MediaTek built its empire inside mobile phones. Its new multibillion-dollar push into custom AI chips shows that the fight over artificial intelligence is moving beyond general-purpose GPUs.
MediaTek is best known for designing the processors inside smartphones, televisions, routers and consumer electronics.
Now the Taiwanese chipmaker wants a place inside the world’s largest AI data centers.
MediaTek has approved a $5 billion financing program to support long-term expansion, with custom AI chips among its main targets. The company expects more than $2 billion in data-center AI chip revenue during 2026 and believes the market could reach $80 billion in 2027. It is aiming to capture 15% to 20% of that opportunity.
A Reuters report on MediaTek’s financing and expansion plan details the company’s move into custom data-center silicon.
That does not mean MediaTek is preparing to sell a direct replacement for Nvidia’s latest GPU.
Its strategy is more targeted—and potentially more disruptive.
The Alternative to Nvidia Is a Chip Built for One Customer
Nvidia’s GPUs are designed to support many models, frameworks and computing workloads. That flexibility helped make them the default hardware for the AI boom.
A custom application-specific integrated circuit, or ASIC, takes a different approach. It is designed around a narrower set of tasks for a particular customer or computing environment.
The customer may be a cloud provider operating millions of AI requests. By eliminating functions it does not need and optimizing the hardware around its models, networking and software, the provider may reduce power consumption and the cost of generating each AI response.
MediaTek says its data-center business can supply more than the processor alone. Its offering includes custom ASICs, high-speed interconnect technology, advanced packaging and rack-level system design.
The company describes this as an end-to-end data-center platform rather than a return to selling isolated chips. MediaTek’s overview of its data-center strategy says it is targeting a $70 billion to $80 billion market and has raised its 2026 ASIC revenue guidance to $2 billion.
That approach uses skills MediaTek developed in smartphones: combining computing, communications and power efficiency inside tightly integrated systems.
A phone processor must perform substantial work without draining the battery or producing excessive heat. Those same constraints now matter inside AI data centers, where electricity and cooling have become major operating costs.
Google Is Designing Hardware Around Gemini
MediaTek is not alone in pursuing specialized AI silicon.
Google has designed Tensor Processing Units for more than a decade. Its current TPU families are built for both training models and serving them to users.
Google is reportedly taking specialization further with a server chip known internally as Frozen v2. The project would incorporate elements optimized specifically for Gemini, potentially delivering six to ten times more generated tokens per unit of power than Google’s existing AI hardware.
The logic is straightforward.
When a company operates one model at enormous scale, even a modest efficiency improvement becomes valuable. Saving a fraction of a cent on one AI request means little. Saving it across billions of requests can determine whether a consumer AI service becomes profitable.
This is pushing the industry toward hardware-and-model co-design. Engineers creating the model work directly with engineers creating the chip, allowing both sides to be optimized together.
Nvidia’s Position Is Still Formidable
The custom-chip trend should not be confused with Nvidia suddenly losing the market.
Nvidia reported $75.2 billion in quarterly data-center revenue in May, up 92% from the previous year. Its advantage includes more than processors: CUDA software, networking products, complete server systems, developer tools and a large ecosystem of engineers trained on its technology.
A company can purchase Nvidia systems and begin running many established AI frameworks relatively quickly. Developing a custom chip may require years, billions of dollars and enough computing demand to justify the investment.
That restricts the strongest custom-chip economics to hyperscale cloud providers, frontier AI laboratories and other customers operating enormous workloads.
The threat to Nvidia is therefore unlikely to arrive as one competitor selling a universally superior processor.
It will arrive through large customers gradually moving predictable workloads onto their own hardware while continuing to buy Nvidia GPUs for frontier training, unfamiliar models and applications requiring flexibility.
Nvidia may retain a dominant share of a rapidly expanding market while losing portions of the most repetitive and cost-sensitive work.
The Smartphone Market Is Driving the Transition
MediaTek has another reason to move quickly.
Its mobile-chip revenue recently fell about 20% amid weak smartphone demand and higher component costs. The global smartphone market also suffered its sharpest quarterly contraction in more than a decade.
Data-center AI offers a larger, faster-growing market and customers willing to spend aggressively for improved efficiency.
Arm is following a related path. The architecture behind most smartphone processors is increasingly being adopted in cloud servers and AI infrastructure because of its power efficiency.
The companies that learned to place powerful computing inside a battery-operated device now see an opportunity to reduce the energy cost of warehouse-sized computers.
What This Means for Businesses Buying AI
Most companies will never purchase an AI chip directly. They will experience this competition through cloud prices, model availability and software performance.
More custom silicon should eventually create additional computing capacity and reduce reliance on one hardware supplier. Cloud platforms may offer different pricing tiers based on whether a workload runs on Nvidia GPUs, proprietary accelerators or lower-cost inference chips.
Businesses should avoid designing applications around hardware they do not control. Their priority should be measuring response quality, speed, reliability and cost per completed task across the cloud services available to them.
Nvidia created the infrastructure foundation for the current AI boom. MediaTek, Google, Amazon, Microsoft and other chip designers are now searching for the profitable pieces they can remove from that foundation and operate themselves.
The smartphone chipmakers are not storming Nvidia’s castle through the front gate.
They are quietly redesigning the plumbing underneath it.
