# NVIDIA Rubin Ultra

> Source: https://aiwiki.ai/wiki/nvidia_rubin_ultra
> Updated: 2026-10-16
> Fact-checked: 2026-10-16
> Categories: AI Compute, AI Hardware, AI Infrastructure, NVIDIA
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "NVIDIA Rubin Ultra." aiwiki.ai, 16 Oct 2026. https://aiwiki.ai/wiki/nvidia_rubin_ultra
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**NVIDIA Rubin Ultra** is a planned data-center GPU platform from [Nvidia](https://aiwiki.ai/wiki/nvidia), the mid-cycle "Ultra" refresh of the Rubin generation. Chief executive [Jensen Huang](https://aiwiki.ai/wiki/jensen_huang) first put it on NVIDIA's roadmap, for 2027, in his Computex keynote in June 2024, and gave detailed targets during the GTC 2025 keynote in March 2025, when NVIDIA scheduled it for the second half of 2027.[24][6][1][2][3] Rubin Ultra follows the base [Vera Rubin](https://aiwiki.ai/wiki/nvidia_vera_rubin) platform, which began production shipments in August 2026.[22] In 2025 it was tied to a rack-scale system called **NVL576** in a redesigned rack codenamed **Kyber**; at GTC 2026 NVIDIA recast the line-up into three scale-up options (NVL72, a Kyber-based NVL144, and a multi-rack NVL576).[1][4][12]

NVIDIA's 2025 roadmap gave Rubin Ultra four reticle-sized compute dies per GPU package (twice the base Rubin's two), 1 TB of HBM4e per package, and 15 exaflops of FP4 inference and 5 exaflops of FP8 training per NVL576 rack. NVIDIA put that rack at roughly 14 times its Blackwell-generation [GB300 NVL72](https://aiwiki.ai/wiki/nvidia_gb300_nvl72), and Huang said it would draw about 600 kilowatts.[1][2][6][8] NVIDIA showed a Rubin Ultra compute tray with four compute chiplets and 1 TB of HBM4E at GTC 2026.[13] Since then, analyst and press reports have said NVIDIA dropped the four-die package for a two-die design, is weighing far smaller HBM configurations, and has seen the Kyber rack slip to 2028. NVIDIA has not confirmed any of these changes. It has said its roadmap is "intact" and that it is "optimizing across compute, networking and memory" for Rubin Ultra.[17][18][19][20] As of October 2026 Rubin Ultra is unreleased, and its final specifications have not been published.

## Position on the NVIDIA roadmap

At GTC 2025, NVIDIA restated the yearly data-center cadence it had first shown at Computex 2024,[24] in which each architecture is followed roughly twelve months later by an "Ultra" variant that increases compute and memory within the same family before the next architecture arrives.[2][6] The sequence ran from [Blackwell](https://aiwiki.ai/wiki/nvidia_blackwell) (2024) and [Blackwell Ultra](https://aiwiki.ai/wiki/nvidia_blackwell_ultra) (2025) to Rubin (2026), Rubin Ultra (2027), and then an architecture named for physicist Richard Feynman ([Feynman](https://aiwiki.ai/wiki/nvidia_feynman), targeted for 2028).[2][6][7] Rubin and Rubin Ultra are named for the American astronomer Vera Rubin, and both pair their GPUs with NVIDIA's Arm-based [Vera (CPU)](https://aiwiki.ai/wiki/nvidia_vera_cpu).[2][11]

The 2025 roadmap also separated the two Rubin racks. Base Rubin would use the existing Oberon-style rack, at the time called NVL144, while Rubin Ultra would move to the larger Kyber-based NVL576.[4][5] NVIDIA later dropped the die-count naming for the base system: its January 2026 launch materials call it Vera Rubin NVL72, a rack of 72 Rubin GPUs and 36 Vera CPUs.[11] In September 2025 NVIDIA also announced the [Rubin CPX](https://aiwiki.ai/wiki/nvidia_rubin_cpx), a GPU for the long-context phase of inference that works alongside Vera CPUs and Rubin GPUs in a Vera Rubin NVL144 CPX platform.[23]

The table below lists the headline figures NVIDIA presented for each step. Rubin Ultra's entries are the GTC 2025 targets, and later reporting suggests some of them may change.

