Nvidia bets $3.5 billion on MediaTek to lock in AI

An investment that goes far beyond financial logic

Nvidia is investing $3.5 billion in MediaTek, the Taiwanese group known for its processors intended in particular for smartphones, connected devices and automotive applications. Reported by TechCrunch in an article titled “Nvidia’s $3.5B MediaTek bet reveals its plan for tackling Big Tech’s AI chip buildout”, the deal should be viewed as an industrial and strategic move as much as a financial one.

The amount reflects the scale of the stakes. Nvidia is not merely seeking to strengthen a relationship with a component supplier or diversify its investments. The group led by Jensen Huang is moving closer to a chip designer with useful expertise in a segment that datacenter GPUs do not cover, or do not cover on their own: the design of integrated systems tailored to specific devices, particular energy constraints and very high production volumes.

This dimension is essential in the current artificial intelligence landscape. Nvidia has become the most visible player in AI infrastructure thanks to its graphics processors, computing systems and CUDA software platform. But the market is no longer limited to purchases of standardized GPUs by labs, cloud providers and large companies. Nvidia's main customers are now seeking to bring part of their hardware roadmaps in-house. Google, Amazon Web Services, Microsoft and Meta are developing or supporting their own accelerators for certain uses, notably to better control their costs, performance and supply.

In this landscape, specialized chips are not necessarily a complete alternative to Nvidia GPUs. They can target inference, meaning the execution of already-trained models, clearly defined internal tasks or cloud infrastructures designed around each group's own software. They can also be integrated into servers, consumer devices, vehicles or industrial equipment. This is precisely where the alliance with MediaTek takes on particular importance.

MediaTek has extensive experience in designing systems on a chip, or SoCs. These components typically bring together several functions within a single chip: processors, graphics elements, connectivity, power management or capabilities dedicated to data processing. This approach is very different from simply supplying a computing accelerator intended for a server rack. It requires mastering trade-offs between power, consumption, cost, component size and integration with other hardware elements.

For Nvidia, the challenge is therefore to better position itself in a phase where AI is gradually moving beyond datacenters alone to spread across all digital equipment. Smartphones, PCs, vehicles, robots, connected objects, enterprise systems and industrial devices have constraints that differ greatly from those of a training cluster. In all these cases, local data processing, low energy consumption, connectivity and the ability to design a suitable chip can matter as much as raw performance.

The wording of TechCrunch's headline is revealing: the bet on MediaTek lays out how Nvidia intends to respond to Big Tech's AI chip buildout. The group does not appear content with defending its positions by selling ever more powerful generations of GPUs. It is moving closer to a partner that could help it take part in the rise of customized chips, including when customers seek to reduce their dependence on its best-known products.

This strategy is defensive because major cloud operators account for a crucial share of global demand for AI computing. But it is also offensive. If Nvidia can provide, directly or with MediaTek, technology building blocks, architectures, software or design capabilities to these same customers, it can remain present in their infrastructures even when they adopt specific components.

The tie-up also comes at a time when the semiconductor value chain is becoming more fragmented. Companies no longer simply buy a finished chip: they assemble complete systems, selecting memory, interconnects, processors, software and cooling solutions. Value is shifting toward integration. For a player such as Nvidia, mastering the accelerator alone may no longer be enough to retain a central position across all AI markets.

MediaTek, a Taiwanese partner at the heart of embedded systems

MediaTek does not occupy the same place as Nvidia in the imagination surrounding generative AI, but its know-how is particularly relevant to the market's next phase. The Taiwanese group is widely identified as a chip designer for mobile and connected devices. Its experience concerns platforms where success depends less on isolated maximum performance than on the balance between hardware integration, battery life, thermal considerations, connectivity and final price.

This expertise is difficult to reproduce quickly. Building an SoC intended for a connected device requires coordinating numerous subsystems. Compatibility with communication modules must be ensured, power resources must be managed, different computing units must be integrated, and the whole must be capable of being produced at scale. Shorter product cycles and certification constraints, which do not always exist in the same way in datacenters, must also be taken into account.

In the AI world, this expertise can translate into processors capable of running certain models locally, processing sensor data or offering intelligent functions without constantly relying on a remote service. This embedded AI, or edge AI, logic is becoming more important as models become more efficient and manufacturers seek to limit cloud inference costs.

