On August 31, 2026, Nvidia announced an investment in MediaTek's $3.5 billion convertible bond issuance. A key aspect of the expanded cooperation between the two companies is to enable MediaTek's customers to develop custom AI processors that can be connected to Nvidia's rack-level systems via NVLink Fusion. The investment vehicle is a convertible bond, which should not be directly interpreted as Nvidia obtaining a corresponding proportion of common stock, let alone acquiring MediaTek.
The notable aspect of this deal is that the partnership is not entirely complementary. The custom chips developed by MediaTek with the help of its clients may replace GPUs in certain workloads. If judged solely by the sales opportunity of individual chips, it seems that Nvidia is funding a potential competitor. However, considering the CPU, memory, interconnect, and rack system as a whole, the question becomes: after clients choose to develop their own computing chips, will they still adopt Nvidia's other technologies?
By examining its investment record over the past few years, it can be inferred that Nvidia does not require every partner to exit the competition for computing chips, but rather strives to develop more different computing architectures together with its interconnects and system platforms. Capital investment is a means to facilitate such long-term cooperation.
Mediatek Not New to Partnership, But Depth of Collaboration Has Changed
This is not MediaTek's first involvement with NVLink Fusion. When Nvidia released the platform in May 2025, the initial partners already included MediaTek, Marvell, Unisoc, Astera Labs, Synopsys, and Cadence. MediaTek subsequently introduced a custom ASIC collaboration model for cloud-based AI. Therefore, describing this deal as "MediaTek's first entry into Nvidia's AI ecosystem" would overlook the cooperation foundation established over the past year.
This announcement further outlined a set of technologies surrounding the customized XPU: the NVLink Fusion chiplet connecting the expansion network, NVLink-C2C connecting the processor, and the customized memory technology NVHBM. The two parties will also collaborate on packaging, manufacturing, and rack-level integration, and continue their cooperation on PC and automotive platforms. These are the directions of cooperation disclosed by the company, which are not equivalent to actual orders and revenue that have been achieved.
For customized chip clients, designing the computing unit is only part of the project. The chip also needs to exchange data with memory and other processors, undergo packaging and system verification, and ultimately enter a cluster that can run continuously. The appeal of the cooperative model lies in the fact that clients can reserve more engineering resources for the parts they truly want to differentiate, while reusing technology in other stages. Whether this approach really shortens the development cycle and reduces total costs remains to be proven by the project results.
What MediaTek gained is not just funding. A reusable system integration path also gives it an opportunity to win over customers that would have otherwise required it to undertake more infrastructure development work on its own. However, the more customers rely on external platforms for system deployment, the more they need to clarify what design assets they possess and how much room for maneuver they have for next-generation products. This is the actual trade-off between convenience and autonomy.
Why is capital needed, rather than just signing a technology cooperation agreement? A reasonable explanation is that multi-generation product cooperation requires partners to invest engineering resources in advance and bear market risks that have not yet materialized. Capital investment can support such investment and increase long-term financial interests between Nvidia and its partners. It helps to improve the sustainability of cooperation, but the public announcement did not specify how the $5 billion is linked to specific R&D expenditures or procurement metrics, and it cannot be concluded that Nvidia has bought exclusive cooperation or fixed orders.
Nvidia's investment record shows it hasn't put all its eggs in one computational basket
The following are representative historical transactions verified as of September 2, 2026, covering chip design, processor IP, and optical interconnect companies, and are not an exhaustive portfolio. Intel also has a manufacturing business, while SiFive primarily provides processor IP, and neither can be simply classified as a pure chip design company.
The total amount of financing rounds listed cannot be taken as NVIDIA's investment amount. The three disclosed investments by NVIDIA in Intel, Marvell, and MediaTek total $10.5 billion, but they were made through common stock, convertible preferred stock, and convertible bonds, which cannot be further interpreted as the same level of company control.
These transactions are not all investments in "competitors." Processors, networks, and optical interconnects perform different tasks in a system; some companies have alternative relationships with Nvidia at the product level, while others primarily complement foundational technologies. To understand this group of investments, it's necessary to look at how specific products are combined, rather than labeling each company as a friend or enemy from the outset.
Marvell and MediaTek: Competition in Computing Chips, Cooperation in Systems Still Possible
The Marvell partnership best illustrates this layered relationship. In March 2026, the two companies explicitly disclosed their respective roles: Marvell supplies custom XPUs and expanded networking compatible with NVLink Fusion, while Nvidia provides supporting technologies including the Vera CPU, ConnectX network adapters, BlueField DPUs, NVLink, and Spectrum-X switches, in addition to collaborating on silicon photonics.
From this, one can infer Nvidia's commercial calculus: even if customers offload some computing tasks to custom XPUs, they may still purchase Nvidia's processors, networking, and other system components. Conversely, if customers shift to an alternative solution encompassing interconnect and system architecture, Nvidia's potential loss could extend beyond GPUs. Therefore, supporting custom chips within its own ecosystem allows Nvidia to retain a portion of business that might otherwise migrate to competing platforms.
This does not mean that the loss from selling fewer GPUs will necessarily be offset by network products. Different components have different revenue and profit levels, and partners may also gain stronger bargaining power. A more accurate statement would be that Nvidia has increased its participation in customer customization projects, rather than obtaining a guarantee of "making money regardless of who wins".
Another notable aspect is that Marvell, like MediaTek, offers customized chip capabilities, and the two may compete for client projects. Nvidia has established capital ties with both companies, allowing clients to choose between different design partners without having to switch entire technology platforms. From Nvidia's perspective, this may also reduce dependence on a particular design partner. This is an inference based on the cooperative structure, rather than an exclusive arrangement disclosed by the company.
