NVLink 6 is the interconnect that lets 72 GPUs behave like one. It moves 3.6 TB/s to and from every GPU — double NVLink 5 — and 260 TB/s across a single Vera Rubin NVL72 rack. That is not a spec-sheet flex. It is the reason the next generation of AI models will be shaped differently than the last one, and the reason a lot of other hardware in the data center is about to get rebuilt around it.
Jensen Huang unveiled the Rubin platform at CES 2026 and filled in the roadmap at GTC in March. NVLink 6 is the quiet centerpiece. Here is what it does, and what it makes possible.
The numbers, briefly
- 3,600 GB/s per GPU, bidirectional — 2× NVLink 5, via 36 links per GPU instead of 18.
- 260 TB/s aggregate in one NVL72 rack, up from 130 TB/s on Blackwell.
- 28.8 TB/s per switch tray, built from four NVLink 6 switch chips.
- 130 TFLOPS of FP8 compute inside the network itself — the switch does math, it does not just forward packets.
- 3× lower latency and 10× the packet rate of off-the-shelf Ethernet, by NVIDIA's own measure.
- The switch chip runs hot enough that liquid cooling is mandatory, not optional.
One more detail that engineers will appreciate more than analysts: the Rubin rack is a cable-free modular tray design. NVIDIA says rack assembly drops from about 100 minutes on Blackwell to roughly 6 minutes. When you are standing up thousands of racks, that is a schedule, not a convenience.
Why in-network compute is the sleeper feature
Training and inference at scale are dominated by collective operations — all-reduce, broadcast, gather. Traditionally the GPUs do that math and the network just hauls the results around. NVLink 6 pushes the reduction into the switch fabric. The data gets summed on the way through.
The practical effect is that the tax you pay for splitting a model across many GPUs goes down. Every generation of that tax reduction moves the line on what is worth parallelizing at all. Which brings us to the interesting part.
What this makes possible — some projections
These are informed extrapolations, not announcements. Take them as the shape of the argument rather than a spec sheet.
Mixture-of-experts models get much wider. MoE architectures are bottlenecked by all-to-all expert routing — every token has to find its expert, and that traffic is brutal. A non-blocking all-to-all fabric at 3.6 TB/s per GPU makes far more aggressive expert counts economically sane. Expect sparser, wider models with more total parameters and similar activated cost.
Disaggregated inference becomes the default. NVIDIA is already shipping Rubin CPX, a separate accelerator aimed at the prefill/context phase, alongside standard Rubin GPUs for decode. Splitting a single request across specialized silicon only works if the interconnect between them is effectively free. NVLink 6 is what makes that architecture viable instead of academic.
Memory stops being per-GPU. With 288 GB of HBM4 per Rubin GPU and a rack-wide fabric this fast, the useful unit of memory becomes the rack, not the card. That changes how you think about context windows, KV cache retention, and how long an agent can hold state before something has to be evicted.
Long-running agents get cheaper to keep alive. Most of the cost of an agent that thinks for an hour is memory pressure and coordination overhead. Both are interconnect problems in disguise.
The knock-on effects on everything else
An interconnect this fast is a forcing function. Everything adjacent has to move.
Memory. HBM4 at 288 GB per GPU is table stakes for Rubin; Rubin Ultra is projected at 1 TB of HBM4e per package in 2027. Feeding a 3.6 TB/s fabric with anything slower would be pointless.
Power and cooling. Rubin GPUs land near 1.8 kW each. The Rubin Ultra NVL576 “Kyber” rack is expected around 600 kW in the second half of 2027. Air cooling is finished at the rack level, and facility power design has to be planned years ahead. Huang has said as much publicly — land and power now get planned two to three years out.
Copper versus optics. This is the one we find most interesting. NVIDIA's networking SVP Gilad Shainer has been blunt that copper wins wherever it physically can: "There is nothing better than copper. Copper is zero power." Passive, no active components, no transceiver failure modes. The catch is reach — as signaling rates climb, usable copper distance shrinks. NVL72 and NVL144 fit inside copper's envelope. NVL576 and the projected NVL1152 do not, because they span multiple racks. So the roadmap is hybrid: copper stays for the in-rack layer, and co-packaged optics takes the spine, arriving in force with Feynman around mid-to-late 2028. Copper is not being replaced. It is being concentrated into the layer where it is unbeatable.