| Platform | Target window | Rack / system | GPU memory (per package) | Per-rack FP4 (dense) | Notes |
|---|---|---|---|---|---|
| Blackwell | 2024 | GB200 NVL72 | 192 GB HBM3E | 0.72 exaflops | B200, 10 PFLOPS FP4 per package |
| Blackwell Ultra | 2025 | GB300 NVL72 | 288 GB HBM3E | 1.1 exaflops | B300, 15 PFLOPS FP4 per package |
| Rubin | 2026 | Vera Rubin NVL72 (announced as NVL144) | 288 GB HBM4 | 3.6 exaflops | 50 PFLOPS NVFP4 per package; 2 dies |
| Rubin Ultra (2025 plan) | 2H 2027 | NVL576 (Kyber) | 1 TB HBM4e | 15 exaflops | 100 PFLOPS FP4 per package; 4 dies |
| Feynman | 2028 | (not detailed) | next-generation HBM | not disclosed | named for Richard Feynman |

Sources for the table: Blackwell and Blackwell Ultra figures from Tom's Hardware's roadmap tables;[2][8] Rubin memory and NVFP4 figures from NVIDIA's Rubin architecture blog;[15] rack figures from the GTC 2025 slides as reported.[2][6]

## The GTC 2025 announcement

NVIDIA used the GTC 2025 keynote, held in San Jose in March 2025, to extend its published roadmap and to put hardware mockups of future systems on stage.[1][2] GTC 2025 gave the systems names (NVL144 for Rubin, NVL576 for Rubin Ultra), put Rubin Ultra in the second half of 2027, and revealed the Kyber rack and the Feynman architecture after it.[2][6][7]

Rubin Ultra was more than two years away, so NVIDIA showed it mainly through renderings, rack mockups, and roadmap slides. Tom's Hardware wrote at the time that "there are no hard specifications yet for Rubin Ultra, but there are performance targets."[1] The main physical exhibit was the Kyber rack and its midplane. SemiAnalysis reported that the rack on the show floor was a Blackwell-based test vehicle and that the architecture was not yet final.[4][5]

## The NVL576 rack and the Kyber design (2025 version)

As presented in 2025, the NVL576 used a counting convention NVIDIA introduced for the Rubin generation, in which each reticle-sized compute die counted as a "GPU."[4][8] An NVL576 rack would hold 144 Rubin Ultra GPU packages. With four compute dies per package, that came to 576 GPU dies in one [NVLink](https://aiwiki.ai/wiki/nvlink) domain.[4][8] SemiAnalysis called this die-counting rule the latest instance of "Jensen math," and predicted it would be "a constant point of clarification."[4]

Kyber is the codename for the rack architecture that succeeds Oberon. NVIDIA describes it as rotating compute blades vertically, "like books on a shelf," with up to 18 compute blades per chassis and NVLink switch blades joined at the back through a cable-free midplane.[9] In SemiAnalysis's description of the 2025 design, each rack held four canisters of 18 compute blades, and each blade carried two Rubin Ultra GPU packages and two Vera CPUs.[4] ServeTheHome photographed the NVL576 Kyber midplane at GTC 2025. It had 18 columns and four rows of connectors and replaced the cable cartridges of the Blackwell-era racks such as the [GB200 NVL72](https://aiwiki.ai/wiki/nvidia_gb200_nvl72). Because the whole rack is liquid cooled, the midplane did not need cutouts for airflow.[5] NVIDIA's Ian Buck told The Next Platform that scale-up links would stay on copper up to and including the Kyber racks.[6]

Huang confirmed during a question-and-answer session that a Rubin Ultra NVL576 rack could draw up to 600 kilowatts. Tom's Hardware put a Blackwell B200 rack at up to about 120 kilowatts.[1][3] NVIDIA paired Kyber with an 800-volt DC power architecture: it exhibited an 800 V sidecar for a Kyber rack at GTC 2025 and said full-scale production of 800 VDC data centers would coincide with Kyber systems in 2027.[10] The scale-up fabric uses NVLink 7, which NVIDIA's slides put at 1.5 petabytes per second for the rack, and ConnectX-9 links carry scale-out traffic between racks.[2][6]

