On August 31, 2026, Nvidia announced a $3.5 billion investment in convertible bonds issued by MediaTek. A key part of the expanded cooperation is enabling custom AI processors developed by MediaTek's clients to access Nvidia's rack-scale systems through NVLink Fusion. The investment vehicle is convertible bonds, so it should not be described as Nvidia having already acquired a corresponding stake in common shares, nor as an acquisition of MediaTek.
What makes this deal worth discussing is that the two partners are not complementary in every respect. The custom chips MediaTek helps clients develop could replace GPUs in certain workloads. Judging only by single-chip sales opportunities, Nvidia appears to be funding a potential competitor. But when CPUs, memory, interconnects and rack systems are considered together, the question becomes: after clients choose their own compute chips, will they still adopt Nvidia's other technologies?
A review of Nvidia's investment record in recent years supports a more evidence-based judgment: Nvidia is not demanding that every partner exit the compute-chip competition, but rather is seeking to have more different computing architectures develop alongside its interconnect and system platforms. Capital investment is one means of fostering this long-term cooperation.
MediaTek's Deeper Ties, Not a New Entry
This is not MediaTek's first involvement with NVLink Fusion. When Nvidia launched the platform in May 2025, the initial partners already included MediaTek, Marvell, Alchip, Astera Labs, Synopsys and Cadence. MediaTek subsequently introduced a custom ASIC cooperation model for cloud AI. Therefore, describing this transaction as 'MediaTek's first entry into Nvidia's AI ecosystem' would overlook more than a year of prior cooperation.
The announcement also lists a set of technologies surrounding custom XPUs: NVLink Fusion chiplets for connecting to expansion networks, NVLink-C2C for connecting processors, and custom memory technology NVHBM. The two companies will also collaborate on packaging, manufacturing and rack-scale integration, and continue cooperation on PC and automotive platforms. These are the cooperation directions disclosed by the companies, and cannot yet be equated with realized orders and revenue.
For custom-chip clients, designing the compute unit is only part of the project. The chip must also exchange data with memory and other processors, undergo packaging and system validation, and finally enter clusters that can run continuously. The appeal of the cooperation model is that clients can devote more engineering effort to the parts they truly want to differentiate, while reusing technology from other segments. Whether it actually shortens development cycles and lowers total costs still needs to be proven by project results.
MediaTek gains more than just capital. A reusable system-integration path could help it win clients that would otherwise need to take on more infrastructure development themselves. However, the more clients rely on external platforms for system deployment, the more they need to clarify which design assets they control and how much choice they have for next-generation products. That is the real trade-off between convenience and autonomy.
Why is capital needed rather than just a technology cooperation agreement? One reasonable explanation is that multi-generation product cooperation requires partners to commit engineering resources in advance and bear market risks that have not yet materialized. Capital investment can support that commitment and also strengthen the long-term financial interest between Nvidia and its partners. It helps make cooperation more durable, but the public announcement does not specify how the $3.5 billion is tied to specific R&D spending or procurement targets, so it cannot be concluded that Nvidia has bought exclusive cooperation or guaranteed orders.
Nvidia's Investment Record Shows It Didn't Bet on Just One Computing Path
The following representative historical transactions, verified as of Sept. 2, 2026, cover chip-design, processor-IP and optical-interconnect companies and are not a complete portfolio. Intel also has manufacturing operations, while SiFive mainly provides processor IP, so neither can be simply categorized as a pure chip-design company.
The total of the financing rounds in the table should not be treated as Nvidia's investment amount. The disclosed Nvidia investments in Intel, Marvell and MediaTek alone total $10.5 billion, but they were made through common stock, convertible preferred stock and convertible notes respectively, and cannot be further interpreted as the same degree of corporate control.
These transactions are not all investments in competitors. Processors, networking and optical interconnects perform different functions in a system; some companies are substitutes for Nvidia at the product level, while others mainly supplement its foundational technology. Understanding this group of investments requires looking at how specific products combine, rather than labeling each company as friend or enemy.
Marvell and MediaTek: Competing in Chips, Cooperating in Systems
The Marvell partnership best illustrates this layered relationship. In March 2026, the two companies disclosed that Marvell would provide custom XPUs and scale-out networking compatible with NVLink Fusion, while Nvidia would supply supporting technology including Vera CPUs, ConnectX network adapters, BlueField DPUs, NVLink and Spectrum-X switches, and they would collaborate on silicon photonics.
This allows an inference of Nvidia's commercial rationale: even if customers assign some computing tasks to custom XPUs, they may still buy Nvidia's processors, networking and other system components. Conversely, if customers move their interconnect and system architecture to another solution, Nvidia's lost opportunity could extend beyond GPUs. Therefore, supporting custom chips within its own system can retain some business that might otherwise go to other platforms.
This does not mean the loss from fewer GPU sales will necessarily be recovered by networking products. Revenue and profit margins vary by component, and partners may also gain stronger bargaining power. A more accurate description is that Nvidia has increased its ways to participate in customers' custom projects, rather than obtaining a guarantee that it will profit no matter who wins.
