NVIDIA's AI AI server evolution: GB300 production ramps up as Vera Rubin looms on the horizon

NVIDIA's GB300 "Blackwell Ultra" Enters Production

NVIDIA's next-generation GB300 "Blackwell Ultra" AI servers have entered production, with large-scale shipments expected to begin in September 2025 1

1

2

2

. This rollout is anticipated to proceed more smoothly than previous generations, thanks to strategic design reuse and improved coordination across the supply chain 1

1

.

Design Reuse and Modular Approach

A key factor enabling the faster transition is NVIDIA's decision to retain the motherboard design used in the current GB200 platform, known as the Bianca board 1

1

2

2

. This move has significantly shortened the learning curve for suppliers, many of which were struggling to keep up with NVIDIA's rapid product update cycle 2

2

.

For the GB300, NVIDIA is shifting to a more modular approach:

• The B300 GPU is provided on an SXM Puck module
• The Grace CPU comes in a separate BGA package
• The hardware management controller (HMC) is supplied by Axiado
• Customers source remaining motherboard components themselves
• CPU memory uses standard SOCAMM memory modules from various vendors 1

  1

This reuse and modular design eliminate the need for a complete redesign, streamlining production processes and reducing risk 1

1

.

Production and Shipment Timeline

ODMs report no significant hurdles in the validation and early production phases of GB300 1

1

. Feedback from partners indicates that component qualification is progressing as planned, and NVIDIA is on track to increase output steadily throughout the third quarter of 2025 1

1

2

2

. By the fourth quarter of 2025, shipment volumes are expected to ramp up significantly 1

1

3

3

.

Challenges with GB200 Liquid Cooling

Despite high-volume shipments of GB200 to data centers, persistent problems with its liquid cooling systems have been reported 1

1

. The primary failures occur with quick-connect fittings, which have shown a tendency to leak despite undergoing factory stress tests 1

1

. Data center operators have responded by adopting measures like localized shutdowns and extensive leak testing, prioritizing deployment speed and performance over hardware reliability 1

1

.

Looking Ahead: Vera Rubin Platform

Even as GB300 production ramps up, NVIDIA is already preparing for its next-generation AI servers, codenamed Vera Rubin 1

1

4

4

. This platform will roll out in two phases:

1. Replace Grace CPUs with Vera CPUs and Blackwell GPUs with Rubin GPUs, retaining the current Oberon rack (NVL144)
2. Introduce an all-new Kyber rack with Vera CPUs and Rubin Ultra GPUs featuring four compute chiplets 1

   1

The Vera Rubin platform is expected to be available in the market from 2026 to 2027 4

4

. It will utilize next-generation HBM4 chips, adopt TSMC's 3nm (N3P) process, and implement CoWoS-L packaging 4

4

. Notably, Rubin will be NVIDIA's first implementation of a chiplet design, featuring a 4x reticle design compared to Blackwell's 3.3x 4

4

.

NVIDIA's Relentless Innovation Pace

NVIDIA's rapid product cycle, releasing new AI server architectures every six to eight months, has set a blistering pace that competitors struggle to match 2

2

4

4

. This relentless standard, driven by CEO Jensen Huang, has positioned NVIDIA at the forefront of AI innovation 4

4

. However, it also raises questions about how the supply chain can adapt to such frequent architectural changes 4

4

.

As NVIDIA continues to push the boundaries of AI server technology, the industry watches closely to see how this rapid innovation cycle will shape the future of artificial intelligence and high-performance computing.