## The GPU package: four dies and HBM4e

NVIDIA's 2025 roadmap gave each Rubin Ultra package four reticle-sized GPU dies, double base Rubin's two.[2][6][8] SemiAnalysis and Tom's Hardware described the package as a row of four compute dies flanked by two I/O chiplets on TSMC CoWoS-L packaging.[4][8] Tom's Hardware wrote that such a package would need about 3,600 W.[8]

For memory, NVIDIA specified 1 TB of HBM4e per Rubin Ultra package across 16 HBM stacks. The base Rubin package had 8 stacks, and NVIDIA's slides put its capacity at 288 GB.[2][6][8] NVIDIA's later Rubin documentation confirms 288 GB of [HBM4](https://aiwiki.ai/wiki/hbm4) per Rubin GPU, built from 12-Hi stacks with up to 22 TB/s of bandwidth.[15] On these figures a four-die Rubin Ultra package would hold more than 3.5 times the memory of a Rubin package.[4] Tom's Hardware listed 32 TB/s of memory bandwidth per Rubin Ultra package.[8] At rack level, the GTC 2025 slides listed 4.6 PB/s of HBM4e bandwidth and 365 TB of "fast memory" for NVL576. For comparison, the base Rubin NVL144 rack was listed at 75 TB of fast memory, a figure that combines GPU and CPU memory.[2] Tom's Hardware's roadmap table puts the NVL576's HBM capacity alone at 147 TB.[8]

## Updates at GTC 2026

At GTC 2026 in March 2026, NVIDIA showed a Rubin Ultra compute tray. Tom's Hardware reported that the GPU had four compute chiplets and 1 TB of HBM4E, and that it was due in 2027. The tray had almost no cables, and the heatspreader covered the package, so it was not possible to see how the GPU is built. Tom's Hardware added that it was not known whether Rubin Ultra had taped out.[13]

NVIDIA's technical blog for the event changed how the systems are counted and grouped. It lists three NVLink scale-up options for Vera Rubin Ultra:[12]

| Option | Configuration described by NVIDIA |
|---|---|
| NVL72 | MGX NVL rack with 72 Rubin Ultra GPUs |
| NVL144 | Kyber rack, which "will double the NVLink domain per rack to fit 144 GPUs"; Kyber's first appearance, as a standalone system |
| NVL576 (flagship) | Eight MGX NVL racks of 72 Rubin Ultra GPUs each, joined into a single 576-GPU NVLink domain "with copper and direct optical connections" in a two-layer all-to-all topology |

Under this scheme, "NVL576" no longer means 576 dies in one Kyber rack. It means 576 GPUs spread across eight racks. NVIDIA said Kyber would later scale into an all-to-all NVL1152 system for the Feynman generation, using direct optical rack-to-rack links.[12] SemiAnalysis reported that in the updated Kyber prototype each compute blade carries four Rubin Ultra GPUs and two Vera CPUs, with two canisters of 18 blades for 144 GPUs per rack and 72 NVLink 7 switches. It said Rubin Ultra NVL144 would use only copper for scale-up, and that [co-packaged optics](https://aiwiki.ai/wiki/co_packaged_optics) would first be used to link the eight racks of NVL576. It expected NVL576 to ship in low volume, mainly for testing.[14]

## Reported changes in 2026

After GTC 2026, a series of supply-chain and analyst reports said Rubin Ultra's design and memory configuration were being cut back. NVIDIA has not officially announced any of these changes. They are summarized here as reported claims.