Another point worth noting: Marvell and MediaTek both offer custom-chip capabilities, so they may also compete for customer projects. Nvidia's capital ties to both companies help customers choose among different design partners without having to switch the entire technology platform. From Nvidia's perspective, this may also reduce its reliance on any one design partner. This is an inference drawn from the partnership structure, not an exclusivity arrangement disclosed by the companies.
For Marvell and MediaTek, being in the same ecosystem does not eliminate competition. Whoever can more accurately understand customer workloads, complete complex integration and deliver on time will still determine which company wins design orders. A common interface may expand the addressable market, but it may also concentrate design companies' differentiation on architecture, engineering execution and customer relationships.
Intel and SiFive: Winning the CPU race isn't a prerequisite for expanding accelerated computing.
Intel's collaboration follows a different path. The arrangement announced in September 2025 includes Intel developing custom x86 CPUs for Nvidia to integrate into AI infrastructure; in the PC segment, Intel is developing x86 SoCs that integrate Nvidia RTX GPU chiplets. The announcement did not confirm that Nvidia GPUs would be manufactured at Intel's fabs, so the processor partnership should not be interpreted as a foundry order.
SiFive announced on Jan. 15, 2026 that it would integrate NVLink Fusion into its high-performance RISC-V offerings for data centers, followed by a new funding round in April that included Nvidia. The technology collaboration preceded the funding, underscoring that capital investment can strengthen an existing partnership rather than necessarily mark its starting point.
Nvidia already has its own CPU products, yet it continues to support x86 and RISC-V collaboration paths. The rationale is straightforward: which CPU a customer uses is a different question from whether they adopt Nvidia accelerators. Supporting more CPU options can lower the barrier to adopting its accelerated computing platform, even if that creates some competition with Nvidia's own CPUs.
The PC market follows similar logic. Intel and MediaTek each have distinct SoC capabilities and customer bases, and Nvidia can use both collaboration paths to broaden the reach of its graphics and AI technologies. The dynamic involves competition among partners as well as cooperation to win new markets, so "defense against ASICs" alone cannot explain all investment motives.
Optical Interconnect and EDA: Opening Complex Systems to More Chips
Nvidia's focus on networking did not begin with NVLink Fusion. In April 2020, it completed its acquisition of Mellanox, combining computing with high-performance networking capabilities. Its recent investments in interconnect companies represent another way to participate in external technology development; the roles of M&A and minority equity investments should not be conflated.
Enfabrica and Ayar Labs offer a lens into improving data-exchange efficiency among computing chips. Ayar Labs in particular drew Nvidia, AMD Ventures and Intel Capital as participants in its 2024 funding round; its 2026 round added MediaTek and Alchip. Such cross-investment shows that a single underlying technology supplier can attract support from competing companies and does not inherently belong to one closed camp.
Extending the lens to design tools, Nvidia announced on Dec. 1, 2025 that it had bought $2 billion of Synopsys common stock. Synopsys is an EDA, IP and engineering-software supplier, not directly another GPU or ASIC maker. The collaboration focuses on using technologies such as CUDA-X to accelerate design and simulation, and the companies explicitly said the partnership is non-exclusive.
This investment expands the use of accelerated computing in engineering software. One likely outcome is that companies developing rival chips will use tools running on Nvidia hardware in their own design and verification processes. That doesn't mean Nvidia controls customer designs, or that it gains access to proprietary design data. EDA partnerships, compute purchases, and design IP are separate layers of relationship.
Broader Cooperation, But Openness Isn't the End of Competition
Nvidia's system-level competition has a concrete reference point. On Oct. 13, 2025, OpenAI and Broadcom unveiled plans for custom AI accelerators targeting 10 gigawatts, with rack designs that include Broadcom networking. The announcement is a deployment plan, not completed installations, but it illustrates that customers can design their own silicon alongside systems and networking, rather than merely swapping out one compute chip.
Interconnect has other technical routes. On April 7, 2026, the UALink consortium issued a new spec covering in-network computing, chiplets, and management functions. Notably, Synopsys — an Nvidia-backed company — continues to publicly back UALink. Equity ties don't automatically foreclose participation in other technology ecosystems.
Expanding the partner ecosystem through NVLink Fusion isn't the same as embracing an open standard with multi-stakeholder governance. Similarly, interconnect compatibility doesn't translate into software compatibility: an XPU's ability to connect to a system doesn't mean it can execute CUDA programs. Compilers, operators, and application migration remain separate workstreams.
Customers must weigh total cost, delivery risk, real-world performance, and future choice. Pre-integrated packages hold appeal if they cut engineering work; excessive dependence on one platform for interfaces, components, and upgrades creates longer-term worries. Publishing a spec isn't proof that alternatives are ready, and a growing partner list doesn't mean Nvidia has won all deals.
For MediaTek, the key question after the $3.5 billion isn't the size of the partner roster. It's whether verifiable customer production programs materialize: what value custom compute brings to clients, which functions Nvidia's accompanying technology handles, how engineering duties are divided, and whether future generations stick with the same combination.
Nvidia is attempting to make chip diversification and platform expansion happen simultaneously. That strategy can accommodate competition, but not eliminate it. Its success ultimately hinges on whether partners such as MediaTek and Marvell deliver systems customers will keep purchasing — and whether Nvidia's share of value offsets the cost of giving up some compute opportunities.