How the Nvidia MediaTek investment turns NVLink Fusion into a custom-chip tollbooth

Nvidia has placed a $3.5 billion bet on MediaTek that is really a bet on its own interconnect standard. The Nvidia MediaTek investment, announced last week, takes the form of convertible bonds that MediaTek will use to expand a custom AI chip business serving the very hyperscalers trying to reduce their dependence on Nvidia GPUs. MediaTek shares rose about 10 percent on the news and remain up close to 200 percent for the year.

The commercial terms are only part of the story. Under the expanded partnership, MediaTek formally adopts NVLink Fusion, Nvidia's rack-scale interconnect platform, as the foundation for the custom AI accelerators it designs for outside customers. The collaboration now spans data center AI, RTX Spark and DGX Spark processors aimed at PCs and workstations, and Dimensity Auto platforms for vehicles.

NVLink Fusion is the element that makes the deal larger than a typical financing round. The platform combines NVLink-C2C die-to-die links with electrical and photonic connectivity and carries support for NVHBM, Nvidia's high-bandwidth memory architecture. NVHBM relocates the memory controller into the HBM stack itself, an arrangement Nvidia says frees about 25 percent of the die for compute. For a chip designer, the practical outcome is that a custom accelerator built on the platform connects directly into Nvidia's rack-scale data center fabric and communicates with Nvidia GPUs and CPUs without extra bridging hardware.

MediaTek is not new to Nvidia's silicon plans. The two companies have co-developed products for years, with MediaTek contributing packaging and integration work around Nvidia's Grace CPU line. What changes now is the business model. Instead of Nvidia pitching its technology to chip designers one at a time, MediaTek becomes the integration and manufacturing partner that offers NVLink Fusion to hyperscalers, cloud providers and frontier model developers as a deployment path already validated against Nvidia's rack-scale systems.

What the Nvidia MediaTek investment buys

MediaTek's custom AI unit is projected to generate about $2 billion in revenue this year, and the company estimates the wider custom AI chip market it serves could reach $80 billion by 2027. Those figures are the commercial rationale for the bond issue and the reason Nvidia's notes convert rather than simply accrue interest.

Structured as convertible debt, the Nvidia MediaTek investment gives Nvidia a share of that growth if the platform succeeds. Rising adoption of NVLink Fusion would push the bonds into equity that captures the upside. The setup has drawn criticism as a form of circular financing, since part of MediaTek's data center growth case now depends on a standard Nvidia owns and controls.

The deal fits a broader pattern in Nvidia's recent financing. Nvidia has repeatedly invested across the AI supply chain, putting capital into partners whose products flow back into its own ecosystem, and it is doing so at a moment when hyperscalers are spending heavily on bespoke silicon that is meant to take workloads off Nvidia GPUs. What is unusual this time is the target: an ASIC design house that serves those same hyperscalers. Investing in that designer buys a seat at the table where non-Nvidia AI chips are designed.

Why NVLink Fusion is the real prize

The stated goal of the two companies is to keep custom AI chips compatible with Nvidia's networking, CPUs and software stack. The strategy behind that goal is defensive, aimed at the bespoke silicon boom rather than against it. Nvidia is buying a position inside the supply chain that designs those chips, so the finished silicon still plugs into Nvidia infrastructure.

The deeper motive is control of the control plane. If custom XPUs had to be managed as foreign devices inside Nvidia-built data centers, Nvidia's software would gradually lose its role as the layer that orchestrates heterogeneous clusters. By folding bespoke silicon into the NVLink Fusion ecosystem, Nvidia keeps its software, networking and CPU line at the center of AI factories that will contain fewer Nvidia GPUs over time.

NVHBM shows how far the integration reaches. With the memory controller living inside the HBM stack, a custom accelerator's memory pathway itself runs on Nvidia IP, and the chip has to follow Nvidia's memory architecture to reach the fabric. The toll extends beyond the interconnect into how the XPU is designed, because NVLink Fusion and NVHBM are the route any bespoke accelerator must take to sit inside an Nvidia-based rack.

A hyperscaler or AI lab can commission MediaTek to build an XPU tuned to a specific model architecture or inference workload, then deploy it in an Nvidia-connected facility where it talks to Nvidia GPUs and CPUs over NVLink Fusion. Every such accelerator pays an Nvidia toll at the fabric, the memory and the software layers, even when its purpose is to displace Nvidia GPUs in part of the rack. That is the difference between this move and a conventional portfolio investment: the return rides on the spread of the standard as much as on MediaTek's own execution.

The partnership also extends the standard's reach outside the data center. RTX Spark and DGX Spark bring NVLink-based computing to PCs and workstations, and the Dimensity Auto line carries the collaboration into vehicles, so the NVLink Fusion footprint is meant to stretch from cloud to edge devices. Each market adds another pool of hardware designed around Nvidia's interconnect.

The trade-offs for hyperscalers and AI labs

Custom silicon buyers now have three routes. They can keep buying Nvidia GPUs outright. They can commission an ASIC house such as Broadcom or Marvell, the incumbents MediaTek is now chasing in the data center. Or they can design through the NVLink Fusion channel MediaTek will operate, an option that did not exist in this form before the deal.

The third route buys compatibility at the price of independence. A chip built around NVLink Fusion inherits Nvidia's memory architecture and software stack and connects into Nvidia facilities with its interfaces already validated, but its roadmap follows Nvidia's, and its value depends on terms Nvidia sets for the platform. The independent route preserves strategic freedom but forfeits drop-in compatibility with the installed base of Nvidia infrastructure.

The right choice depends on the buyer. A frontier model developer without its own data center fabric gains enormous leverage from plugging a tailored XPU into existing Nvidia systems, and MediaTek's entry gives it a second design partner alongside Broadcom and Marvell. A hyperscaler operating its own fabric weighs the same compatibility against the cost of building around a rival's proprietary standard.

The arrival of a platform-backed rival changes the design market itself. MediaTek can now offer customers drop-in Nvidia compatibility at the moment the custom AI chip market is heading toward the $80 billion scale it projects, which pressures Broadcom and Marvell from two directions at once. For buyers, more competition among design houses usually means better commercial terms and shorter design cycles.

For Nvidia, the $3.5 billion buys optionality in both directions. If custom XPUs take share from GPUs, Nvidia still earns on the fabric, the CPUs, the networking and the software around them, converting some lost GPU revenue into platform revenue; if GPUs keep winning, the bond is a modest cost of defending the franchise. For MediaTek, the capital and the platform access turn a commercial relationship into a strategic alliance, and the market reaction, a 10 percent jump on top of a near-200 percent run for the year, shows investors assigning real value to ambitions beyond the phone chip business.

Why this matters

The Nvidia MediaTek investment is Nvidia's answer to hyperscaler vertical integration, and it moves the custom chip contest onto the fabric and memory standards between accelerators. Data center buyers should now plan any bespoke XPU program around an interconnect component controlled by the vendor the program was meant to displace. Whether the tollbooth holds depends on how much custom silicon volume materializes by the $80 billion market the partners project, and on whether Broadcom and Marvell build independent ecosystems around their own design wins.

Related Articles

• MediaTek AI Data-Center Chip Expansion: Inside the $5B Bet and the 20% Phone Slump
• Google's $150B Anthropic Chip Financing Machine Beats Nvidia on Borrowing Costs
• Marvell and Alphabet Co-Develop Custom AI Inference Silicon

✔Human Verified

Researched and cross-referenced against primary sources by the Bytevyte editorial team. This article was generated with the assistance of artificial intelligence and reviewed by the Bytevyte editorial team.