Nvidia's investment, as reported by TechCrunch, therefore strengthens its access to markets where custom architectures and integrated devices play a decisive role. This includes connected endpoints, but also specialized infrastructure. Between the smartphone and the very large datacenter lies a vast range of products: servers designed for a specific task, telecommunications equipment, automotive systems, point-of-sale terminals, medical devices, industrial machines or robotics tools. Not all of them will need a very high-capacity GPU. Many will instead require an integrated chip suited to a limited power envelope.

The complementarity with Nvidia lies in this division of strengths. Nvidia brings its standing as a benchmark in accelerated computing, its software environment and its experience with large-scale AI systems. MediaTek brings in-depth knowledge of compact platform design and mass-market electronics markets. Such a combination can more closely connect cloud computing and local computing, without assuming that a single type of processor will rule over every use case.

The American group is no stranger to the issue of complete systems. Its activities already cover GPUs, computing architectures, networking, software platforms and several specialized markets, including automotive. But the rise of AI in devices requires partners with industrial relationships, design expertise and a product culture distinct from those of the datacenter. The MediaTek deal is a way to consolidate this capability.

It also serves as a reminder that the chip sector rests on global interdependencies. Nvidia is an American company, MediaTek is Taiwanese, and design, manufacturing, assembly, memory and software are distributed across many territories. In this context, a multi-billion-dollar investment is not simply a rapprochement between two brands. It reflects Nvidia's need to secure capabilities and industrial options at a time when advanced components are as much a geopolitical issue as an economic one.

The relationship between the two companies is also part of a trajectory in which Nvidia has sought to extend its influence beyond the traditional GPU. The company founded in 1993 built its identity around graphics processors before capitalizing on their ability to perform parallel calculations. CUDA, its programming platform launched in the mid-2000s, helped turn these processors into general-purpose computing tools. The rise of deep learning then made GPUs a cornerstone of training large models.

This history explains Nvidia's current strength, but it also outlines its potential limitation. The GPU has become a de facto standard for a large share of AI workloads, particularly because it comes with a mature software ecosystem. However, use cases are multiplying, and some may be better served by components designed around a specific model, algorithm or environment. MediaTek gives Nvidia more direct access to this logic of customization.

The question is therefore not only whether Nvidia will make more chips for devices. It is whether the group can become an indispensable player when manufacturers and operators want to combine general technologies with components tailored to their own needs. From this perspective, MediaTek's expertise is not merely a peripheral complement: it can become a central instrument for retaining a place in markets moving away from the model of the GPU sold as a standard product.

The response to Google, AWS, Microsoft and Meta

The main strategic signal from this deal concerns the major technology groups developing their own accelerators. Google has its TPUs, designed for its machine-learning computing needs. Amazon Web Services develops Trainium and Inferentia chips. Microsoft has introduced Maia, while Meta is working on its own accelerators through internal programs. These initiatives do not mean that these groups are abandoning Nvidia. They do show, however, that they do not want to depend on a single supplier for all their AI workloads.

This diversification responds to several imperatives. Cloud giants operate infrastructure on a considerable scale and seek to optimize every layer of their technology stack. An in-house chip can be calibrated for a type of model, a software library, a data format or a particularly widespread inference task. It can also be more tightly integrated with internal servers and networks. At very large scale, even a targeted improvement in efficiency can have significant consequences for total operating cost.

Using internal accelerators is also a means of negotiating and planning. When a customer has a credible alternative, it has greater room for maneuver in dealing with external suppliers. It can distribute its workloads, limit supply risks and decide which services should run on which hardware. The choices are not solely technical: they concern infrastructure sovereignty, component availability and control of cloud margins.

Nvidia nevertheless remains difficult to replace entirely. Its strength does not stem only from the power of its GPUs. It comes from the combination of hardware, interconnects, systems and the software ecosystem. Many researchers, developers and software providers have built their tools around CUDA and libraries compatible with Nvidia platforms. Changing architecture can require time, skills and sometimes substantial software rewriting.