For Marvell and MediaTek, participating in the same ecosystem does not eliminate competition. Whoever can more accurately understand customer workloads, execute complex integrations, and deliver on schedule will still determine where design wins land. Common interfaces may expand the addressable market, but they could also sharpen differentiation among design houses around architecture, engineering execution, and customer relationships.
Intel and SiFive: No need to win the CPU battle to expand accelerator computing
Intel's collaboration is unfolding along a different path. The arrangement announced by both parties in September 2025 includes: Intel developing customized x86 CPUs for Nvidia, which Nvidia will integrate into its AI infrastructure; in the PC domain, Intel developing x86 SoCs that integrate Nvidia's RTX GPU cores. The announcement did not confirm that Nvidia's GPUs would be transferred to Intel's fabs for manufacturing, and the processor collaboration should not be misconstrued as a foundry order.
SiFive announced on January 15, 2026, that it would integrate NVLink Fusion into its high-performance RISC-V solutions for data centers, followed by a new round of funding in April that included NVIDIA. The technical cooperation preceded this round of funding, once again demonstrating that capital investment can strengthen existing partnerships rather than being the starting point for cooperation.
Nvidia already has its own CPU products, yet it continues to support collaboration paths with x86 and RISC-V. This choice is not hard to understand: which CPU a customer uses is a different question from whether that customer adopts Nvidia's accelerators. Supporting a wider range of CPU options could lower the barrier to adopting its accelerated computing platform, even if that means some competition with Nvidia's own CPUs.
The PC market follows a similar logic. Intel and MediaTek each possess distinct SoC capabilities and customer bases, giving Nvidia two partnership paths to broaden the reach of its graphics and AI technologies. These relationships involve both competition among partners and cooperation in pursuing new markets together—so reducing all investment motives to "defending against ASIC" would be an oversimplification.
Optical Interconnects and EDA: Enabling More Chips to Enter Complex Systems
Nvidia's emphasis on networking did not start with NVLink Fusion. In April 2020, it completed the acquisition of Mellanox, combining computing with high-performance networking capabilities. Recent investments in interconnect companies represent another way to participate in the development of external technologies, and the roles of acquisitions and minority equity investments should not be conflated.
Enfabrica and Ayar Labs offer insights into how to improve data exchange efficiency between computing chips. Notably, Ayar Labs' 2024 funding round included participants such as Nvidia, AMD Ventures, and Intel Capital, while its 2026 funding round saw the addition of MediaTek and SMIC. This cross-investment at least suggests that a single underlying technology supplier can receive support from competing companies without being inherently tied to a specific closed camp.
Extending the observation scope to design tools, Nvidia also announced on December 1, 2025 that it had purchased $2 billion of Synopsys common stock. Synopsys is a provider of EDA, IP, and engineering software, and cannot be directly classified as another GPU or ASIC vendor. The two companies' cooperation focuses on accelerating design and simulation using technologies such as CUDA-X, with the explicit clarification that the partnership is non-exclusive.
This investment expands the use of engineering software for accelerated computing. One possible outcome is that companies developing competing chips will also use tools running on Nvidia hardware during the design and verification process. However, this does not mean Nvidia controls the customer's design, nor does it have access to other companies' confidential design data. EDA collaboration, computing resource procurement, and design intellectual property are relationships on different levels.
Wider Cooperation Means Openness Doesn't Signal End of Competition
Nvidia's system-level competition now has a concrete reference point. On October 13, 2025, OpenAI and Broadcom announced a collaboration plan for 10-gigawatt custom AI accelerators, with rack solutions that also incorporate Broadcom networking technology. This is a published deployment plan, not completed installed capacity, but it demonstrates that customer-designed chips can be built together with systems and networking—not merely as a replacement for a single compute chip.
There are also other technical routes for interconnection. The UALink alliance released a new specification on April 7, 2026, which includes in-network computing, chiplets, and management functions. Notably, Synopsys, which has received investment from NVIDIA, still openly supports UALink. Capital relationships do not automatically preclude a company's choice to participate in other technical ecosystems.
NVLink Fusion's expansion of its partner ecosystem and adoption of an open standard with multiple stakeholders are two separate things. Similarly, interconnect compatibility does not directly imply software compatibility: just because an XPU can access a system does not mean it automatically has the ability to run CUDA programs. Compiler, operator, and application migration work still needs to be done separately.
For clients, it's necessary to compare the overall cost of the solution, delivery risk, actual performance, and subsequent options. Pre-integrated solutions have practical appeal if they can reduce engineering work, but they also raise long-term concerns if interfaces, component supply, and product iteration are overly dependent on a single platform. The release of standards does not mean that alternative solutions are already mature, and an increase in partners does not mean that Nvidia has won all projects.
For MediaTek, the most noteworthy follow-up to this $3.5 billion deal is not about adding more partners to the list, but whether verifiable customer mass production projects can emerge: what benefits do customers' customized computing bring, what functions do Nvidia's supporting technologies undertake, how do both parties allocate engineering responsibilities, and whether next-generation products will continue to use this combination.
Nvidia is attempting to diversify its computing chips while expanding its own platform, a layout that can accommodate competition but not eliminate it. Whether this will ultimately succeed depends on whether partners like MediaTek and Marvell can deliver systems that customers are willing to continue buying, and whether the value Nvidia gains from this can offset the cost of ceding some computing opportunities.