SemiAnalysis: The Fire and Ice of NVIDIA's Rubin Platform
- Core Viewpoint: SemiAnalysis has released two contrasting analyses, indicating that NVIDIA's performance in the second half of fiscal 2027 will exceed expectations due to the resolution of HBM4 supply bottlenecks. However, the flagship product, Rubin Ultra, has been significantly scaled back for technical reasons, indirectly reflecting the erosion of its CUDA ecosystem by custom ASICs.
- Key Elements:
- SemiAnalysis predicts that NVIDIA's data center revenue in the second half of fiscal 2027 will be approximately 20% higher than the Wall Street consensus. The core drivers are the resolution of HBM4 supply issues for the Rubin platform and the reserves of front-end wafer production capacity.
- SemiAnalysis's predictive model is based on primary research across the industry chain, covering materials, wafers, components, servers, and procurement data from cloud service providers, distinguishing it from the more conservative estimates of traditional sell-side firms.
- NVIDIA's original plan for a 4-chip Rubin Ultra was canceled approximately three months after its announcement at GTC 2026. The new version has half the scale and half the performance, with the cause linked to the difficulty of advanced packaging manufacturing.
- Shift in NVIDIA's competitive landscape: Hyperscalers and AI companies (like Anthropic) are adopting custom ASICs (e.g., Google TPU, Amazon Trainium) for both training and inference, eroding the CUDA moat.
- Anthropic has adopted a multi-platform architecture comprising TPU, Trainium, and NVIDIA GPUs. Claude model training and inference are gradually being migrated from GPUs to TPUs and Trainium.
Original source: Wall Street News
Semiconductor research firm SemiAnalysis has released two analyses, outlining a mixed picture of opportunities and challenges for NVIDIA's prospects, akin to a "fire and ice" duality.
SemiAnalysis's latest forecast, published on platform X on June 30, indicates that NVIDIA's data center computing revenue for the second half of fiscal year 2027 will be approximately 20% higher than Wall Street's consensus estimates. The core support for this optimistic outlook lies in the resolution of the HBM4 memory supply issues that previously constrained the mass shipment of the Rubin platform, coupled with the securing of front-end wafer capacity, clearing a significant hurdle for a strong performance in the second half of the year.
However, on the morning of the same day, SemiAnalysis disclosed another piece of bearish news: NVIDIA's original 4-chip Rubin Ultra was canceled approximately three months after its unveiling at GTC 2026. The new "Rubin Ultra" has been scaled down to half its original size, resulting in a halving of its actual performance.
On one hand, optimistic upward revisions in revenue following the resolution of supply bottlenecks; on the other, pessimistic corrections regarding technological direction after a flagship product downsizing. These two diametrically opposed analyses from SemiAnalysis anchor entirely different narratives for NVIDIA from the perspectives of performance delivery and technological moat.
HBM4 Bottleneck Resolved; Rubin Platform Poised for Volume Growth in H2
SemiAnalysis, using its Accelerator Model, has forecast a significant ramp-up in shipments for NVIDIA in the second half of this year.
The firm estimates that driven strongly by the Rubin platform, NVIDIA's data center computing revenue for the second half of fiscal year 2027 will be approximately 20% higher than market consensus expectations. The HBM4 issues that previously impacted Rubin's progress have now been resolved, and front-end wafer supply has been secured in advance, suggesting the delayed Rubin platform will enter a rapid ramping phase.
SemiAnalysis specifically notes that its forecasting logic differs significantly from traditional sell-side analysts. Most Wall Street institutions tend to build relatively conservative earnings forecasts, reserving room for future "surprises." In contrast, SemiAnalysis's conclusions are more grounded in front-line industry research, striving to be closer to actual market dynamics.
Its Accelerator Model constructs a cross-validation system covering the entire supply chain. Data sources include material suppliers, wafer fabrication, key components, server OEMs, and other links in the supply chain. It also integrates actual procurement and deployment data from hyperscale cloud providers and cutting-edge AI labs to conduct multi-dimensional checks on supply-demand dynamics.
Notably, this model not only focuses on NVIDIA but also covers other AI chip firms like Broadcom, AMD, MediaTek, and Marvell, continuously tracking the overall evolution of the AI computing power industry chain in conjunction with the HBM Model.
CUDA Moat Under Erosion; Rubin Ultra Downsizing Reflects Rise of Custom ASICs
However, another earlier commentary from SemiAnalysis regarding the Rubin Ultra sparked widespread market discussion.
The firm stated that approximately three months after Rubin Ultra's unveiling at this year's GTC, NVIDIA adjusted the original plan for the chip, which was initially designed with 4 compute dies. The scale of the new version is significantly reduced compared to the original design, attributed to difficulties in advanced packaging manufacturing.
SemiAnalysis believes that the downsizing of Rubin Ultra itself is less noteworthy than the shift in the competitive landscape it reflects. The firm points out that over the past year, NVIDIA's biggest competitive pressure has no longer just come from traditional GPU players like AMD. A growing number of hyperscale cloud providers and AI model companies have started adopting custom ASICs, building specialized chip systems for specific scenarios like training or inference.
For example, Anthropic currently operates a multi-platform compute architecture comprising Google's TPUs, Amazon's Trainium, and NVIDIA's GPUs. A significant portion of Claude model training runs on TPU platforms, while Claude Code inference is increasingly deployed on Trainium. NVIDIA GPUs are more tasked with general-purpose computing for cutting-edge research. SemiAnalysis notes that a year ago, the growth of TPU and Trainium to their current scale would have been hard to imagine, but now the CUDA moat is being slowly eroded.
