By 2025, AI is no longer a new trend, but rather an infrastructure. Computing power has become the lifeblood of the digital economy, and this algorithm-driven transformation is fundamentally reshaping the power structure of the global semiconductor industry.
Once, NVIDIA built an almost monopolistic moat in the AI training market with its CUDA ecosystem and GPU architecture—holding over 90% market share and surpassing a market value of $4.5 trillion, making it the most valuable chip company in the world. However, as AI transitions from the laboratory to large-scale deployment, moving from training to inference, edge, and endpoint, the ‘golden age’ of general-purpose computing is quietly receding. In its place is a revolution centered on specialization, energy efficiency, and vertical integration, known as the ‘de-GPU’ revolution.
In this changing landscape, former giants are no longer following in lockstep; instead, they are betting on Arm architecture and ASICs (Application-Specific Integrated Circuits), attempting to carve out a second growth curve away from NVIDIA’s shadow. A new game of ‘one strong, many strong’ has already begun.
Intel: From IDM Predicament to ASIC Breakthrough
Once upon a time, Intel was the absolute ruler of the x86 world. However, process delays and a lag in AI strategy left it nearly absent in the early waves of AI. Faced with NVIDIA’s strong rise, Intel chose not to confront it head-on but to take a different path— transforming its IDM (Integrated Device Manufacturing) model into a customized service advantage.
In 2024, Intel established the Central Engineering Group (CEG), led by former Cadence executive Srini Iyengar, targeting the booming ASIC market. CEO Pat Gelsinger clearly stated, ‘We will provide customers with a one-stop solution from IP, design to manufacturing and packaging.’ This means Intel is no longer just selling chips but is selling ‘computing power delivery capabilities.’
Its unique advantages include: possessing x86 IP, advanced packaging technologies (such as Foveros), and an 18A process that is ramping up. Although NVIDIA publicly praises TSMC’s ‘limitless magic,’ it is unlikely to become a major customer for Intel in the short term. However, for cloud providers like Google and Microsoft seeking supply chain diversification, Intel’s ‘full-stack control’ is highly attractive.
More subtly, NVIDIA recently acquired a 4% stake in Intel for $5 billion, and both parties announced a joint development of a customized x86 CPU for AI infrastructure. On the surface, this appears to be cooperation, but it actually conceals competitive tension—Intel may gradually exit the high-end GPU battlefield, focusing instead on CPUs and ASIC foundry. The future of its Arc GPU may be limited to low-power edge scenarios.
Qualcomm: From Mobile Chips to Data Centers, Leveraging NPU to Tap into a Trillion-Dollar Market
If Intel is seeking change through defense, Qualcomm is making a bold cross-industry assault.
This mobile giant, known for its Snapdragon chips, is making a significant entry into AI inference data centers with its long-developed Hexagon Neural Processing Unit (NPU). The AI200, set to launch in 2026, and the AI250 in 2027, will not only support liquid-cooled racks and 160kW high-power designs but will also be equipped with 768GB LPDDR memory—far exceeding similar products.
Qualcomm’s strategy is very clear: to avoid the training market dominated by NVIDIA and focus on the inference track. As the scale of large model deployments grows exponentially, the demand for inference computing power is expected to exceed training by more than ten times before 2030. Qualcomm’s NPU has already demonstrated low power consumption and high energy efficiency on mobile devices, and now it is ‘amplifying’ this to data centers at just the right time.
More critically, there is innovation in the business model: Qualcomm not only sells complete systems but also opens up chips and components, allowing hyperscale customers to ‘mix and match’. It even does not rule out the possibility of NVIDIA or AMD purchasing its interconnect or security modules in the future. This open stance is key to breaking the closed ecosystem.
Saudi Arabia’s Humain company has already committed to deploying a 200MW system, becoming Qualcomm’s first heavyweight customer. Its software stack’s native support for mainstream frameworks like PyTorch, vLLM, and LangChain significantly reduces migration costs for developers. Qualcomm is attempting to prove: the future of AI does not necessarily have to run on GPUs.
MediaTek: From Second-Tier Mobile to First-Tier Cloud
MediaTek’s transformation is more subtle yet more disruptive. This company, long regarded as a representative of ‘cost performance’ in mobile chips, has now become an ASIC partner for top cloud service providers like Google and Meta.
