MediaTek-built XPU still needs a CUDA-equivalent software stack, and Nvidia has shown no sign of opening CUDA to third-party XPUs at the level required for serious inference workloads. The chip might ship; the workloads might not follow.

The second skepticism is competitive. Google's TPUs succeeded because Google controlled both the silicon and the workload (search, ads, YouTube, Gemini). Amazon's Trainium succeeded because Amazon could mandate its own internal customers use it. Meta's MTIA is succeeding because Meta's recommendation systems are unique enough to justify custom silicon. None of those companies needed Nvidia's fabric to make their chips work. By inviting Nvidia into the rack, a hyperscaler is implicitly accepting that Nvidia will continue to set the terms on how accelerators talk to each other — and that the hyperscaler's negotiating leverage on future GPU pricing has just been reduced. That is a real trade-off, and it is not obvious that every hyperscaler will accept it.

The third, more cynical read is that Nvidia is buying optionality against a future where the largest customers defect en masse. As we covered when Nvidia backs $500B data center deal with GPU value guarantee, Nvidia's free cash flow has grown roughly 18x in three years, reaching $48.5 billion in its most recent reported quarter. The company is using that cash to buy options on every plausible future of the AI compute stack — power plants, model labs, networking startups, custom-silicon design houses. The MediaTek deal is one more option in a portfolio that already includes Groq, Enfabrica, Poolside, and a near-30-billion-dollar stake in Intel. The bull case is that at least one of these bets produces outsized returns. The bear case is that Nvidia is overpaying for optionality because its core GPU market is closer to peaking than the headlines suggest.

What this changes for everyone else

Three things to watch. The first is whether MediaTek announces a specific customer for its NVLink Fusion-based custom XPU program in the next six months. The partnership is plausible in the abstract, but it only becomes a real product when a hyperscaler, sovereign AI program, or large model lab commits to buy a MediaTek-designed accelerator. Until then, $3.5 billion is a credible statement of intent, not a revenue stream.

The second is whether NVLink Fusion becomes a real standard or stays a Nvidia-controlled ecosystem. The difference matters. If hyperscalers can mix and match XPUs from different vendors inside the same NVLink fabric without renegotiating with Nvidia every time, the platform becomes a genuine alternative to proprietary accelerator programs. If Nvidia retains veto power over who gets to plug in, the platform is just a softer version of the closed stack it already sells.

The third is what this means for the broader custom-silicon competitive set. Marvell, Broadcom, and Alchip already build custom AI silicon for hyperscalers; Astera Labs and others build the connectivity glue. NVLink Fusion raises the bar — designing an XPU that plugs into Nvidia's fabric requires capabilities that the existing custom-silicon shops do not all have, particularly in advanced packaging and high-bandwidth memory integration. MediaTek brings all of those. Expect the next twelve months to look like Nvidia picking favorites among the custom-silicon designers, with MediaTek as the preferred partner and everyone else competing for the residual demand.

The takeaway is bigger than the press release. Nvidia has spent three years selling the world on the idea that its rack-scale systems are the only sensible answer to large-scale AI compute. The MediaTek deal concedes, quietly, that custom silicon is going to exist no matter what Nvidia sells, and that the better fight is over the fabric those custom chips plug into. For a company that built its reputation on controlling the entire stack, that is a significant concession — and a $3.5 billion one.

What do you think — does Nvidia's fabric strategy make custom silicon safer, or does it just make Nvidia harder to escape? Tell us in the comments.

Sources: Nvidia Newsroom · SiliconANGLE · Nvidia Blog (NVLink Fusion) · CNBC on Nvidia's balance sheet · Techmeme on Nvidia/Intel stake