Nvidia scraps original four-chip Rubin Ultra design

SemiAnalysis reported June 30 that Nvidia's original four-die Rubin Ultra was canceled, replaced by a half-size, half-performance design.moomoo+1
Substrate warping in TSMC's CoWoS-L packaging made the monolithic four-chip package unviable, according to the firm.digg
SemiAnalysis warned that manufacturing setbacks will accelerate market share losses to Amazon's Trainium, Google's TPUs, and AMD chips.moomoo+1

NVIDIA has canceled the original four-chip design of its Rubin Ultra GPU just three months after unveiling it at GTC 2026, according to semiconductor research firm SemiAnalysis, which posted the findings on June 30. The revised chip reverts to a two-chip configuration with roughly half the performance of what was originally promised, a setback that arrives as the company faces mounting competition from hyperscalers building their own AI silicon.

Manufacturing Challenges Force a Redesign

SemiAnalysis reported that substrate warpage in TSMC's Taiwan Semiconductor Manufacturing Company Limited CoWoS-L advanced packaging process made the original four-die design unviable. The initial Rubin Ultra had planned a single massive package with approximately 1 TB of HBM4E memory. The scaled-back version uses dual-die GPUs with 384 to 768 GB of HBM4E each, achieving comparable total performance only through a board-level two-plus-two configuration rather than the monolithic package NVIDIA had envisioned.moomoo+2

"Execution issues at the manufacturing level will only accelerate further market share loss," SemiAnalysis stated. The firm had previously clashed with NVIDIA over product timelines, including a dispute in early June about co-packaged optics roadmaps that prompted an official NVIDIA rebuttal.news.futunn+1

CUDA Moat Under Pressure

The cancellation lands amid broader questions about NVIDIA's dominance in AI accelerators. While the company still commands an estimated 80 to 85 percent of the market by revenue, competitors are gaining ground. Anthropic announced in April a $100 billion, ten-year commitment to Amazon Web Services Amazon.com, Inc. technologies, including current and future Trainium chips for training and deploying its Claude models. A majority of overall Trainium allocation is currently dedicated to Claude workloads, according to Futurum Group's analysis of the deal.anthropic+3

The software side of NVIDIA's competitive advantage is also facing challenges. Hardware-agnostic frameworks such as OpenAI's Triton compiler and Google's Alphabet Inc. JAX framework allow developers to deploy models across AMD Advanced Micro Devices, Inc. , Intel , and custom ASICs without rewriting CUDA code. Several online commentators disputed SemiAnalysis's framing, accusing the firm of bias and noting that some of the competitive dynamics described were not new.builtin+2

What Comes Next

NVIDIA's data center business remains formidable — fiscal year 2026 revenue reached $215.9 billion, up 65 percent year over year. But the Rubin Ultra setback raises questions about whether the company can sustain the pace of its generational leaps as packaging technology approaches physical limits. TSMC is already exploring panel-level packaging for future NVIDIA designs, moving from CoWoS to CoPoS to handle ever-larger chip assemblies.nvidianews.nvidia+1

SemiAnalysis framed the cancellation not as an isolated manufacturing hiccup but as a symptom of a shifting landscape where hyperscalers are less willing to wait for NVIDIA's next-generation silicon when they can build their own.moomoo+1

Sources (39)