Reports indicate that the original design for NVIDIA's next-generation flagship AI chip, Rubin Ultra, has been significantly scaled down due to manufacturing and packaging challenges.
On June 30th, semiconductor research firm SemiAnalysis posted on platform X, stating that the original 4-chip version of Rubin Ultra was canceled approximately three months after its unveiling at GTC 2026. The revised "Rubin Ultra" is said to be half the size and scale of the original, with a corresponding halving of its expected performance.
Examining the Broader Competitive Context
SemiAnalysis framed this development within a larger competitive landscape, suggesting that NVIDIA's market share is being eroded by chips from Amazon's Trainium, Google's TPU, and AMD. The firm stated that "manufacturing execution issues will only cause more market share to be lost."
The post quickly ignited debate. Supporters view it as a sign of declining execution at NVIDIA, while skeptics argue the information was already public months ago and accuse SemiAnalysis of having a biased stance.
The Original Design's Aggressive Goals
To understand the reported scaling back, one must look at the ambition of the initial plan. According to a prior report by TechPowerUp on March 31st, the standard Rubin GPU was designed with a 2-chip compute die and 8 HBM4 memory modules. The original Rubin Ultra aimed to double that configuration to 4 compute chips and 16 HBM4E modules, all integrated into a single package for a planned 2027 launch.
This was akin to fitting two complete Rubin chips into one package, placing immense demands on packaging technology. TSMC was to utilize its CoWoS-L packaging process for this. However, according to Global Semi Research, the 4-chip configuration (in a 2+2 arrangement) caused warping issues in the package substrate. The substrate bent in multiple directions, preventing the compute dies from making proper contact, which would render the chip inoperable due to signal transmission failure.
Alternative Solutions and Timeline Challenges
Faced with the warping problem in CoWoS-L, TSMC is exploring a new solution called CoPoS (Chip-on-Panel-on-Substrate). The core idea is to use a large square or rectangular panel to replace the roughly 300mm silicon interposer, with early versions measuring around 310x310mm and later versions potentially scaling to 515x510mm or even 750x620mm. Larger panels could accommodate more chips and HBM memory while reducing edge waste.
The issue, however, is timing. As reported by TechPowerUp, TSMC originally planned to establish a CoPoS pilot line by 2026 at the earliest, with mass production targeted for late 2028 to the first half of 2029. This timeline is misaligned with the original 2027 launch target for Rubin Ultra, leaving uncertainty about whether CoPoS could be ready in time.
Competitive Pressures on the CUDA Ecosystem
In its post, SemiAnalysis further argued that competitors are rising faster than anticipated. "Claude Code, the most successful AI agent, has a significant portion of its inference running on Trainium, while Claude is trained on TPU," the firm wrote. "A year ago, it was unimaginable that TPU and Trainium would grow this fast while the CUDA moat is slowly being eroded."
This statement points directly at NVIDIA's core competitive barrier. The CUDA ecosystem has long been considered its most difficult-to-replicate advantage, but SemiAnalysis suggests this moat is weakening. The firm also noted that the reported changes have systemic implications for the HBM memory market and NVIDIA's future rack-scale products.
Questions Over the Report's Novelty
The post on X elicited polarized reactions. Skeptics argued it was merely rehashing old news. Several users commented that this was "old news from three months ago" and that the cancellation of the 4-chip version was "already known" information shared earlier in the year.
Some users criticized SemiAnalysis directly, with one stating: "SemiAnalysis posts another anti-NVIDIA piece, this time a misleading report on 4-chip cancellation (chip count didn't actually change). More interestingly – SA's entire credibility is built on being supplier-agnostic, not a mouthpiece for AMD and ASIC." Another suggested the firm had taken a short position on NVIDIA.
Supporters of the report, however, found it noteworthy, with one commenter stating, "NVIDIA is collapsing under its own arrogance."
NVIDIA has not commented on the reported developments. As one user noted, the company "remains silent as usual on these relentless FUD rumors."
Comments