NVIDIA's Rubin Ultra 4-Chip Solution Canceled, Performance Scale Cut in Half
N.R. Finch
Nvidia's original 4-die Rubin Ultra package has been cancelled roughly three months after its unveiling, with the replacement halved in scale and performance — the root cause is a manufacturing wall at TSMC, and competitors are closing in.
Why was the original design killed?
Chip-research firm SemiAnalysis reported on June 30 that the original Rubin Ultra planned to pack 4 compute dies and 16 HBM4E memory modules into a single package — essentially two full Rubin chips fused together.
TSMC hit a wall: using its CoWoS-L process — a packaging method that bonds multiple dies onto one substrate — the substrate warped in the 4-die (2+2) layout, bending in multiple directions so the dies could not make full contact. Signals failed.
This means → the design wasn't the problem. Manufacturing was. The bigger you build the package, the harder it is to control physical stress.
Can the backup plan close the gap?
TSMC is developing an alternative called CoPoS — Chip-on-Panel-Substrate — replacing the traditional ~300 mm round silicon interposer with a large square panel, initially around 310 × 310 mm and scalable to 750 × 620 mm.
The timing does not line up: the CoPoS pilot line starts in 2026 at the earliest, with volume production targeted for late 2028 to first-half 2029. The original Rubin Ultra was slated for 2027.
In plain terms = the fix exists on paper, but the factory isn't ready — and the deadline has already passed.
Why does SemiAnalysis say the CUDA moat is eroding?
SemiAnalysis framed the cancellation inside a broader competitive picture: Nvidia's share is being chipped away by Amazon's Trainium, Google's TPU, and AMD.
The firm cited a specific case: Claude Code, the most successful AI agent, runs a significant portion of its inference on Trainium, while Claude's training runs on TPUs.
This means → something hard to imagine a year ago is now happening — top-tier AI workloads are bypassing Nvidia's CUDA ecosystem and running on rival silicon.
What is the market actually arguing about?
Skeptics say the 4-die cancellation was public knowledge three months ago and accuse SemiAnalysis of bias. One commenter called the firm "a mouthpiece for AMD and ASICs"; another said it was "obviously short Nvidia."
Supporters see this as a signal that Nvidia's execution is slipping — manufacturing troubles compounding competitive pressure, making the strain structural.
This reflects a deeper split: the debate is no longer about whether Nvidia is overvalued, but about whether the moat still holds. Nvidia has not commented on the report.
What to watch next?
One variable matters above all: whether TSMC can bring CoPoS to volume production in time for the 2027 window.
If it cannot, the new Rubin Ultra remains a compromise — half the performance of the original blueprint.
In plain terms = Nvidia's design capability is not in question. Only what can actually be built counts — and that timeline is in TSMC's hands.
Content is for reference only, not financial advice.