Author: Xiao Meng
TrendForce recently released the Top 10 Global Fabless Companies ranking for the second quarter of 2026, presenting a highly impactful report card to the industry. The list shows that the top ten fabless companies achieved a combined revenue of USD 141.45 billion, a year-on-year increase of 73%. While the 73% YoY growth rate is certainly eye-catching, it is merely an aggregate result. What deserves more attention than this figure is exactly who is growing, by how much, and where the growth is coming from. Looking further down the list, several industry shifts are already clearly visible.
01. What Exactly Happened in the Top 10?
Looking at the entire list, combined revenue continues to grow, but the distribution of the increment is significantly uneven. NVIDIA remains at the top, accounting for 64% of the total revenue of the top ten, leaving the other nine far behind. Broadcom holds steady at second place; although its revenue is much lower than NVIDIA's, its YoY growth rate ranks first, exceeding 110%. AMD surpassed Qualcomm to rise to third place with USD 11.536 billion, while Qualcomm dropped to fourth, with both YoY and QoQ growth rates declining. The remaining positions on the list are MediaTek, Marvell, Realtek, MPS, Novatek, and OmniVision.
The middle and lower tiers are equally noteworthy. Marvell achieved a 34% YoY growth and a 12% QoQ growth; MPS saw a 48% YoY growth and a 22% QoQ growth. Although their revenue scale is not as large as the top players, the quality of their growth is relatively high. Overall, the "reshuffle" in this list is not only reflected in the rise and fall of rankings: NVIDIA has further widened the gap, AMD has completed its overtaking of Qualcomm, and the middle and lower-tier manufacturers show a more pronounced divergence in growth rates.
02. How Did NVIDIA Capture 64% of the Revenue?
We previously mentioned that NVIDIA is in a position where the other nine "cannot even see its taillights." So, what has enabled it to form its current leading advantage?
There are mainly two answers.
First, it still stems from the sustained growth in AI GPU demand. Given the continuous expansion of large model training, inference, and Agentic AI applications, cloud service providers, AI companies, and enterprise customers are continuously increasing their investments in AI infrastructure. GPUs remain the primary computing accelerators in current AI data centers. TrendForce points out that in addition to traditional hyperscale cloud service providers, NeoClouds, sovereign AI projects, and enterprise customers are also beginning to contribute more to NVIDIA's revenue. The expansion of the customer base further supports GPU demand.
Looking at NVIDIA's own financial report, the AI data center remains the company's primary source of growth. NVIDIA's total revenue for the quarter reached USD 96.22 billion, of which data center business revenue reached USD 89 billion, a YoY increase of 117%, accounting for approximately 92% of the company's total revenue. This business scale has far surpassed other businesses such as gaming, professional visualization, and automotive.
NVIDIA's advantages also stem from its AI infrastructure layout beyond GPUs. Given the continuous expansion of AI computing scale, high-speed data exchange is required among multiple GPUs. The communication efficiency between GPUs, CPUs, networking equipment, and other computing nodes will also directly affect the overall system's operational efficiency. Therefore, NVIDIA continues to strengthen its high-speed interconnect and networking capabilities, integrating GPUs, NVLink, networking chips, and complete systems to gradually form an AI infrastructure solution covering computing, interconnects, and systems.
Among them, NVLink is the key technology in this system. It is a high-speed interconnect technology used to improve data transmission efficiency between GPUs and between processors and accelerators. In NVIDIA's AI computing platform, NVLink can connect multiple GPUs to form a larger-scale high-speed computing domain; NVLink-C2C further undertakes high-speed, low-latency interconnects between processors and accelerators. Given the continuous expansion of AI cluster scale, the importance of high-speed interconnects also increases. Building on this, NVIDIA further launched NVLink Fusion, extending NVLink capabilities to third-party custom chips. Through this platform, partners can develop custom CPUs and XPUs and connect them to NVIDIA's NVLink-based AI infrastructure. NVIDIA provides supporting capabilities such as interconnects, networking, and rack-scale systems, while partners can design processors for specific workloads.
Currently, Marvell has joined the NVLink Fusion ecosystem to provide relevant interconnect and networking solutions for custom XPUs; AI inference chip company d-Matrix has also announced the adoption of NVLink Fusion to connect its next-generation Raptor inference processors to NVIDIA's AI infrastructure. NVIDIA has also engaged in relevant cooperation with MediaTek, with MediaTek developing custom XPUs for data center customers.
From GPUs to NVLink, and then to networking and rack-scale systems, NVIDIA is continuously expanding the coverage of its AI infrastructure. For customers, this means that the deployment of AI computing power can be carried out around a more unified hardware and software system, and NVIDIA has thus further strengthened its platform position in AI data centers.
03. AMD Rises to Third: AI Dividends Transmit to CPUs
If NVIDIA is reaping the most direct wave of dividends from AI computing power demand, the change in AMD's ranking reflects that this demand is transmitting to broader data center computing segments.
