NVIDIA's Next-Gen GPU 'Rubin' Moved Up to 2025, Expected To Create Major Industry Impact

According to Morgan Stanley, NVIDIA's next-generation Rubin GPU is already being prepared six months ahead of schedule, with the expected launch time moving from the first half of 2026 to the second half of 2025. Utilizing 3nm technology, CPO (Chip on Package Optics), and HBM4 (sixth-generation high-bandwidth memory), the new generation GPU's chip area will be double that of the previous generation Blackwell, and it is anticipated that companies in the related industry chain will benefit.

Morgan Stanley analyst Charlie Chan stated in a report on December 2nd that, despite the output of Blackwell chips still growing, considering the complexity of the new chip, TSMC and the supply chain have already begun preparations for the launch of NVIDIA's next-generation Rubin chip. The 3nm Rubin GPU is expected to enter the tape-out phase in the second half of 2025, about six months earlier than previously scheduled.

Morgan Stanley has said that due to the migration to 3nm process technology, and the adoption of CPO and HBM4, Rubin will be a powerful chip. According to supply chain surveys, the chip size of Rubin will be nearly double that of Blackwell, and the Rubin chip may contain four compute chips, twice that of Blackwell.

Analysts currently estimate that, due to the increased chip area, TSMC will further expand its CoWoS (Chip on Wafer Substrate) capacity in 2026. The company is expected to start placing orders for 2026 with equipment suppliers in the middle of 2025, depending on the sustainability of AI capital expenditures.

In terms of the supply chain, analysts believe that ASMPT is slightly lagging in the TCB tool certification for Rubin. TCB (Thermal Compression Bonding) is the process required for the CoW (Chip on Wafer) of the Rubin GPU. K&S previously announced that it had received TSMC's lead-free TCB orders for CoW in November, while ASMPT stated that it is still communicating with TSMC for CoW certification.

Morgan Stanley believes that ASMPT still has the opportunity to complete certification before the end of this year, and low-volume purchases of the company's equipment may be realized in the second half of 2025. Industry surveys show that TSMC's first batch of orders for CoW lead-free TCB was placed with fewer tools through K&S, with an expected volume of less than 10 units in 2025. As the Rubin GPU is expected to enter mass production in 2026, the order volume may increase in 2026.

Another structural opportunity brought by the new GPU is the extension of test time. Morgan Stanley analysts have stated that the current Blackwell (with test times three times longer than Hopper) and MI355 (twice as long) both require longer test times. Due to strong demand, the delivery time for new Advantest testers has been extended from 3 months to 6 months.

In the long run, some AI-specific integrated circuits (ASICs), such as AWS's 3nm AI accelerators, may begin burn-in testing. According to supply chain surveys, the full final testing for Blackwell will continue to be conducted at KYEC, with shipments of the B200/300 (dual-chip version) expected to reach about 5 million units by 2025, based on TSMC's CoWoS-L capacity.