NVIDIA is already lining up its next big leap in AI infrastructure. While Blackwell Ultra GB300 systems are ramping into mass production, the company and its manufacturing partners are preparing the Vera Rubin generation of AI servers, positioned to deliver a major jump in data center performance and efficiency.
At the heart of this push is the Vera Rubin NVL144 built on the MGX platform, which is designed to bring tremendous compute density and a revamped rack design to hyperscale deployments. Industry reports indicate that Foxconn has begun development on these next‑gen NVL144 MGX servers, with mass production targeted for the second half of 2026. That timeline underscores NVIDIA’s aggressive cadence: Vera Rubin is expected to follow just six to eight months after the GB300 ramp, signaling faster upgrade cycles for AI infrastructure.
For the near term, GB200 Blackwell systems remain the primary revenue engine for the ecosystem. Suppliers have already moved into volume production, and Blackwell Ultra is expected to account for a substantial share of AI server shipments through the end of this year and into the first half of 2026. Vera Rubin then steps in as the next wave, bringing a reworked stack and significantly higher AI compute to keep pace with expanding model sizes and training demands.
Foxconn is emerging as a central player in this transition. Estimates suggest it could command around 60% share among NVIDIA’s next‑gen AI server manufacturing partners. The company is also ramping investment in US‑focused facilities, a move that aligns with NVIDIA’s push to expand American manufacturing for AI hardware. If those plans hold, expect tangible shifts in where high‑end AI servers are built and assembled over the next few years.
Momentum on the customer side is already building. Vera Rubin has reportedly been selected for large-scale, multi‑gigawatt deployments, including deals with OpenAI, indicating that early access and pilot deployments could reach mainstream visibility as soon as next year. That sets the stage for a broader rollout culminating with mass production of specific NVL144 MGX configurations in late 2026.
The takeaway for data center operators is clear: the upgrade path from Blackwell Ultra to Vera Rubin is coming quickly. With a redesigned rack architecture, denser compute per rack unit, and a focus on end-to-end platform efficiency, Vera Rubin aims to compress training times and improve inference throughput while optimizing power and floor space. For organizations planning capacity two to three years out, aligning roadmaps with NVIDIA’s GB300-to‑Vera Rubin transition could prove decisive in staying ahead of AI workload growth.
