In this note
For Vera Rubin, NVL144 and NVL72 are the same rack. NVIDIA called it Vera Rubin NVL144 through 2025, counting 144 GPU dies, then renamed it Vera Rubin NVL72 (often shortened to VR200) just before its CES launch in January 2026. The new name counts the 72 Rubin GPU packages, each of which holds two dies.
The confusion comes from NVIDIA reusing the number. NVL144 now also names a Kyber rack with 144 Rubin Ultra GPUs, which comes after Vera Rubin. NVL576 has changed meaning as well, and Rubin CPX, the chip behind "NVL144 CPX", was pulled in March 2026.
What does the NVL number mean?
The number after NVL is the size of the NVLink domain: how many GPUs share one all-to-all NVLink switch fabric. Any GPU in the domain can read and write any other GPU's memory at full NVLink bandwidth, so a model can be split across them as if they were one large accelerator. Traffic that leaves the domain goes over the scale-out network (InfiniBand or Ethernet), which is much slower.
What NVIDIA hasn't been consistent about is what counts as one GPU. Since Blackwell, its top data center GPU has been a package with two reticle-sized compute dies joined by a die-to-die link. A GB200 rack holds 72 of those packages, or 144 dies, depending on how you count.
The NVL numbers in NVIDIA's current documents:
- NVL72: one Oberon rack of 72 packages (GB200, GB300, Vera Rubin, and one of the Rubin Ultra options)
- NVL8: an HGX server such as HGX Rubin NVL8, with eight GPUs on one NVLink switch board
- NVL4: Vera Rubin NVL4, four Rubin GPUs on an NVLink bridge with two Vera CPUs, aimed at scientific computing
- NVL144, NVL576 and NVL1152: the larger Rubin Ultra and Feynman domains covered below
Older GB200 material also mentions NVL36x2, which split the 72 GPUs across two racks of 36 for sites that couldn't take a full-density rack.
How the names changed from 2024 to 2026
| Date | What happened |
|---|---|
| March 2024 | GTC: GB200 NVL72 announced, counting packages |
| March 2025 | GTC: Jensen Huang says "Blackwell was named wrong" and the rack should have been NVL144. Rubin is shown as Vera Rubin NVL144 for the second half of 2026 and Rubin Ultra as NVL576 for the second half of 2027, both counting dies |
| September 2025 | Rubin CPX announced, with a Vera Rubin NVL144 CPX rack due at the end of 2026 |
| Late December 2025 | Name changed back to Vera Rubin NVL72, counting packages (per SemiAnalysis) |
| January 2026 | CES: NVIDIA launches Vera Rubin NVL72, with 72 Rubin GPUs and 36 Vera CPUs |
| March 2026 | GTC: Kyber shown as a 144-GPU rack for Vera Rubin Ultra NVL144; NVL576 redefined as eight racks of 72; Kyber NVL1152 previewed for Feynman; CPX missing from the roadmap |
| June and July 2026 | SemiAnalysis reports that Rubin Ultra drops from four dies to two and that Kyber slips to 2028. NVIDIA's only comment: "Our roadmap is intact" |
At GTC 2025 NVIDIA counted dies, so a Rubin rack with "the same 144 total GPU dies" as Blackwell became NVL144, as Tom's Hardware reported (opens in a new tab). NVIDIA's own keynote recap (opens in a new tab) listed Vera Rubin NVL144 for the second half of 2026 and Rubin Ultra for the second half of 2027.
According to SemiAnalysis (opens in a new tab), the switch back to packages happened in late December 2025, and the CES announcement (opens in a new tab) used the new name. By GTC 2026 NVIDIA was counting packages everywhere. Its Vera Rubin POD write-up (opens in a new tab) gives Vera Rubin Ultra three scale-up options: "NVL72, NVL144, and the flagship NVL576".
Oberon vs Kyber
Both are rack designs. Oberon is the name most coverage uses for the rack behind every NVL72 system so far; NVIDIA's 2026 documents call it the third-generation MGX rack. Kyber is what NVIDIA calls "the next-generation MGX NVL rack design", which doubles the NVLink domain per rack to 144 GPUs. It will arrive first with Vera Rubin Ultra and then carry Feynman.
| Oberon | Kyber | |
|---|---|---|
| Systems | GB200, GB300 and Vera Rubin NVL72; Rubin Ultra NVL72 and NVL576 | Vera Rubin Ultra NVL144, Feynman NVL1152 |
| GPU packages per rack | 72 | 144 |
| Compute layout | 18 horizontal compute trays, 4 GPUs each | 36 vertical blades in two canisters, 4 Rubin Ultra GPUs and 2 Vera CPUs per blade (GTC 2026 prototype) |
| NVLink fabric | Copper cable spine at the back of the rack; Vera Rubin trays have no internal cables | Two PCB midplanes between compute and switch blades, 72 NVLink 7 switches per rack (prototype) |
| Rack power | About 120 kW for GB200 and 135 kW for GB300; 187 kW (Max-Q) to 227 kW (Max-P) for Vera Rubin | Up to 600 kW cited at GTC 2025 for the first design; no current figure |
| Cooling | Liquid, 45°C inlet on Vera Rubin | Liquid |
| Status | Shipping | Due with Rubin Ultra; reported slip to 2028 |
The Kyber layout above comes from SemiAnalysis's GTC 2026 recap (opens in a new tab), which describes the hardware on stage as prototypes and expects parts of it, the connectors for one, to change for production. The 2025 design looked different again, with four canisters instead of two.
