Vera Rubin Is More Than a New GPU: Inside NVIDIA’s Next AI Factory
TLDR: Vera Rubin is not simply NVIDIA's next GPU generation. It is a co-designed AI-factory platform spanning the Vera CPU, Rubin GPU, NVLink 6, ConnectX-9, Spectrum-6, BlueField-4, optics, power and cooling. NVIDIA is even changing how the infrastructure is deployed: with DSX POD, a pre-integrated AI factory can be deployed onsite in as little as 14 days once site utilities are ready.
Blackwell made the rack the computer. Rubin pushes the idea further.
When people hear Vera Rubin, the first reaction is usually:
NVIDIA has a new GPU generation.
That is true, but it misses the bigger story.
Vera Rubin is not simply a faster GPU replacing Blackwell.
NVIDIA is redesigning the CPU, GPU, scale-up fabric, scale-out NIC, Ethernet fabric, DPU, optics, power, cooling and even the way an AI factory gets deployed.
The easiest way to understand Vera Rubin is to stop thinking about one GPU.
Think about the entire AI factory as one computer.
Blackwell → Vera Rubin
| Blackwell generation | Vera Rubin generation |
|---|---|
| Grace CPU | Vera CPU |
| Blackwell GPU | Rubin GPU |
| NVLink 5 | NVLink 6 |
| ConnectX-7 / ConnectX-8 era | ConnectX-9 |
| BlueField-3 | BlueField-4 |
| Spectrum-X | Spectrum-6 + Spectrum-X |
| HBM3E | HBM4 |
The important point is not any individual component.
It is that NVIDIA designed these pieces to operate together.
That is why Vera Rubin should be viewed as an AI system architecture, not simply a GPU architecture.
What is Vera Rubin NVL72?
At the center of the platform is the Vera Rubin NVL72 rack.
One rack contains:
- 72 Rubin GPUs
- 36 Vera CPUs
- 18 compute trays
- 9 NVLink 6 switch trays
- networking, power, management and liquid-cooling infrastructure
Each compute tray contains:
2 Vera CPUs + 4 Rubin GPUs
and the 18 trays together create the 72-GPU system.
A simplified view:
VERA RUBIN NVL72
36 Vera CPUs
│
72 Rubin GPUs
│
NVLink 6
│
9 NVLink Switch Trays
│
One 72-GPU NVL Domain
The rack is no longer simply a collection of independent servers.
Through NVLink and NVSwitch, the GPUs operate as a tightly connected accelerator system.
NVIDIA lists the NVL72 system with 20.7 TB of HBM4 and approximately 216 TB/s of aggregate NVLink bandwidth.
But eventually the workload has to leave the rack.
That is where the networking story gets especially interesting.
NVLink 6 and ConnectX-9 solve different problems
A common question is:
If Vera Rubin has NVLink, why does it also need ConnectX?
Because they solve two different networking problems.
NVLink 6 = scale-up
NVLink connects GPUs inside the tightly coupled NVL72 domain.
GPU ←→ GPU ←→ GPU ←→ GPU
NVLink 6
Its job is extremely fast GPU-to-GPU communication inside the rack-scale computer.
ConnectX-9 = scale-out
ConnectX-9 connects the rack to GPUs in other racks.
Rubin GPU
│
ConnectX-9
│
RDMA / RoCE
│
Spectrum-X
│
Remote NVL72
NVIDIA says Vera Rubin can provide up to 1.6 Tb/s of scale-out networking per Rubin GPU.
Once an NCCL collective crosses racks, traffic moves through the data-center fabric.
The simplest way to remember the split is:
NVLink makes GPUs inside the rack behave like one tightly coupled accelerator.
ConnectX-9 connects that accelerator to the rest of the AI factory.
Then comes Spectrum-6
ConnectX-9 is the endpoint.
The scale-out fabric still needs switches.
That is where Spectrum-6 comes in.
Spectrum-6 provides 102.4 Tb/s of Ethernet switching capacity with 200G SerDes.
The path becomes:
Rubin GPU
│
NCCL
│
ConnectX-9
│
RDMA / RoCE
│
Spectrum-6
│
Leaf / Spine Fabric
│
ConnectX-9
│
Remote Rubin GPU
For network engineers, this is where Rubin becomes especially interesting.
Imagine thousands of GPUs entering an AllReduce or All-to-All phase almost simultaneously.
If too many elephant flows converge on the same path:
Queue builds
↓
ECN marks increase
↓
Congestion control reacts
↓
RDMA throughput changes
↓
One NCCL rank slows
↓
Other GPUs wait
Nothing necessarily failed.