It is precisely this tension that the investment in MediaTek appears to seek to exploit. If Nvidia were to stick to general-purpose GPUs, the in-house chips of Google, AWS, Microsoft or Meta could gradually gain ground in the most standardized tasks within their infrastructures. By moving closer to an expert in integrated chips, Nvidia can position itself nearer to the process through which these companies design specific solutions. The implicit goal is not to be pushed out of the value chain when customers customize their hardware.

The deal should not, however, be equated with taking control of the custom-chip market. Major groups have their own design teams, priorities and varied industrial partnerships. Development cycles are long. Moreover, an in-house chip is useful only if it comes with software tools, compilers, libraries and robust operational procedures. Hardware design is an essential element, but it is not sufficient on its own.

Nvidia's case nevertheless shows that the dominant supplier has an interest in participating in these trajectories rather than confronting them head-on. A company can retain a role by providing computing, networking or intellectual-property building blocks, software, or co-design expertise. The partnership then becomes a form of response to disintermediation: instead of seeing a customer become fully autonomous, the supplier seeks to remain indispensable within its architecture.

The scope of the tie-up with MediaTek can also be measured at the inference level. Training the largest models requires massive capabilities and remains associated with vast clusters. But inference is set to occur at very high frequency as assistants, software agents and generative functions are deployed in everyday products. Some of this computing can remain in the cloud; another part can move to devices or local infrastructure. These uses create a more diverse market in which integrated chips can play a larger role.

For Big Tech, the ideal strategy could be hybrid: using Nvidia GPUs and systems where they offer the best flexibility or the greatest power, while deploying specific components for repetitive and predictable services. For Nvidia, the danger would be to be absent from this second category. TechCrunch thus presents the bet on MediaTek as part of its strategy in response to the AI chip buildout of major platforms.

Competition is therefore no longer played out solely between Nvidia and another GPU manufacturer. It pits visions of infrastructure against one another. On one side are general-purpose, extremely high-performance platforms used for a variety of models and workloads. On the other are specialized architectures designed to reduce the cost and power consumption of a specific set of tasks. The $3.5 billion investment suggests that Nvidia sees both approaches as compatible, and that it wants influence in each of them.

A major industrial issue for Europe and the French-speaking market

For France and Europe, the announcement does not directly change the available offering overnight. It does, however, provide useful insight into the choices that the continent's companies, public administrations and digital service providers will have to make. European organizations deploying AI will have to choose between centralized cloud computing, local infrastructure and embedded processing. This decision will depend on costs, performance, data confidentiality, energy consumption and sector-specific obligations.

In this context, the spread of more integrated chips can support certain use cases. An industrial company may want to analyze images, sensor data or video streams locally without continuously sending all information to a remote datacenter. A player in healthcare, transport or energy may seek low latency and better data control. Equipment manufacturers may want to integrate AI functions directly into their products. These needs do not necessarily overlap with those of companies training very large language models.

France has active players in software, cloud, research and industrial AI applications. It is also interested in the issue of digital sovereignty and computing capabilities. But the Nvidia-MediaTek announcement is a reminder of a reality: control over the most advanced components and their design remains largely concentrated outside Europe. European companies are often customers of American or Asian platforms for processors, cloud services and development tools.

This observation does not mean that Europe is absent from the semiconductor industry. The continent has recognized expertise in several segments, from production equipment to specialized chips, including automotive and industrial electronics. Nevertheless, competition around large-scale AI accelerators, proprietary architectures and software platforms remains dominated by a limited number of global companies. Nvidia's investment in MediaTek illustrates the intensification of this concentration around cross-border partnerships.

For French-speaking customers, the practical interest lies in watching whether this tie-up speeds up the arrival of systems combining local computing and cloud services. In regulated sectors, embedded processing can address certain confidentiality or connectivity concerns. In consumer uses, it can reduce dependence on a permanent connection. In industry, it can improve equipment responsiveness. But these benefits will depend on the products actually brought to market, their consumption and the integration conditions offered to European manufacturers.

Software will remain decisive. A high-performance chip has practical value only if developers have tools capable of exploiting its capabilities. Nvidia has a historical advantage with CUDA, while companies deploying local solutions often have to contend with a diversity of processors and development kits. If collaboration with MediaTek results in coherent platforms, it could facilitate the adoption of certain architectures. Conversely, if it adds a layer of fragmentation, integrators will have to manage additional environments.