Looking at AMD's own financial report, what is attractive is not only the quarterly revenue reaching USD 11.5 billion, but also the data center revenue reaching USD 6.7 billion, accounting for 58% of the total and a YoY increase of 107%. CEO Lisa Su specifically emphasized that the "success" of the data center this time is largely due to the strong market demand for EPYC processors and Instinct GPUs. In other words, AMD's performance growth this time cannot be entirely attributed to CPUs; GPUs are also an important source of growth.
However, alongside the rapid growth of the data center business, AMD's server CPUs have also seen continuous expansion. Server CPU revenue has hit a record high for the fifth consecutive quarter, with related sales to cloud and enterprise customers increasing by over 70% YoY. Meanwhile, with the continued expanded deployment of the 5th Gen EPYC Turin and 4th Gen EPYC Genoa, AMD's revenue share in the x86 server market continues to rise. The 5th Gen EPYC Turin now covers nearly one-third of the more than 1,600 EPYC public cloud instances globally, and cloud service providers are using it for more workloads such as databases, storage, and AI.
The logic behind this is not entirely the same as how GPUs benefit from AI demand. The demand for parallel computing power in large model training and inference first directly drove up the procurement of GPU accelerators, making GPUs the first chips to feel the demand growth in the expansion of AI infrastructure. As more and more GPUs enter data centers, the server systems running around these accelerators also need more CPUs to handle data processing, task scheduling, storage access, and system management. Thus, AI computing power demand begins to transmit to general-purpose computing segments.
AMD has already begun to adjust its product layout around this change. The company launched the 6th Gen EPYC server processor this year, explicitly targeting it at Agentic AI, general-purpose computing, and enterprise workloads. At the same time, AMD is also integrating EPYC CPUs, Instinct GPUs, and networking products into a more complete AI infrastructure through the Helios rack-scale AI system.
AI first drives a rapid increase in GPU demand, and then transmits the increment to general-purpose computing segments such as CPUs through data center scale expansion and workload changes. Happening to possess both GPU and CPU product lines allows AMD to capture these two layers of demand, which also becomes an important support for the continuous expansion of its data center business.
04. Mobile Chip Manufacturers Also Start to Grab a Slice of the AI Computing Power Cake
Compared to NVIDIA and AMD, Qualcomm and MediaTek have chosen to "take a different path."
The original businesses of both companies are mainly focused on mobile phone chips. However, given the increasing pressure in the mobile phone chip market, AI is becoming an important breakthrough for both to find growth space.
Qualcomm's route is relatively diversified, with automotive, IoT, and data centers becoming the growth directions where Qualcomm continues to exert effort.
According to Qualcomm's financial report, its revenue for the quarter reached USD 9.947 billion, of which USD 1.588 billion came from the automotive portion of QCT revenue. This part saw a 61% YoY increase, and the automotive business has maintained double-digit YoY growth for 23 consecutive quarters; IoT brought in USD 1.83 billion in revenue, a 9% YoY increase. Regarding this, Qualcomm stated that the automotive segment is expected to reach USD 10 billion by 2029, and the IoT segment will exceed USD 14 billion.
Moreover, Qualcomm is treating the data center as an important growth direction for the next stage.
In June this year, Qualcomm officially unveiled its complete AI roadmap for data centers, setting a business revenue target of over USD 15 billion for the data center in fiscal 2029. The company launched a series of brand-new data center solutions, including the Dragonfly C1000 CPU, High Bandwidth Compute (HBC), Dragonfly AI300 inference accelerator, and connectivity products. It also provides custom chip solutions and further acquired AI software platform company Modular, hoping to extend from chips to hardware-software collaborative platform solutions.
MediaTek's route is more concentrated on AI ASICs. Compared to continuing to rely on mobile SoCs, MediaTek is leveraging its accumulation in high-speed SerDes, low-power processors, and connectivity technologies to develop custom AI chips for cloud service providers. At the Q1 2026 earnings call, the company raised its 2027 ASIC market size estimate to USD 70 billion to USD 80 billion, and expects this market to continue expanding as cloud service providers continuously increase their self-developed computing power.
This quarter, MediaTek further raised its data center business target. The company disclosed that the 2026 data center revenue target has been increased from over USD 1 billion to over USD 2 billion, and the first AI accelerator ASIC for large US cloud service providers is expected to enter mass production in the fourth quarter. MediaTek's Smart Edge business revenue accounted for 53% this quarter, exceeding the 41% of the Mobile Phone business for the first time, which also reflects that the company's business focus is gradually expanding beyond mobile phones.
More noteworthy is that MediaTek has established deeper AI infrastructure cooperation with NVIDIA. On August 31, NVIDIA announced an investment of USD 3.5 billion in MediaTek, and the two parties will further cooperate around NVLink Fusion. MediaTek's customers can develop custom AI chips based on this and connect them to NVIDIA's data center infrastructure. This means that while developing ASICs, MediaTek can also leverage NVIDIA's existing interconnect and system ecosystem to accelerate its entry into the AI data center market.
05. Traditional Consumer Electronics Chip Manufacturers Are Also Looking for AI Entry Points
The industrial increment brought by AI has not stopped solely at core computing power chips such as GPUs, CPUs, and ASICs. Given the continuous expansion of AI server scale, infrastructure segments such as power management, optical communication, and network connectivity are also simultaneously gaining new demand. Some fabless companies that originally deeply engaged in the consumer electronics market are also entering new growth tracks leveraging their own advantages. MPS and OmniVision are excellent examples.
