
NVIDIA Computex Preview: What AI Computing Power "Bombs" Will Jensen Huang Unveil?
Morgan Stanley expects NVIDIA to comprehensively showcase its new generation of AI computing architecture, centered on the Rubin GPU and Vera CPU, at Computex 2026, with a focus on advancing AI factory solutions aimed at achieving the "lowest cost per token." The power consumption plan for the Rubin GPU and its HBM4e packaging design have become focal points, while the Vera CPU is viewed as a new growth driver opening up a $20 billion market
Morgan Stanley expects that at the upcoming Computex exhibition, NVIDIA CEO Jensen Huang will comprehensively showcase the new generation of AI computing architecture centered on the Rubin GPU and Vera CPU, further consolidating its dominant position in the AI infrastructure sector.
According to Zhuifeng Trading Desk, COMPUTEX 2026 will be held in Taipei from June 2 to 5 under the theme "AI together," with market focus rapidly shifting toward NVIDIA's rack-level system designs and advancements in advanced packaging.
In its latest research report, Morgan Stanley expects that NVIDIA will showcase AI factory solutions aimed at achieving the lowest cost per token, a move that is directly driving a reassessment of supply chain capacity and bringing significant growth expectations to related hardware manufacturers.
Driven by strong demand for AI GPUs and CPUs, TSMC is accelerating the expansion of its CoWoS advanced packaging capacity. Morgan Stanley has significantly raised its forecast for TSMC's CoWoS capacity in 2027 and maintains a positive outlook on the investment prospects of NVIDIA's core supply chain.
As the exhibition approaches, investors are closely monitoring the power control solutions for the Rubin chip and the commercialization potential of the Vera CPU. These technical details will determine the direction of capital expenditure and the competitive landscape in the global AI computing market over the next few years.
Rubin Rack Design and Power Solutions
Morgan Stanley analysts, including Charlie Chan, pointed out in their report that one of the key highlights of Computex is the server rack design for the Rubin series. NVIDIA is expected to showcase a combination of NVL, LPU, Vera CPU, BlueField, and NVSwitch racks, aiming to provide the lowest cost per token for future "AI factories."
Addressing market concerns about whether the Rubin GPU can achieve its 2.3kW Thermal Design Power (TDP) target, Morgan Stanley revealed that Rubin has reverted to its original chip cooling solution, achieving a TDP of 1.8kW. However, through coordinated cooling within the server system, the 2.3kW target remains achievable.
Regarding packaging design, for the Rubin Ultra version, NVIDIA needs to adopt new masks in 2027 to introduce HBM4e memory. To maintain good chip yield and address warpage issues, Rubin Ultra is expected to maintain a two-die-per-package design, rather than the previously rumored four-die configuration.
Vera CPU Opens Up a $20 Billion Market Opportunity
In addition to GPUs, NVIDIA's expansion into the CPU sector is also attracting significant attention. Joe Moore, Morgan Stanley's US semiconductor analyst, expects NVIDIA to disclose more details about the Vera CPU at the exhibition, representing a market opportunity worth up to $20 billion.
Supply chain surveys indicate that TSMC is allocating additional CoWoS-R capacity and 3nm wafers for the Vera CPU. Based on current capacity planning, Morgan Stanley believes a reasonable shipment assumption for standalone Vera CPUs is 1.5 million units.
The firm expects that end customers for the Vera CPU will include major tech companies and cloud service providers such as Microsoft, Meta, CoreWeave, and Oracle. The launch of this product will further perfect NVIDIA's closed-loop computing ecosystem, posing direct competitive pressure on traditional CPU manufacturers.
Significant Upward Revision to TSMC CoWoS Capacity Forecast
Strong demand for AI chips is reshaping the landscape of the upstream foundry and packaging markets. In its report, Morgan Stanley significantly raised its forecast for TSMC's CoWoS capacity in 2027 from 160,000–170,000 wafers per month to 200,000 wafers per month.
To meet surging orders, TSMC is converting 28/22nm capacity space at its Fab 15A plant to produce 55nm interposers, expecting to add approximately 10,000 wafers per month in capacity this year. Due to this capacity crowding-out effect, UMC is expected to take on overflow orders for Sony's 28/22nm Image Signal Processors (ISPs) in 2027, leading Morgan Stanley to maintain its "Overweight" rating on UMC.
Regarding SoIC (System on Integrated Chips) capacity, although Morgan Stanley slightly adjusted its capacity construction forecasts for 2027 and 2028 to 40,000 and 70,000 wafers per month respectively, the actual production volume forecasts remain unchanged at 30,000 and 60,000 wafers per month respectively. Giants such as Apple, AMD, and NVIDIA are all core customers of SoIC technology.
Based on optimistic expectations for the Rubin and Vera product lines, Morgan Stanley maintains its "Overweight" rating on NVIDIA's core supply chain.