The midplane is the hard part. Tom's Hardware (opens in a new tab), citing trade analyses, describes a passive board close to a square meter in area and 78 layers thick that carries every GPU-to-GPU link in the cabinet. Doing the same job with cables would take more than 20,000 of them.
In July 2026 SemiAnalysis reported that manufacturing problems with that board had pushed Kyber back by more than a year, to 2028. It also reported that NVIDIA had dropped NVL72x2, a stopgap that put two Oberon racks back to back, after its largest customers objected to running two linked cabinets as one unit. NVIDIA hasn't confirmed either report.
For site planning, Oberon generations differ mostly in power per rack. Kyber is a new rack, and nothing NVIDIA has published yet tells you what it needs. Our note on building a data center for GPU clusters covers what today's racks ask of a site.
NVL72 vs NVL576
NVL576 has kept its number and changed what it describes.
At GTC 2025 it was a single Kyber rack of Rubin Ultra: 144 packages with four dies each, 576 dies in total, rated at 15 exaflops of FP4 inference and due in the second half of 2027.
At GTC 2026 NVIDIA redefined it as eight separate racks, "each with 72 Rubin Ultra GPUs, all in a single 576-GPU NVLink domain with copper and direct optical connections." Copper stays inside each rack, and optics connect the racks in a two-layer all-to-all topology. NVIDIA has built a working GB200-based prototype of the multi-rack design, called Polyphe.
So older coverage of a 576-GPU rack means one cabinet, and current coverage means eight. SemiAnalysis expects NVL576 to slip, or ship in low volume, until the co-packaged optics it depends on are ready. NVIDIA plans the same trick one size up for Feynman: Kyber NVL1152, which works out to eight 144-GPU racks.
Every system shipping today is a 72-GPU Oberon rack. If the reported Kyber delay holds, SemiAnalysis's reading is that the largest single Rubin Ultra NVLink domain you can get in 2027 will also be 72 GPUs.
What happened to Rubin CPX?
Rubin CPX was a cheaper GPU for the prefill (context) phase of inference: a single die with 128 GB of GDDR7 instead of HBM, rated at up to 30 petaflops of NVFP4. NVIDIA announced it (opens in a new tab) in September 2025 along with the Vera Rubin NVL144 CPX rack, quoted at 8 exaflops, 100 TB of fast memory and 1.7 PB/s of memory bandwidth, for the end of 2026.
SemiAnalysis (opens in a new tab) worked out that the rack kept the usual 72 Rubin packages and added 144 CPX GPUs across the same 18 trays. The "144" in the name counted Rubin dies, and the CPX chips came on top.
CPX was missing from the GTC 2026 slides. In a press session (opens in a new tab) afterwards, NVIDIA's Ian Buck said: "we've pulled CPX. It's still a good idea," and that NVIDIA would be "thinking about CPX more in the next generation." The 2026 platform pairs Vera Rubin NVL72 with Groq 3 LPX racks of 256 LPUs, which handle decode, the other phase of inference.
If a vendor quotes an NVL144 CPX rack for delivery in 2026 or 2027, ask what's actually in it.
Reference table
| Name | Rack | What the number counts | GPU packages | Dies per package | Status |
|---|---|---|---|---|---|
| GB200 NVL72 | Oberon | Packages | 72 | 2 | Shipping |
| GB300 NVL72 | Oberon | Packages | 72 | 2 | Shipping |
| Vera Rubin NVL144 | Oberon | Dies | 72 | 2 | 2025 name for Vera Rubin NVL72 |
| Vera Rubin NVL72 (VR200) | Oberon | Packages | 72 | 2 | Shipping since the second half of 2026 |
| Vera Rubin NVL144 CPX | Oberon | Rubin dies, plus 144 CPX | 72 Rubin, 144 CPX | 2 (Rubin), 1 (CPX) | Pulled in March 2026 |
| Rubin Ultra NVL576 (2025 plan) | Kyber | Dies | 144 | 4 | Superseded |
| Vera Rubin Ultra NVL72 | Oberon | Packages | 72 | 4 announced, 2 reported | Planned for 2027 |
| Vera Rubin Ultra NVL144 | Kyber | Packages | 144 | 4 announced, 2 reported | Planned; reported slip to 2028 |
| Vera Rubin Ultra NVL576 | 8 Oberon racks | Packages | 576 | 4 announced, 2 reported | Planned; low volume expected first |
| Feynman NVL1152 | 8 Kyber racks | Packages | 1,152 | Not published | After Rubin Ultra |
| HGX Rubin NVL8 | Server | Packages | 8 | 2 | In NVIDIA's current datasheet |
The two-die Rubin Ultra comes from a SemiAnalysis report (opens in a new tab) in June 2026. We haven't seen NVIDIA confirm it.
