Original source: Wall Street Insights

Semiconductor research organization SemiAnalysis has released two studies, outlining the "ice and fire" sides of Nvidia's prospects that coexist with opportunities and challenges.

SemiAnalysis's latest forecast released on the X platform on June 30 shows that Nvidia's data center computing business revenue in the second half of fiscal year 2027 will be approximately 20% higher than Wall Street's consensus expectations. The core support for this optimistic judgment is that the HBM4 memory supply problem that previously restricted large-scale shipments of the Rubin platform has been resolved. At the same time, front-end wafer production capacity reserves are in place, clearing substantial obstacles for the performance explosion in the second half of the year.

However, in the morning of the same day, SemiAnalysis disclosed another bad news: Nvidia’s original 4-chip Rubin Ultra was canceled about three months after its release at GTC 2026. The size of the new version of “Rubin Ultra” was reduced to half of the original, and the actual performance was also halved.

On the one hand, there is an optimistic upward revision of revenue after the supply bottleneck is lifted, and on the other hand, there is a pessimistic revision of the technical roadmap after the flagship product shrinks - SemiAnalysis. These two completely opposite judgments anchor completely different narrative coordinates for Nvidia from the two dimensions of performance realization and technical moat.

The bottleneck of HBM4 has been solved, and the Rubin platform is expected to increase its volume in the second half of the year

SemiAnalysis made the latest prediction through its Accelerator Model that Nvidia will usher in a large-scale increase in volume in the second half of this year.

The agency expects that Nvidia's data center computing business revenue in the second half of fiscal 2027 will be approximately 20% higher than market consensus expectations, driven by the strong Rubin platform. The HBM4 problem that once affected the progress of Rubin has now been resolved, and the front-end wafer supply has also been reserved in advance, which means that the Rubin platform, which had been delayed, will enter a rapid ramp-up stage.

SemiAnalysis specifically noted that its forecasting logic is significantly different from that of traditional sell-side analysts. Most Wall Street institutions tend to establish relatively conservative profit forecasts to leave room for the company's subsequent "exceeding expectations" performance; while SemiAnalysis's conclusions are based more on front-line research in the industry chain and strive to be closer to real market dynamics.

Its Accelerator Model has built an information cross-validation system covering the entire chain. The data sources include material suppliers, wafer manufacturing, key components, server manufacturers and other supply chain links. At the same time, it combines the actual procurement and deployment status of ultra-large-scale cloud service providers and cutting-edge AI laboratories to conduct multi-dimensional verification of the supply and demand relationship.

It is worth noting that this model not only focuses on NVIDIA, but also covers AI chip manufacturers such as Broadcom, AMD, MediaTek, and Marvell, and combines it with the HBM Model to continue to track the overall evolution of the AI computing power industry chain.

The CUDA moat is eroding, and the shrinkage of Rubin Ultra reflects the rise of self-developed ASICs

However, another previous review by SemiAnalysis about Rubin Ultra triggered widespread discussion in the market.

The agency stated that Nvidia originally planned to use Rubin Ultra designed with 4 computing chips. About three months after the release of GTC this year, it adjusted the original plan. The scale of the new version was significantly reduced compared with the original design. The reason was related to the difficulty of advanced packaging manufacturing.

SemiAnalysis believes that what deserves more attention is not the shrinkage of Rubin Ultra itself, but the changes in the industry competition pattern reflected by this incident. The agency pointed out that in the past year, Nvidia’s biggest competitive pressure is no longer traditional GPU manufacturers such as AMD, but more and more ultra-large-scale cloud manufacturers and AI model companies have begun to use self-developed ASICs to build dedicated chip systems for specific scenarios such as training or reasoning.

For example, Anthropic currently has a multi-platform computing architecture composed of Google TPU, Amazon Trainium and NVIDIA GPU. Among them, a large number of Claude model training runs on the TPU platform, Claude Code inference is increasingly deployed on Trainium, and NVIDIA GPUs are more responsible for general computing tasks such as cutting-edge research. SemiAnalysis points out thatit would have been unimaginable a year ago that TPU and Trainium would be able to grow to the scale they are today, and now the CUDA moat is slowly eroding.