The European market will also be attentive to supply conditions. Global demand for AI hardware has highlighted the importance of component availability, production capacity and delivery times. A closer alliance between a leader in accelerated computing and a Taiwanese designer may strengthen development capabilities, but it also confirms the dependence of many markets on Asian and global supply chains. Technology investment decisions in Europe will therefore have to incorporate supply resilience, not just announced performance.

The energy question cannot be separated from this development. Large AI datacenters require considerable electrical and cooling infrastructure. Embedded computing does not eliminate these needs, because training and part of inference remain centralized, but it can distribute the load differently. For French and European players subject to energy-efficiency targets or network-capacity constraints, the efficiency of models and hardware will be as important a criterion as nominal power.

Finally, the deal highlights the increasingly blurred boundary between consumer electronics, cloud and industry. The same models can be trained in a datacenter, fine-tuned in an enterprise infrastructure and then run on an endpoint. Nvidia and MediaTek sit on either side of this chain. Their tie-up could reinforce an already visible trend: decisions concerning AI will no longer be made solely by chief information officers or data teams, but also by product teams, hardware engineers and supply-chain managers.

Toward an Nvidia less dependent on the GPU-only model

The $3.5 billion investment in MediaTek must be interpreted in light of a lasting transformation in the sector. The AI market is not heading toward a single architecture. GPUs will remain fundamental for many complex tasks, for training and for environments where flexibility is essential. But specific chips should gain importance where volumes are high, tasks predictable and energy constraints strong.

Nvidia appears to be seeking to avoid a scenario in which the largest buyers of computing gradually become their own suppliers. Google, AWS, Microsoft and Meta have the financial means, workloads and teams needed to develop accelerators suited to their uses. They do not need to replace every Nvidia GPU to influence the market: it is enough for them to absorb internally a significant share of the most profitable or most recurrent workloads.

In this context, the best defense for Nvidia is not necessarily to deny the value of custom chips. It is to become useful to their development. MediaTek brings design capabilities that can broaden Nvidia's role in connected devices and infrastructure. The American group can thus seek to position itself not only as an accelerator seller, but as a partner across the entire chain from the datacenter to the endpoint.

This evolution is not without risk. Industrial partnerships require precise alignment of interests. MediaTek retains its own customers, markets and priorities. Major technology groups, for their part, will want to preserve their strategic independence. Nvidia will therefore have to demonstrate that its involvement in customized architectures delivers greater value than fully in-house design or the use of other partners.

Success will also depend on the ability to make the hardware and software layers work together. Nvidia's historical advantage rests largely on a widely used development ecosystem. As chips diversify, ease of programming, model porting and performance optimization will become major differentiating factors. Customers will not choose a chip based only on its specifications: they will assess the time needed to integrate it into their services and AI tools.

For TechCrunch, this bet on MediaTek therefore reveals a broader ambition in the face of the rise of accelerators developed by Big Tech. Nvidia wants to retain influence in the segments where its customers are specifically seeking to emancipate themselves. The deal illustrates a form of pragmatism: rather than exclusively defending a model based on general-purpose GPUs, the company is investing in expertise that can open the door to specialized chips and embedded systems.

Over the long term, the question will be whether this strategy truly transforms the balance of power with hyperscalers. If cloud giants generalize their own components, Nvidia will have to prove it can capture value through means other than selling its most visible platforms. If, conversely, customized architectures remain complementary to GPUs, the investment in MediaTek could give the group a stronger presence at several levels of the market.

For French and European companies, this dynamic above all signals a multiplication of technical options, but also greater complexity. Tomorrow's AI will be distributed across datacenters, sovereign or public clouds, industrial sites and connected devices. The tie-up between Nvidia and MediaTek could contribute to this reshaping, in which hardware will no longer be merely an invisible support for software, but a strategic choice that determines cost, autonomy and data control.

The next step will therefore not be measured solely by the scale of the announced investment. It will play out in the products, architectures and tools that result from it. Nvidia is clearly seeking to remain at the center of a market where major customers no longer want merely to buy computing: they want to design it, integrate it and control it. The alliance with MediaTek is a response to this ambition for independence, and perhaps a sign that the AI battle is gradually shifting from the GPU to the entire system surrounding it.