A semiconductor research firm has published two separate analyses, painting a contrasting picture of both opportunity and challenge for NVIDIA's future.
The firm's latest forecast, published on June 30, indicates that NVIDIA's data center compute revenue for the second half of fiscal 2027 will be approximately 20% higher than the Wall Street consensus. This optimistic outlook hinges on the resolution of previously constraining HBM4 memory supply issues for the Rubin platform, coupled with secured front-end wafer capacity, which are seen as clearing the path for a significant performance surge later this year.
However, on the same morning, the firm disclosed contrasting negative news: the original 4-chip NVIDIA Rubin Ultra, unveiled at GTC 2026, was reportedly canceled roughly three months after its announcement. The revised "Rubin Ultra" is said to be scaled down to half its original size, with a corresponding halving of its actual performance.
On one side is an optimistic revenue revision following the removal of supply bottlenecks; on the other is a pessimistic reassessment of the technology roadmap after the flagship product's downgrade. These opposing assessments from the research firm anchor NVIDIA's narrative at vastly different coordinates, focusing on earnings potential and technological moat, respectively.
HBM4 Hurdle Cleared, Rubin Ramp-Up Anticipated
Using its proprietary Accelerator Model, the firm predicts that NVIDIA is poised for a major production ramp-up in the second half of this year.
It is projected that, driven strongly by the Rubin platform, NVIDIA's data center compute revenue for the latter half of fiscal 2027 will surpass the market consensus by about 20%. The previously hindering HBM4 issue affecting Rubin's progress has now been resolved, and front-end wafer supply has been secured in advance. This suggests the once-delayed Rubin platform will enter a phase of rapid scaling.
The firm notes that its forecasting logic differs significantly from that of traditional sell-side analysts. Most Wall Street firms tend to build relatively conservative earnings estimates to leave room for subsequent "outperformance," while the firm's conclusions are more grounded in direct supply chain research, aiming to reflect actual market dynamics more closely.
Its Accelerator Model constructs a comprehensive information cross-verification system covering the entire supply chain. Data sources include material suppliers, wafer manufacturing, key components, and server OEMs, combined with actual procurement and deployment insights from hyperscale cloud providers and cutting-edge AI labs to validate supply and demand from multiple angles.
It is worth noting that this model does not focus solely on NVIDIA; it similarly covers other AI chip players like Broadcom, AMD, MediaTek, and Marvell, and tracks the overall evolution of the AI compute supply chain in conjunction with its HBM Model.
CUDA Moat Under Pressure, Rubin Ultra Downgrade Reflects Rise of Custom Chips
However, another earlier commentary from the firm regarding Rubin Ultra has sparked widespread market discussion.
The firm stated that NVIDIA's originally planned Rubin Ultra, designed with four compute chips, saw its original blueprint altered roughly three months after its GTC unveiling this year. The new version is significantly scaled down from the original design, with reasons linked to the manufacturing complexity of advanced packaging.
The firm believes the more critical issue is not the downgrade of Rubin Ultra itself, but the shift in the competitive landscape it signals. It points out that over the past year, NVIDIA's greatest competitive pressure no longer comes solely from traditional GPU rivals like AMD, but increasingly from hyperscale cloud providers and AI model companies developing their own Application-Specific Integrated Circuits (ASICs) tailored for specific tasks like training or inference.
For instance, Anthropic has reportedly built a multi-platform compute architecture comprising Google's TPUs, Amazon's Trainium chips, and NVIDIA GPUs. Within this, a significant portion of Claude model training runs on TPU platforms, Claude Code inference is increasingly deployed on Trainium, while NVIDIA GPUs handle more general-purpose computing tasks like frontier research. The firm notes that a year ago, the current scale of TPU and Trainium adoption would have been difficult to imagine, suggesting that the CUDA ecosystem's competitive moat is being gradually eroded.
Comments