Its killer feature is high-speed SerDes technology—the 224G SerDes has completed silicon validation, supporting HBM4E memory integration and advanced packaging, meeting the extreme requirements of AI chips for bandwidth and energy efficiency. In a direct confrontation with Broadcom, MediaTek successfully secured the order for Meta’s next-generation inference ASIC ‘Arke,’ utilizing TSMC’s 2nm process, expected to be mass-produced in 2027.
Even more noteworthy is its deep binding with NVIDIA. The jointly designed GB10 Grace Blackwell superchip will power NVIDIA’s DGX Spark personal AI workstation. This chip integrates the Blackwell GPU with the Arm architecture Grace CPU, achieving 1 PFLOP computing power and 128GB unified memory, allowing developers to run 200 billion parameter models on their desktops.
This collaboration not only opens the door to the high-end AI market for MediaTek but also marks its transition from a ‘consumer-grade chip manufacturer’ to a ‘cloud infrastructure supplier.’ In an era where CSPs (Cloud Service Providers) increasingly value TCO (Total Cost of Ownership), MediaTek’s high-cost performance ASIC solutions are becoming a significant variable in breaking NVIDIA’s monopoly.
AMD: Arm’s Return, Edge Breakthrough
In contrast, AMD appears more cautious. Although its MI300 series GPUs have gained some market share in AI training, it still struggles to shake NVIDIA’s ecological barriers.
Thus, AMD quietly restarted its Arm strategy. According to leaked information, its Arm architecture APU, codenamed ‘Sound Wave,’ is set to be released in the second half of 2026, featuring a compact package of 32mm x 27mm, integrating a 6-core CPU (2P+4E) and RDNA GPU, clearly targeting AI PCs and edge devices.
This is not AMD’s first attempt at Arm. Back in 2016, its A1100 server CPU failed due to an undeveloped ecosystem. Now, with a vastly different background—Arm’s acceptance in servers and endpoints has significantly increased, and the chiplet architecture has made heterogeneous integration possible—AMD has sufficient financial resources to advance both x86 and Arm lines simultaneously.
This move is both defensive and offensive—while NVIDIA and Intel join forces to solidify the x86 AI ecosystem, AMD is opening a new battlefield with Arm, avoiding being completely locked into the traditional PC/server red sea.
Why the Collective ‘Change of Heart’? AI Enters Deep Waters
The collective shift of these giants is not coincidental but a necessary result of the evolution of AI development logic.
Early AI relied on the general parallel computing capabilities of GPUs, which, due to their flexibility, could adapt to various model training. However, as AI enters the stage of large-scale deployment, energy consumption, latency, cost, and security have become core metrics. At this point, the ‘large and comprehensive’ nature of general-purpose GPUs has become a burden.
ASICs achieve hardware-level customization, precisely directing transistors to specific tasks, achieving tenfold or even hundredfold energy efficiency improvements; Arm, with its low power consumption, licensable, and modular characteristics, has become the ideal carrier for edge and endpoint AI. The combination of the two forms the golden combination of ‘dedicated computing.’
Deeper changes lie in: the value focus of chips is shifting from ‘products’ to ‘capabilities’. Cloud providers are no longer satisfied with purchasing off-the-shelf chips; they want to deeply participate in defining—from IP selection, architecture design to manufacturing and packaging. Whoever can provide end-to-end customized service capabilities will bind the best customers.
The Next Decade: The Era of Dedicated Computing
The glory of GPUs will not end, but a silent revolution has already occurred at the bottom of AI infrastructure. The future landscape of computing power will present a ‘dual-track parallelism’:
- High-end training: still dominated by NVIDIA, pursuing extreme performance and ecological closure;
- Inference, edge, and endpoint: led by ASIC+Arm, pursuing extreme energy efficiency and scene adaptation.
On this new track, Intel relies on IDM integration, Qualcomm extends with NPU, MediaTek breaks through with SerDes, and AMD tests the waters with Arm—each is redefining the meaning of ‘chip giant’ in its own way.
As AI transitions from ‘can it run’ to ‘is it cost-effective to run,’ specialization is efficiency, customization is competitiveness. This wave of ‘de-GPUization’ may determine who the true king of computing power will be in the next decade.