Nvidia Vera Rubin: единая система CPU и GPU, чтобы владеть каждым чипом в AI-дата-центрах
Nvidia превращает Vera Rubin в оружие вертикальной интеграции: фирменный Arm-процессор Vera и GPU Rubin с памятью HBM4 работают как единая система, а стойка NVL144 выдаёт до 3,6 эксафлопс в FP4-инференсе. Как пишет Wired, цель компании — владеть каждым чипом внутри AI-дата-центра. Поставки стартуют во второй половине 2026 года, Rubin Ultra — в 2027-м.
AI-processed from Wired; edited by Hamidun News
Nvidia is making the Vera Rubin platform the centerpiece of its strategy: the system combines the company's own Vera processors (CPU) and Rubin accelerators (GPU) into a single product, with shipments scheduled for the second half of 2026. As Wired writes, the company's ambition is to produce every chip inside an AI data center — from compute superchips to the networking fabric.
What is Vera Rubin
Vera Rubin is Nvidia's next-generation AI platform, which CEO Jensen Huang announced at the GTC conference on March 18, 2025 as the successor to Grace Blackwell. The platform is named after astronomer Vera Rubin and, for the first time, makes the "own CPU plus own GPU" pairing the basic unit of the server rack rather than an option for select configurations.
- Announcement — March 18, 2025, Jensen Huang's keynote at GTC
- Configuration: a Vera CPU with 88 custom Arm cores and a Rubin GPU with HBM4 memory
- The Vera Rubin NVL144 rack — up to 3.6 exaflops in FP4 inference
- Start of shipments — second half of 2026
- Next step — Rubin Ultra NVL576 with 15 exaflops in 2027
According to Nvidia, the NVL144 rack is roughly 3.3 times more performant in inference workloads than the current Grace Blackwell NVL72 generation — and it was designed as a single system, not as a set of compatible components.
Why Nvidia wants to own every chip
Control over every layer of the data center protects Nvidia's margins and ties customers to its ecosystem. The data center segment brought the company $115.2 billion in revenue in fiscal year 2025, according to Nvidia's reporting, and the portfolio now extends far beyond GPUs: the Arm-based Vera processor, the NVLink interconnect, the Spectrum-X networking platforms and the BlueField DPU cover virtually the entire list of silicon in a server rack. In May 2025 the company also unveiled NVLink Fusion — a technology that lets third-party CPUs and accelerators into its interconnect, but on Nvidia's own terms.
Wired sees a systemic shift here: the more tightly the CPU and GPU are integrated into a single superchip, the less room is left for Intel and AMD processors in AI servers.
"The
Vera Rubin platform combines the CPU and GPU in a single system and reflects the company's growing ambition to provide every layer of AI infrastructure," the Wired article says.
What will change for the market
Buying Nvidia accelerators will increasingly pull the rest of the stack along with it — from the central processor to the network cards. The Vera processor with its 88 cores is, according to Nvidia, roughly twice as fast as the previous Grace, so customers are left with fewer arguments for putting someone else's CPU next to it. At the GTC conference in Washington in October 2025, Huang was already showing an assembled Vera Rubin superchip and saying the platform was on schedule.
Hyperscalers are responding with silicon of their own: Google is developing TPUs, Amazon — Trainium, while Microsoft and Meta are designing their own accelerators. Nvidia's bet on fully integrated racks is a way to keep its largest customers inside its ecosystem before these alternatives mature.
What this means
Nvidia is ceasing to be a supplier of standalone accelerators and is turning into a vertically integrated manufacturer of "AI factories" in which it owns every chip. Customers get maximum performance out of the box — and a growing dependence on a single vendor, with all the pricing consequences that entails.
Frequently asked questions
When will Vera Rubin be released?
Nvidia promises to start shipping the Vera Rubin platform in the second half of 2026; the beefed-up Rubin Ultra NVL576 version is slated for 2027.
How does Vera Rubin differ from Blackwell?
Vera Rubin combines the company's own CPU and GPU in a single system and uses HBM4 memory, whereas the Grace Blackwell generation was assembled from more autonomous components. In inference, the NVL144 rack is, according to Nvidia, roughly 3.3 times faster than Grace Blackwell NVL72.
- Meta has been designated an extremist organization and is banned in Russia.
Want to stop reading about AI and start using it?
AI News is a curated feed of AI/tech news. Hamidun Academy teaches you to use AI systematically in your work.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.