The rest of the network. NVLink 6 handles scale-up inside the rack. Scale-out across racks lands on ConnectX-9 SuperNICs at 1.6 Tb/s, BlueField-4 DPUs, and Spectrum-X Ethernet with co-packaged optics and 200G SerDes. Two networks, two jobs, and the boundary between them is where most real-world deployment mistakes happen.
The competitive picture. NVIDIA opened the door with NVLink Fusion, licensing the fabric to MediaTek, Marvell, Alchip, Astera Labs, Synopsys, and Cadence — with Arm now in the mix and a $2B NVIDIA investment in Marvell behind it. The counterweight is UALink, the vendor-neutral consortium spec. The strategic read is simple: NVIDIA would rather your custom accelerator plug into its fabric than into someone else's. If that works, NVLink stops being a product and becomes an industry substrate.
Why we pay attention to this
Resilient Tec does not sell NVL72 racks. What we do is supply the interconnect and infrastructure hardware that surrounds this world — DACs, AOCs, structured cabling, switching gear — often at a fraction of OEM pricing. The reason we track a 600 kW rack we will never stock is that architecture at the top of the market defines the requirements at every tier below it within about three years. The copper-versus-optics boundary NVIDIA is drawing right now is the same boundary a 16-GPU cluster in a university lab will be arguing about in 2028.
Staying current is not a marketing posture. It is how you know which cable someone is actually going to need.
Frequently asked questions
What is NVLink 6?
NVLink 6 is the sixth generation of NVIDIA's scale-up GPU interconnect, introduced with the Vera Rubin platform in 2026. It delivers 3.6 TB/s of bidirectional bandwidth per GPU using 36 links, double the 1.8 TB/s of NVLink 5 on Blackwell.
How much bandwidth does the NVLink 6 Switch provide?
Each NVLink 6 switch tray contains four switch chips and provides 28.8 TB/s of bandwidth plus 14.4 TFLOPS of FP8 in-network compute. A full Vera Rubin NVL72 rack reaches 260 TB/s of aggregate GPU bandwidth and roughly 130 TFLOPS of in-network compute.
What is in-network compute and why does it matter?
In-network compute lets the switch fabric perform collective operations such as all-reduce, reduce and broadcast as data passes through it, rather than making the GPUs do that math. It reduces the overhead of splitting a model across many GPUs, which is the main cost of large-scale training and distributed inference.
Is NVLink 6 copper or optical?
Copper, within the rack. NVIDIA has stated that copper is preferred wherever distance allows because it is passive and consumes no power. NVL72 and NVL144 configurations fit within copper's reach. Larger multi-rack scale-up domains such as NVL576 and the projected NVL1152 exceed it, so those move the spine layer to co-packaged optics, expected in volume with the Feynman generation around 2028.
How does NVLink compare to Ethernet or InfiniBand?
They solve different problems. NVLink is a scale-up fabric that makes GPUs inside one domain behave as a single accelerator, with roughly 3x lower latency and 10x the packet rate of off-the-shelf Ethernet. InfiniBand and Ethernet are scale-out fabrics that connect those domains together across a data center. Rubin deployments use both: NVLink 6 inside the rack, ConnectX-9 SuperNICs and Spectrum-X Ethernet between them.
What is NVLink Fusion?
NVLink Fusion licenses the NVLink scale-up fabric to third parties so custom CPUs and accelerators can join an NVIDIA rack. Licensees include MediaTek, Marvell, Alchip, Astera Labs, Synopsys and Cadence. The vendor-neutral alternative is UALink, backed by a multi-company consortium.
Sources
- NVIDIA — NVLink & NVLink Switch
- NVIDIA Developer Blog — NVLink: The Scale-Up Network for AI Factories
- NVIDIA Developer Blog — Inside the Vera Rubin Platform
- ServeTheHome — Rubin platform launch at CES 2026
- Tom's Hardware — Vera Rubin platform in depth
- The Register — Nvidia embraces optical scale-up as copper reaches limits
- SDxCentral — Nvidia backs copper amid CPO push
- DataCenterDynamics — Rubin Ultra NVL576 expected at 600kW
- Network World — Nvidia opens NVLink to competitive processors
Projections in this article are our own analysis, clearly marked as such, and are not statements from NVIDIA.