Despite numerous headwinds currently facing the AI industry outlook, it is undeniable that the AI sector—especially AI servers—will continue to be a cornerstone of Taiwan's ICT industry in 2025. For key players following Nvidia's trajectory, including Foxconn, Quanta, and Wistron, AI servers are expected to drive their growth throughout 2025.
Regarding recent reports about design changes to the GB300, related companies have stated that product design modifications during early development stages are quite common for various reasons. As long as these changes occur before mass production begins, they are considered normal practice.
After all, any adjustments made prior to mass production are far preferable to having to implement fixes after products have already been shipped.
Reversion to socketless GB300 design raises supply chain concerns
Due to unresolved signal transmission issues failing to meet customers' stringent standards during testing, the originally planned socket-enabled GB300 has reportedly reverted to a socketless design similar to the GB200. This shift affects not only suppliers of sockets and NVQD components but also raises concerns about potential impacts on final system assembly and shipment schedules.
In response, major Nvidia system assembly partners such as Foxconn, Quanta, and Wistron have refrained from commenting. However, insiders point out that new products must undergo repeated testing and revisions before entering mass production to ensure the final product meets performance expectations.
This is especially true for high-tech, precision, and high-cost products like AI servers, where rigorous demands on performance and quality require every design element to be thoroughly tested before finalization.
Moreover, each design change often triggers linked adjustments elsewhere, so manufacturers accept that any design modifications before the final production phase are manageable. In fact, there are precedents for making further corrections even after delivery but before actual customer deployment.
Industry sources stress that the critical point is that all corrections and rework should ideally happen before handing over products to clients, rather than recalling them post-deployment.
As for the GB300 design reverting to the GB200 model, suppliers believe this will benefit Quanta and Foxconn, who have already mass-produced the GB200 NVL72. Returning to the GB200 design will facilitate smoother mass production and likely allow them to maintain their positions as lead manufacturers for GB300 rack systems. For Nvidia, this approach could accelerate the GB300 launch and avoid delays experienced with the GB200 that impacted cloud service provider customers.
Foxconn sees AI server revenue outpacing iPhones
Foxconn chairman Young Liu previously remarked that within two years, AI server revenue would surpass iPhone sales. In 2025, cloud network products might even approach the revenue share of consumer smart devices.
In 2024, Foxconn's AI server revenue surged more than 150%year-over-year. The company aims to exceed NT$1 trillion (US$30.8 billion) in AI server revenue for 2025, targeting over 50% of its total server revenue and capturing more than 40% market share in the overall AI server segment.
Quanta projects strong AI server growth in 2025
Quanta has declined to comment on chip supplier design changes. According to its recent earnings call, Quanta expects robust AI server growth in 2025, with GB200 NVL racks entering mass production starting in the first quarter. Annual AI server growth is projected to multiply significantly.
Quanta reported first-quarter revenue of NT$485.671 billion, up 16.36% quarter-over-quarter and 87.56% year-over-year, setting a single-quarter record and breaking the traditional slow season pattern, largely driven by AI server demand. Shipments of Quanta's GB200 NVL72 units are reportedly proceeding smoothly, with second-quarter deliveries expected to increase further.
Wistron expands global capacity amid soaring demand
Wistron holds a crucial role in the GB200 system as a compute board supplier alongside Foxconn, accounting for over half of the supply volume. Wistron remains optimistic about AI server growth in 2025, forecasting triple-digit percentage increases amid strong customer orders exceeding supply.
To meet client demand, Wistron is actively expanding AI-related capacity worldwide, including facilities in Taiwan, Mexico, and Texas, US. Recently, Nvidia announced plans to build an AI supercomputer domestically in the US with supply chain partners, naming Wistron among the selected vendors.
Article translated by Charlene Chen and edited by Jack Wu