| Date (2026) | Source | Reported claim |
|---|---|---|
| April 1 | TrendForce, citing Commercial Times | Foundry planning was focused on a dual-die Rubin Ultra on TSMC N3P with CoWoS-L, because a four-die package would reach about 7.5 to 8 times the reticle limit and hurt yield and cost.[16] |
| June 30 | SemiAnalysis, via Tom's Hardware | NVIDIA cancelled the four-chiplet Rubin Ultra in favor of a two-chiplet design over "manufacturing execution concerns"; the revised part would use eight HBM4E stacks instead of 16.[17] |
| Early July | SemiAnalysis, via Tom's Hardware | The Kyber NVL144 rack slipped by more than a year, from 2027 to 2028, because its PCB midplane was hard to manufacture; a back-to-back two-rack NVL72x2 alternative was cancelled.[18] |
| August 4 | TrendForce | Starting in the third quarter of 2026, NVIDIA widened its evaluation beyond its baseline 12-Hi HBM4e to 8-Hi HBM4e, 12-Hi HBM4, and 8-Hi HBM4; the final specification had not been decided.[19] |
| August | The Information, via Tom's Hardware | NVIDIA tested versions with 192 GB or 256 GB of memory, with fewer than the 16 announced stacks, and with HBM4 in place of HBM4E.[20] |
| September 2 | SemiAnalysis | NVIDIA "has despecced Rubin Ultra's HBM: from HBM4E 12-Hi (384GB) down to HBM4 8-Hi (192GB)," because memory had become about 40% of total cost of ownership after HBM and DRAM price rises.[21] |

TrendForce attributed the memory evaluation to two supply constraints: a DRAM shortage expected through 2027, and uncertainty over the validation schedule and yield ramp of 12-Hi HBM4e. It said NVIDIA's main goal for the generation is higher HBM I/O speed. If HBM4e is validated on schedule, Rubin Ultra's I/O speed could reach 14 to 16 Gbps, compared with 8 to 11.7 Gbps for Rubin; if not, HBM4 optimization might give only 11 to 12 Gbps. TrendForce expected any cut to come from fewer DRAM layers per stack, and it reported that several cloud providers were also considering lower HBM capacities for their in-house AI chips.[19] The three HBM makers are [SK Hynix](https://aiwiki.ai/wiki/sk_hynix), [Samsung](https://aiwiki.ai/wiki/samsung), and [Micron](https://aiwiki.ai/wiki/micron_technology). Tom's Hardware, citing Digitimes, reported that all three had sold out their HBM capacity through 2027.[20] On NVIDIA's August 2026 earnings call, management spoke of "extreme pricing conditions in memory" and said it had "long-standing deep relationships with all three major memory suppliers."[22]

Watch the units when comparing these numbers. NVIDIA's 1 TB figure is for a four-die package with 16 stacks, while Rubin's 288 GB is for a two-die package with 8 stacks.[8][15] Tom's Hardware noted that the 192 GB and 256 GB figures are low "even with a dual-die Rubin Ultra," since each base Rubin GPU already has 288 GB.[20] The Information's report did not say how many compute dies, or which memory type, went with each tested capacity.[20]

NVIDIA's public responses have not confirmed or denied the specific claims. On the Kyber report, a spokesperson told Tom's Hardware, "Our roadmap is intact." Huang separately called reports of Vera Rubin delays "not true," but he did not address the Rubin Ultra rack.[18] On the memory report, NVIDIA said: "With NVIDIA Rubin Ultra, we are optimizing across compute, networking and memory to deliver the best performance and efficiency for our customers' AI deployments. These optimizations let us build more GPUs and deploy more AI systems."[20]

## The Vera CPU pairing

Like base Rubin, Rubin Ultra is a CPU-plus-GPU platform. It pairs the GPUs with NVIDIA's Vera CPU, an Arm-based processor with 88 custom cores and simultaneous multithreading for 176 threads, connected to the GPUs over a 1.8 TB/s NVLink chip-to-chip interface.[2][6][8] Vera arrives with the 2026 Rubin generation and carries over to Rubin Ultra; the Ultra step changes the GPU and the rack.[2]

## Performance claims

NVIDIA stated Rubin Ultra's headline performance at the NVL576 rack level in 2025: up to 15 exaflops of FP4 ([NVFP4](https://aiwiki.ai/wiki/nvfp4)) for inference and 5 exaflops of FP8 for training, using dense matrices.[2][6] Per package, it cited 100 petaflops of FP4, double the 50 petaflops of the base Rubin package.[2][6][8]

NVIDIA described the NVL576 as about 14 times the performance of the Blackwell Ultra GB300 NVL72 for inference and training. The Next Platform calculated about 21 times the GB200 NVL72.[6][8] The base Rubin NVL144 was listed at 3.6 exaflops of FP4 inference and 1.2 exaflops of FP8 training per rack.[6] All of these are vendor projections for unreleased hardware in low-precision formats, not benchmark results. They describe the 2025 four-die, single-Kyber-rack NVL576, which NVIDIA's GTC 2026 materials have since replaced with an eight-rack NVL576.[12] Tom's Hardware wrote that the reported dual-chiplet redesign was projected to deliver half the performance of the four-chiplet part.[18]

