Mizuho Analyst Vijay Rakesh repeated a purchase and $225 target price Nvidia Corp NVDA.
Mizuho hosted NVDA’s VP of Accelerated Computing, Ian Buck, and IR Simona Jankowski during its 4th Annual Mizuho Storage & AI Virtual Series.
US restricted sale of Nvidia’s high-performance AI chips for servers, A100 and H100, to China and Russia, citing national security concerns.
The US also demanded Advanced Micro Devices, Inc AMD to stop the export its MI250 chips to China.
The share prices of the chip manufacturers and the iShares Semiconductor ETF SOXX came under significant pressure on September 7 after the embargo.
However, Rakesh saw that the restriction could allow access and focused primarily on a workaround for the next few months.
He believes Chinese OEMs can still access US cloud AI services. NVDA could still seek selective licenses more focused on a workaround solution for the A100/H100 for transforming AI training in China hyperscalers, though subject to the same geopolitical risks.
Hopper development is on track and shipping to key customers, with the majority of development taking place in the US, Rakesh noted. The hopper ramp remains unaffected by the ban due to the U.S. exceptions to supply chain freedom, he noted.
Compared to the A100, the H100 with HBM3 from HBM2e storage results in a 1.5x improvement in bandwidth to ~3 Tbps.
He also emphasized that Hopper reduces the number of GPUs by 7 times compared to A100, with superior training performance and HPC computing power, with Hopper H100 already being shipped to some key customers and interest is high.
NVDA also noted that data centers are investigating water cooling with DGX H100 >50% power consumption compared to DGX A100
Overall, increasing complexity is leading to a need for more robust software capabilities to optimize, with NVDA finding that there is a higher mix of software engineers than hardware, Rakesh wrote in the analyst note. He believes this is a key reason for NVDA’s dominance in the AI space.
NVDA accounts for >90%…