BGP can remain established.
Interfaces can stay UP.
CRC counters can remain zero.
Yet expensive GPUs can still be waiting on communication.
That is why next-generation AI networking is about much more than raw link speed.
The NIC is becoming part of the fabric
ConnectX-9 is not simply a faster Ethernet adapter.
The Rubin networking architecture brings more intelligence toward the endpoint, including functions around congestion control, traffic shaping and packet scheduling.
That matters because AI traffic is highly synchronized.
Large numbers of GPUs can transmit at almost exactly the same time.
Instead of relying only on switches to react after queues build, the endpoint can participate in controlling how traffic enters the fabric.
Conceptually:
GPU
↓
NCCL
↓
SuperNIC
↕
Congestion awareness
↕
Spectrum-X Fabric
The NIC is becoming an active participant in fabric behavior.
Optics are becoming part of the compute architecture
At 800G and 1.6T-class networking, the physical layer becomes a serious design constraint.
Traditional architecture:
Switch ASIC
↓
SerDes
↓
PCB
↓
Pluggable Optic
↓
Fiber
NVIDIA is also building Spectrum-6 systems around co-packaged optics, bringing optical engines much closer to the switching silicon.
At AI-factory scale, networking has to optimize more than bandwidth.
It also has to optimize:
- power
- thermal load
- signal integrity
- reliability
- useful GPU time
The network is becoming part of the compute-efficiency equation.
The bigger idea: one AI factory
Vera Rubin goes beyond a single NVL72 rack.
The broader platform combines specialized systems for compute, networking and infrastructure services.
VERA RUBIN AI FACTORY
Vera Rubin NVL72
Compute
│
▼
Spectrum-6
Ethernet Fabric
│
┌───────┴────────┐
│ │
BlueField-4 Vera CPU
Infrastructure Compute
│ │
└───────┬────────┘
│
Large AI Factory
The architectural shift is important.
We used to think:
Server → Network → Server
Then:
GPU Server → AI Fabric → GPU Server
Now we are moving toward:
AI Factory = Compute + Network + Storage + Memory + Power + Cooling
all designed together.
From months to as little as 14 days onsite
One of the most interesting Rubin announcements is not actually about the GPU.
It is about deployment speed.
At GTC Berlin 2026, NVIDIA introduced DSX POD.
Instead of building every AI factory entirely onsite, major infrastructure can be pre-integrated before it reaches the final location.
That can include:
- compute
- networking
- power
- cooling
- integration
- validation
The traditional process looks roughly like:
Building
↓
Power
↓
Cooling
↓
Racks
↓
Networking
↓
Cabling
↓
Integration
↓
Validation
Months
With DSX POD:
Factory Integration
↓
Compute + Network
Power + Cooling
↓
Prevalidated POD
↓
Ship to prepared site
↓
Connect utilities
↓
AI Factory
NVIDIA says that once site utilities are ready, onsite deployment can shrink from roughly six months to as little as 14 days.
That distinction matters.
NVIDIA is not saying an NVL72 rack itself takes 14 days to build.
It is saying a pre-integrated Vera Rubin AI factory can potentially be deployed onsite in that timeframe after the site is prepared.
The bottleneck is no longer simply:
How quickly can we manufacture GPUs?
It becomes:
How quickly can we turn power, cooling, racks, networking and fiber into a functioning AI factory?
What changes for network engineers?
Vera Rubin makes one thing clear:
The network can no longer be treated as infrastructure sitting underneath the compute.
It is part of the compute system.
Troubleshooting increasingly needs to follow the workload:
AI Job
↓
GPU
↓
NCCL
↓
NVLink
↓
ConnectX-9
↓
RDMA / RoCE
↓
Spectrum-6
↓
Queue / Congestion
↓
FEC / Optics
↓
Remote GPU
The useful question is no longer simply:
Is the network up?
It becomes:
Is the network allowing the GPUs to make useful progress?
That is a very different operational model.
Final thought
Vera Rubin is easy to misunderstand if we look only at the Rubin GPU.
The more interesting architecture is:
Vera CPU → Rubin GPU → NVLink 6 → ConnectX-9 → Spectrum-6 → optics → thousands of GPUs → complete AI factories
Blackwell helped establish the idea that the rack can become the computer.
Vera Rubin pushes that concept further.
The AI factory itself is becoming the computer.
And perhaps the most important metric is no longer just FLOPS.
It is:
How quickly and efficiently can we turn megawatts, racks and fiber into useful tokens?