How to read a spec sheet or quote
Most of the confusion comes from mixed units. Five things to check:
- Whether GPU counts are packages or dies. NVIDIA's October 2026 datasheet (opens in a new tab) counts packages: 72 Rubin GPUs and 20.7 TB of HBM4 per rack, or 288 GB per GPU. Per-GPU memory, FLOPS and prices should use the same unit.
- Which rack it is. A Vera Rubin rack is Oberon. Kyber NVL144 is a different cabinet and will need a different site.
- Whether it's one NVLink domain. NVL576 and NVL1152 span eight racks, so for any multi-rack system, ask whether it's one NVLink domain or several racks joined over the scale-out network.
- Which power profile the figures assume. The datasheet lists Vera Rubin NVL72 at 187 kW in Max-Q and 227 kW in Max-P, and 156 kW with NVIDIA's MaxLPS power management.
- Which generation it is. "Rubin" can mean the 2026 GPU or Rubin Ultra, and "NVL144" can mean the renamed Vera Rubin rack or a Kyber rack that won't ship before 2027.
If you're deciding between the generation that ships now and the next one, our GB300 or wait for Vera Rubin note works through it.
Building or reserving an NVL72 rack
Every NVL72 generation so far fits the same Oberon footprint but asks more of the site each time: more power per rack, all of it liquid-cooled. We design, source, build and operate NVL72 clusters, on your site or in a Tier III facility in the region you choose. See GB300 NVL72, Vera Rubin NVL72, cluster deployment and reserved clusters, or talk to us.
Questions this note answers
NVL72 means 72 GPUs in one NVLink domain, so every GPU can reach every other at full NVLink bandwidth through the rack's switch trays. In NVIDIA's current naming each GPU is one package, which on Blackwell and Rubin holds two compute dies.
Yes. NVL144 was the 2025 name, counting 144 GPU dies. NVIDIA renamed the rack in late December 2025 and launched it at CES in January 2026 as Vera Rubin NVL72, counting its 72 GPU packages.
Oberon is the 72-GPU rack used by GB200, GB300 and Vera Rubin, with horizontal compute trays. Kyber is the next rack, built for 144 GPUs in one NVLink domain, with vertical compute blades connected through PCB midplanes. It's planned for Rubin Ultra and then Feynman.
At GTC 2025 NVIDIA used NVL576 for one Kyber rack of Rubin Ultra, counting 576 dies. Since GTC 2026 it means eight racks of 72 Rubin Ultra GPUs joined into one 576-GPU NVLink domain, with optical links between the racks.
NVIDIA dropped Rubin CPX at GTC in March 2026 to focus on Groq 3 LPX racks for decode. Ian Buck said NVIDIA would think about CPX more in the next generation.
Sources
- Tom's Hardware: Nvidia announces Rubin GPUs in 2026, Rubin Ultra in 2027 (March 18, 2025) (opens in a new tab)
- NVIDIA Blog: GTC 2025 keynote recap (March 18, 2025) (opens in a new tab)
- Tom's Hardware: Rubin Ultra with 600,000-watt Kyber racks (March 19, 2025) (opens in a new tab)
- NVIDIA Newsroom: NVIDIA unveils Rubin CPX (September 9, 2025) (opens in a new tab)
- SemiAnalysis: Rubin CPX specialized accelerator rack (September 10, 2025) (opens in a new tab)
- NVIDIA Newsroom: NVIDIA kicks off the next generation of AI with Rubin (January 5, 2026) (opens in a new tab)
- SemiAnalysis: Vera Rubin, Extreme Co-Design (February 25, 2026) (opens in a new tab)
- NVIDIA Technical Blog: NVIDIA Vera Rubin POD (March 16, 2026) (opens in a new tab)
- Tom's Hardware: Nvidia removes Rubin CPX from its roadmap (March 17, 2026) (opens in a new tab)
- Tom's Hardware: GTC 2026 press Q&A with Ian Buck (March 23, 2026) (opens in a new tab)
- SemiAnalysis: GTC 2026, The Inference Kingdom Expands (March 24, 2026) (opens in a new tab)
- Tom's Hardware: Nvidia reportedly cancels quad-die Rubin Ultra (June 30, 2026) (opens in a new tab)
- Tom's Hardware: Nvidia's Kyber rack for Rubin Ultra reportedly slips to 2028 (July 6, 2026) (opens in a new tab)
- OEM product guide: GB300 NVL72 rack-scale AI (opens in a new tab)
- NVIDIA Vera Rubin NVL72 product page (opens in a new tab)
- NVIDIA Vera Rubin datasheet (October 2026) (opens in a new tab)