NVIDIA Ships First Vera Rubin AI Racks to Hyperscalers in July

NVIDIA has begun July shipments of its Vera Rubin NVL72 AI platform to Microsoft, Google, AWS, Meta, Oracle and CoreWeave, with mass production ramping through Q3.

Key Takeaways

  • NVIDIA began shipping first Vera Rubin NVL72 racks in July to Microsoft Azure, Google Cloud, AWS, Meta, Oracle and CoreWeave, with mass production ramping through Q3.
  • Each Rubin NVL72 rack contains 72 Rubin GPUs, 36 Vera CPUs and delivers 260 TB/s of scale-up bandwidth.
  • Rubin shipped just three months after unveiling, twice as fast as Blackwell's six-month gap from unveiling to first shipments.
  • TSMC is mass-producing Rubin dies on 3nm with capacity committed through 2027; Foxconn, Quanta and Wistron will scale full rack production in H2 2026.
  • SK hynix has shipped 12-layer HBM4E samples that will supply later Rubin variants, and the early ramp pressures custom-silicon programs at Anthropic and Meta.

NVIDIA Ships First Vera Rubin AI Racks to Hyperscalers in July

NVIDIA has quietly begun shipping the first production units of its Vera Rubin AI platform in July, delivering to a narrow list of hyperscaler customers ahead of a broader Q3 mass ramp. Foxconn, Quanta and Wistron are lead system integrators, and TSMC has moved Rubin dies into 3nm mass production.

First stops: Microsoft, Google, Amazon, Meta, Oracle

NVIDIA said in a note to partners that the first Rubin NVL72 racks are landing in July at Microsoft Azure, Google Cloud, AWS, Meta and Oracle, joined by neocloud CoreWeave. Each rack packs 72 Rubin GPUs, 36 Vera CPUs and 260 TB/s of scale-up bandwidth. The pace is unusually fast for NVIDIA — Blackwell’s first shipments came six months after unveiling; Rubin is doing it in three.

Supply chain aligned

TSMC began Rubin die production earlier this year on 3nm and has committed capacity through 2027 as part of its physical-AI-driven fab collaboration with NVIDIA. Contract manufacturing partners including Foxconn, Quanta and Wistron will roll out full-scale rack production in H2 2026, and memory partner SK hynix has already shipped 12-layer HBM4E samples that will feed later Rubin variants.

NVIDIA Vera Rubin AI platform

A capex arms race, upgraded

Rubin’s July slot lands in a wider capex arms race that already includes SpaceX committing $6.3 billion to Reflection AI’s Colossus 2, NVIDIA rolling out a revenue-share compute model and AirTrunk’s $21 billion Maharashtra campus. Rubin’s early shipment pushes hyperscaler generation targets to end-of-year and reshapes competitive pressure on custom-silicon programs at Anthropic and Meta.

Reporting based on coverage from Wccftech, TradingKey and VideoCardz.

Category: Partnerships

Tags: AI Foundation Models Chrome Gigafactory

Related Articles

Frequently Asked Questions

Which companies are receiving the first NVIDIA Vera Rubin racks?

The first Rubin NVL72 racks are shipping in July to hyperscalers Microsoft Azure, Google Cloud, AWS, Meta and Oracle, plus neocloud provider CoreWeave.

What are the specs of a Vera Rubin NVL72 rack?

Each rack packs 72 Rubin GPUs, 36 Vera CPUs and 260 TB/s of scale-up bandwidth.

Who is manufacturing the Vera Rubin platform?

TSMC produces Rubin dies on its 3nm process with capacity committed through 2027, while Foxconn, Quanta and Wistron serve as lead system integrators, ramping full-scale rack production in H2 2026. SK hynix supplies HBM4E memory for later variants.

How does Rubin's launch timeline compare to Blackwell's?

Rubin shipped three months after its unveiling, an unusually fast pace for NVIDIA — Blackwell's first shipments came six months after unveiling.