Nvidia expects Rubin platform shipping in second half of fiscal year 2027

Nvidia’s next-generation AI accelerator platform, Vera Rubin, is on track to begin shipping in the second half of fiscal year 2027. For those keeping score on Nvidia’s fiscal calendar, that translates to the second half of calendar year 2026.

The timeline positions Vera Rubin as the natural successor to the Blackwell architecture, which has dominated conversations about AI infrastructure over the past year. But where Blackwell was already a beast, Rubin is shaping up to be something fundamentally different in scope.

Not just a GPU, an entire AI factory

Here’s the thing about Vera Rubin: calling it a “chip” undersells what Nvidia is actually building. The company describes it as a comprehensive system designed for large-scale AI factories. That includes NVL72 GPU racks, CPUs, inference accelerators, and networking components, all bundled into what amounts to a turnkey data center solution.

In English: Nvidia isn’t just selling you the engine anymore. It’s selling you the entire car, the road, and the gas station.

The NVL72 configuration alone is staggering in its resource demands. Each NVL72 server may require approximately 1,152TB of NAND, owing to its integrated storage and memory configurations. To put that in perspective, that’s roughly the equivalent of over a thousand consumer-grade 1TB SSDs packed into a single server rack. The sheer volume of flash memory required for these systems could have meaningful ripple effects across the NAND supply chain.

This architectural decision signals that Nvidia is betting heavily on rack-scale computing as the future of AI infrastructure. Rather than selling individual accelerators that customers then stitch together into clusters, the company is moving toward pre-integrated systems where the GPU, CPU, networking, and storage all ship as a unified product.

Who’s buying and when

Production of Vera Rubin has reportedly already begun, with some initial customer samples being distributed ahead of the broader commercial rollout. That’s consistent with Nvidia’s typical approach of seeding key partners early to ensure software optimization is ready by the time volume shipments start.

One confirmed early adopter is Nebius, which plans to deploy Nvidia Vera Rubin NVL72 systems across the US and Europe starting in the second half of 2026. Nebius, for context, is an AI infrastructure company that spun out of Yandex’s cloud business and has been rapidly expanding its data center footprint.

The Nebius deployment is notable because it suggests Vera Rubin isn’t just aimed at the usual hyperscaler suspects like Microsoft, Google, and Amazon. Nvidia appears to be casting a wider net, targeting the growing tier of AI-focused infrastructure companies that are building out their own compute capacity to serve enterprise customers.

This matters because the AI infrastructure market is no longer just about the biggest players hoarding GPUs. A secondary ecosystem of compute providers has emerged, and they need next-generation hardware just as badly as the hyperscalers do. Vera Rubin’s rack-scale design could be particularly appealing to these companies, since it reduces the engineering burden of assembling and optimizing complex GPU clusters from scratch.

The competitive landscape and what investors should watch

Nvidia’s annual cadence of new AI architectures has become one of the most important product cycles in the tech industry. The company has gone from Hopper to Blackwell, and now to Vera Rubin, with each generation promising significant performance improvements for AI training and inference workloads.

The shift toward rack-scale infrastructure is strategically significant. By selling complete systems rather than individual chips, Nvidia captures more value per customer and makes it harder for competitors to pry away individual components of the stack. If a data center operator buys an NVL72 rack, they’re getting Nvidia silicon, Nvidia networking, and Nvidia software, all tightly integrated. Switching costs go up dramatically.

This is the moat-building strategy that AMD, Intel, and a growing list of custom silicon startups will need to contend with. It’s one thing to compete on raw GPU performance. It’s another thing entirely to compete against a fully integrated system where every component has been co-designed to work together.

For investors, the key metrics to watch will be how quickly Vera Rubin ramps to volume production and whether the massive NAND requirements per rack create any supply constraints. A single NVL72 server consuming over 1,000TB of NAND is the kind of demand signal that could tighten flash memory markets, potentially benefiting NAND producers like Samsung, SK Hynix, and Micron.

There’s also the question of pricing. Rack-scale systems will carry price tags that dwarf individual GPU purchases, which could concentrate Nvidia’s revenue among fewer but larger customers. That’s great for average deal size but introduces concentration risk if any major buyer pulls back on spending.

The broader signal from Vera Rubin is that Nvidia sees AI infrastructure spending accelerating, not plateauing. You don’t design a system that requires 1,152TB of NAND per server if you think demand is about to cool off. Whether that confidence proves justified will likely be one of the defining questions for the semiconductor industry through 2026 and beyond.

Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.