Insights from the ongoing 2026 Taipei International Computer Exhibition indicate that Guangfa Securities has received positive signals from several semiconductor industry leaders, including NVIDIA (NVDA.US), Qualcomm (QCOM.US), Intel (INTC.US), and Marvell Technology (MRVL.US).
Guangfa Securities analyst Jeff Pu noted that for NVIDIA, a key highlight is the RTX Spark superchip for PCs, but there are several other positive developments.
In a report to clients, Pu wrote: "NVIDIA emphasized that the Vera GPU is designed specifically for agent AI, aiming to address CPU bottlenecks in the agent AI era. Vera features 88 Olympus cores, 1.5TB of LPDDR5X memory, 1.2TB/s memory bandwidth, and 3.4TB/s inter-core bisectional bandwidth. According to NVIDIA, its single-core bandwidth is 3 times higher than x86 architecture, with agent sandbox performance 1.8 times higher."
Furthermore, Pu stated that he believes the standalone Vera CPU could generate $20 billion in revenue for NVIDIA (the company is currently in full production of its Vera Rubin product line), and adoption of GPUs by hyperscale data centers should increase when Vera Rubin NVL72 ramps up production in October.
Intel
Regarding Intel, Pu mentioned that the company has reiterated to partners, investors, and industry observers that it is engineer-led and that its 18A manufacturing process is now in volume production.
Pu explained: "Management stressed that agent AI is boosting CPU demand, restoring the strategic importance of x86. The company also introduced a rack-scale blueprint to support AI agents, demonstrated a more efficient decoupled inference architecture developed in collaboration with SambaNova and NVIDIA, and expanded its custom chip business through partnerships with Foxconn, Google, and Ericsson."
During the event, Intel CEO Pat Gelsinger reaffirmed that the company has held discussions with several external partners about joining its foundry business, though no announcements have been made yet. However, he added that if Intel raises its capital expenditure expectations, it would signify that they have secured customers.
The company made several announcements at the event, including a new chip-to-rack-level AI solution. There were also reports that Intel plans to launch a new AI data center chip by the end of the year.
Qualcomm
Pu pointed out that Qualcomm CEO Cristiano Amon stated at the event that 2026 is the year of the agent.
Pu wrote: "The company anticipates that global token consumption will grow from 32 billion per 10 seconds in 2026 to 1.27 trillion by 2030, an increase of approximately 40 times. This shift is expected to reshape computing architecture, where the CPU will no longer be limited to system management but will be responsible for task orchestration; GPUs and NPUs will handle AI models, inference, and content generation. Additionally, devices will increasingly become the perception and execution platforms for AI agents, benefiting not only from persistent context awareness and long-context understanding (which we expect will help reduce memory requirements) but also from seamless connectivity between the cloud, edge, and local devices."
Marvell Technology
Pu stated that at Computex, Marvell Technology emphasized its positioning as an end-to-end AI connectivity platform, rather than just a semiconductor supplier. (The company launched the Teralynx T100 Ethernet switch, the Ara series 1.6T DSP, and COLORZ 1600 products at the show.)
Pu explained: "In summary, these product launches solidify Marvell Technology's ambition to build a complete connectivity stack encompassing switching, optical components, and interconnect infrastructure for AI scaling."
Furthermore, NVIDIA CEO Jensen Huang expressed his view that Marvell Technology has the potential to become the "next trillion-dollar company," which Pu suggested could indicate deeper collaboration between the two firms.
Pu added: "We view NVLink Fusion as a significant long-term option for Marvell Technology, but its commercialization path may depend on whether hyperscale data center operators are willing to adopt NVIDIA's interconnect architecture in their custom AI chip roadmaps. We remain positive on Marvell Technology. Beyond the NVIDIA option, our primary reasons for a favorable view on Marvell Technology remain its market position in custom AI chips and connectivity."
Comments