## Timing and status

NVIDIA scheduled Rubin Ultra for the second half of 2027: one year after the base Rubin platform and one year before Feynman.[1][2][6] At GTC 2026 Tom's Hardware still described it as due in 2027.[13] The base platform kept to its schedule. NVIDIA said in January 2026 that Rubin was in full production with partner products due in the second half of 2026, and on its August 26, 2026 earnings call it said production shipments of Vera Rubin had begun that month.[11][22]

For Rubin Ultra, the 2027 target has been questioned by the reported Kyber slip to 2028, and its memory specification is still under evaluation according to TrendForce. NVIDIA maintains that its roadmap is intact.[18][19] Rubin Ultra is a roadmap product, so its die count, memory capacity, rack configurations, and ship date could all differ from both the GTC 2025 figures and the 2026 reports.

## References

[1] Tom's Hardware, "Nvidia shows off Rubin Ultra with 600,000-Watt Kyber racks and infrastructure, coming in 2027," 19 March 2025. https://www.tomshardware.com/pc-components/gpus/nvidia-shows-off-rubin-ultra-with-600-000-watt-kyber-racks-and-infrastructure-coming-in-2027

[2] Tom's Hardware, "Nvidia announces Rubin GPUs in 2026, Rubin Ultra in 2027, Feynman also added to roadmap," 18 March 2025. https://www.tomshardware.com/pc-components/gpus/nvidia-announces-rubin-gpus-in-2026-rubin-ultra-in-2027-feynam-after

[3] DataCenterDynamics, "Nvidia's Rubin Ultra NVL576 rack expected to be 600kW, coming second half of 2027." https://www.datacenterdynamics.com/en/news/nvidias-rubin-ultra-nvl576-rack-expected-to-be-600kw-coming-second-half-of-2027/

[4] SemiAnalysis, "NVIDIA GTC 2025 - Built For Reasoning, Vera Rubin, Kyber, CPO, Dynamo Inference, Jensen Math, Feynman." https://newsletter.semianalysis.com/p/nvidia-gtc-2025-built-for-reasoning-vera-rubin-kyber-cpo-dynamo-inference-jensen-math-feynman

[5] ServeTheHome, "The NVIDIA Rubin NVL576 Kyber Midplane is Huge." https://www.servethehome.com/the-nvidia-rubin-nvl576-kyber-midplane-is-huge/

[6] The Next Platform, "Nvidia Draws GPU System Roadmap Out To 2028," 19 March 2025. https://www.nextplatform.com/2025/03/19/nvidia-draws-gpu-system-roadmap-out-to-2028/

[7] TrendForce, "GTC 2025: NVIDIA Unveils More on Rubin GPUs, Announces Feynman for 2028 in Roadmap Update." https://www.trendforce.com/news/2025/03/19/news-gtc-2025-nvidia-unveils-more-on-rubin-gpus-announces-feynman-for-2028-in-roadmap-update/

[8] Tom's Hardware, "Nvidia enterprise GPU and CPU roadmaps: Rubin, Rubin Ultra, Feynman, and silicon photonics," 6 August 2025. https://www.tomshardware.com/tech-industry/semiconductors/nvidia-enterprise-roadmap-rubin-rubin-ultra-feynman-and-silicon-photonics

[9] NVIDIA Blog, "NVIDIA, Partners Drive Next-Gen Efficient Gigawatt AI Factories in Buildup for Vera Rubin," 13 October 2025. https://blogs.nvidia.com/blog/gigawatt-ai-factories-ocp-vera-rubin/

[10] NVIDIA Technical Blog, "NVIDIA 800 VDC Architecture Will Power the Next Generation of AI Factories," 20 May 2025. https://developer.nvidia.com/blog/nvidia-800-v-hvdc-architecture-will-power-the-next-generation-of-ai-factories/