As a high-performance analog and power chip manufacturer, MPS is different from the manufacturers mentioned above; it does not directly participate in AI computing power competition. However, the continuous increase in AI server power consumption also means that power supply systems need to handle higher power density and more complex power conversion requirements. The increase in the number of GPUs and the rise in server power will ultimately translate into new demand for PMICs (Power Management ICs) and solutions.
This change is already reflected in MPS's performance. In Q2 2026, MPS revenue reached USD 981 million, a 47.6% YoY increase, hitting a new single-quarter high. Among them, the Enterprise Data business revenue reached USD 381 million, a 164.3% YoY increase and a 44.8% QoQ increase, becoming one of the company's fastest-growing businesses. MPS explicitly stated that the growth of the Enterprise Data business is mainly driven by the demand for AI-related power solutions, and the company is continuously expanding its power product layout for AI and data centers.
As AI data centers develop towards higher power density, MPS's products are also beginning to extend to higher-power power supply architectures. The company has started providing high-voltage AC-DC product samples for 800V data center architectures, further expanding its product coverage in the power supply segment of AI infrastructure. For MPS, the opportunities brought by AI have gradually extended from single PMIC demand to more complete data center power supply solutions.
OmniVision, on the other hand, demonstrates another "circuitous" growth path. According to the financial report, its Q2 revenue reached approximately USD 1.135 billion (CNY 7.611 billion), a 18.65% QoQ increase. Although the image sensor solutions achieved revenue of CNY 9.254 billion, accounting for 66.05% of main business revenue, it decreased by 10.55% YoY; emerging markets performed strongly and grew rapidly, with a combined revenue of CNY 1.726 billion, a 47.12% YoY increase. Among them, professional imaging terminals reached CNY 1.287 billion (a 53.36% YoY increase), machine vision and robotics reached CNY 210 million (a 71.15% YoY increase), and edge AI reached CNY 230 million (an 8.46% YoY increase). Meanwhile, OmniVision is also continuously deploying in directions such as optical communication and AI computing infrastructure, further expanding its analog chip business layout.
For OmniVision, the market space for traditional businesses such as mobile phone and automotive CIS is being squeezed. The company is accelerating its layout in emerging tracks, continuously expanding new application scenarios from professional imaging, machine vision, and edge AI to optical communication and AI computing infrastructure, seeking to build a second growth curve.
06. AI Is Redefining the Growth Logic of Fabless Companies
Looking at the entire list, different manufacturers have given their own unique answers to the AI exam paper. Their answers also represent different development trends of the entire industry.
The first answer is certainly continuing to bet on general-purpose computing power, which can be said to be the most stable and foolproof choice. The expansion of AI model scale and the increase in computing power demand provide continuous market space for general-purpose GPUs and CPUs. NVIDIA and AMD are continuously expanding their data center GPU businesses. TrendForce also points out that the revenue growth of the top ten global fabless companies in Q2 2026 is mainly driven by the rising demand for GPUs, CPUs, ASICs, and interconnect products.
The second answer is extending towards customized computing power. Broadcom, Marvell, and MediaTek are all strengthening their ASIC-related layouts. Given that large cloud service providers are increasingly developing self-developed AI chips, custom ASICs can be optimized for specific models and workloads, finding a balance of performance, power consumption, and cost that is more suitable for their own businesses. TrendForce expects that cloud service providers such as Google and AWS will continue to accelerate the deployment of self-developed ASICs in 2026, and the proportion of ASICs in AI servers will further increase.
The third answer is extending towards AI infrastructure. The changes in Qualcomm and MediaTek represent that fabless manufacturers are bringing computing, connectivity, and chip design capabilities into the larger data center market. Given the continuous expansion of AI computing power scale and the rising market demand, expanding from single accelerator chips to CPUs, interconnects, connectivity, and system-level solutions, the competitive boundaries of fabless companies are further broadened accordingly.
The final answer is exploring emerging markets to seek growth space. The layouts of MPS and OmniVision illustrate that the industrial increment brought by AI is still continuing to diffuse to other segments of the industry chain, intersecting with existing markets such as machine vision, professional imaging, optical communication, and power management. For fabless companies, new demand often extends outward along existing technical capabilities, thereby forming new application scenarios and growth space.
07. Conclusion
AI is further spreading the growth space of the chip industry to different segments. GPUs undertake the most direct AI computing power demand. CPUs gain more increments given the expansion of data center scale and new AI workloads. ASICs continue to develop along the demand for customized computing power. Infrastructure chips such as power management, networking, and optical communication are also beginning to share the market opportunities brought by the expansion of AI data centers.
Ranking changes are merely the result; more importantly behind this is that AI is redistributing the growth opportunities in the chip industry. For fabless companies, the future competitive boundaries are continuously broadening. General-purpose computing power, custom chips, infrastructure, and hardware-software ecosystems could all become new growth fulcrums.