NVIDIA's Blackwell-related woes are well-known by now, though the GPU giant has done much in recent weeks to improve the production cadence of its latest cutting-edge offerings. One such change involves the Blackwell Ultra GPU's compute board, which is now drawing appreciative glances from Wall Street.
To wit, KeyBanc has now noted with manifest positivity NVIDIA's recent switch to the Bianca compute board, which pairs one CPU with 2 GPUs, from the earlier Cordelia compute board, which adhered 2 CPUs to 4 GPUs, for the GB300 Blackwell Ultra GPUs.
The investment bank explains the rationale behind this change in the following words:
"We believe this change was due to a signal loss performance issue related to Cordelia, as it utilizes a SXM (Server PCI Express) socket interface that was intended to support the enablement of a broader supply chain and to provide better serviceability, which is in contrast to Bianca, which has a more fixed board structure with limited flexibility."
In essence, despite offering more convenient maintenance via the Server PCI Express (SXM) socket interface, Cordelia's structural flexibility occasionally led to a loss of signal issue for the socket interface. This then compelled NVIDIA to switch back to the sturdier Bianca compute board.
According to KeyBanc, this change will allow NVIDIA to "maintain its 4Q25 launch schedule for GB300 and to provide a more seamless transition from Blackwell to Blackwell Ultra."
Also, NVIDIA intends to ship around 30,000 GB NVL rack units this year, with only 30 percent of these shipments expected to materialize in H1 2025, while the remaining 70 percent would presumably ship in the second half of the year. Consequently, NVIDIA's reversion to Bianca would "minimize changes as we transition to GB300, and effectively provides a drop-in replacement for NVDA's NVL72 rack structure from GB200 to GB30," as per KeyBanc's assessment.
Of course, today's commentary from KeyBanc is a breath of fresh air for NVIDIA bulls, many of whom have been hammered over the past few days by back-to-back adverse developments.
First, NVIDIA announced that it expects to incur charges of up to $5.5 billion during its fiscal Q1 2026 (which concludes on the 27th of April) after the Trump administration introduced export licensing requirements on the China-specific H20 GPU.
Second, Huawei recently unveiled its latest AI chip, the Ascend 910C, which offers similar performance to NVIDIA's H100 GPUs by combining the two smaller 910B GPUs.