[11] NVIDIA Newsroom, "NVIDIA Kicks Off the Next Generation of AI With Rubin: Six New Chips, One Incredible AI Supercomputer," 5 January 2026. https://nvidianews.nvidia.com/news/rubin-platform-ai-supercomputer

[12] NVIDIA Technical Blog, "NVIDIA Vera Rubin POD: Seven Chips, Five Rack-Scale Systems, One AI Supercomputer," 16 March 2026. https://developer.nvidia.com/blog/nvidia-vera-rubin-pod-seven-chips-five-rack-scale-systems-one-ai-supercomputer/

[13] Tom's Hardware, "Nvidia demonstrates Rubin Ultra tray, the world's first AI GPU with 1TB of HBM4E memory, new chips will slot into Kyber racks," 17 March 2026. https://www.tomshardware.com/pc-components/gpus/nvidia-demonstrates-rubin-ultra-tray-worlds-1st-ai-gpu-with-1tb-of-hbm4e

[14] SemiAnalysis, "Nvidia - GTC 2026" (Groq LP30, LPX Rack, Oberon and Kyber Updates, Nvidia's CPO Roadmap), 24 March 2026. https://newsletter.semianalysis.com/p/nvidia-the-inference-kingdom-expands

[15] NVIDIA Technical Blog, "Inside NVIDIA Rubin GPU Architecture: Powering the Era of Agentic AI." https://developer.nvidia.com/blog/inside-nvidia-rubin-gpu-architecture-powering-the-era-of-agentic-ai/

[16] TrendForce, "NVIDIA's Rubin Ultra Seen Sticking to Dual-Die Design on Packaging Constraints; TSMC 3nm Demand Intact," 1 April 2026. https://www.trendforce.com/news/2026/04/01/news-nvidias-rubin-ultra-seen-sticking-to-dual-die-design-on-packaging-constraints-tsmc-3nm-demand-intact/

[17] Tom's Hardware, "Nvidia reportedly cancels quad-die Rubin Ultra GPU in favor of dual-GPU design, report claims," 30 June 2026. https://www.tomshardware.com/tech-industry/artificial-intelligence/nvidia-reportedly-cancels-quad-die-rubin-ultra-gpu-in-favor-of-dual-gpu-design-report-claims-complex-design-purportedly-scrapped-over-manufacturing-execution-concerns

[18] Tom's Hardware, "Nvidia's Huang vows to deliver 'giant amounts' of Vera Rubin, company says that 'our roadmap is intact'," 15 July 2026. https://www.tomshardware.com/tech-industry/artificial-intelligence/nvidias-huang-vows-to-deliver-giant-amounts-of-vera-rubin-company-says-that-our-roadmap-is-intact

[19] TrendForce, "DRAM Supply to Remain Tight in 2027, Prompting NVIDIA to Lower HBM Configurations for Rubin Ultra, Says TrendForce," 4 August 2026. https://trendforce.com/presscenter/news/20260804-13166.html

[20] Tom's Hardware, "Nvidia reportedly testing lower memory configs of Rubin Ultra as memory shortage bites back," last updated 12 August 2026. https://www.tomshardware.com/pc-components/gpus/nvidia-reportedly-testing-lower-memory-configs-of-rubin-ultra-as-memory-shortage-bites-back-designs-tested-include-as-little-as-192-gb-and-step-back-to-hbm4

[21] SemiAnalysis (@SemiAnalysis_) on X, 2 September 2026. https://x.com/SemiAnalysis_/status/2094953611797151807

[22] NVIDIA, Q2 fiscal 2027 earnings call transcript, 26 August 2026. https://s201.q4cdn.com/141608511/files/content_files/TRANSCRIPT_-NVIDIA-Corp-NVDA-US-Q2-2027-Earnings-Call-26-August-2026-5_00-PM-ET.pdf

[23] NVIDIA Newsroom, "NVIDIA Unveils Rubin CPX: A New Class of GPU Designed for Massive-Context Inference," 9 September 2025. https://nvidianews.nvidia.com/news/nvidia-unveils-rubin-cpx-a-new-class-of-gpu-designed-for-massive-context-inference

[24] The Register, "AMD, Intel, Nvidia detail their datacenter roadmaps," 5 June 2024. https://www.theregister.com/2024/06/05/chipmakers_computex_roadmaps