> ## Content Index
> Fetch the complete content index at: https://www.insidedeeptech.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# NVIDIA Rubin vs Rubin Ultra vs Feynman: A Full Comparison Guide to the 2026, 2027 and 2028 GPUs, Racks and Limits
- URL: https://www.insidedeeptech.com/nvidia-rubin-vs-rubin-ultra-vs-feynman-full-guide/
- Published: 2026-11-05T07:17:34.000Z
- Updated: 2026-11-05T07:17:33.000Z
- Description: NVIDIA Rubin vs Rubin Ultra vs Feynman, compared from NVIDIA primary material: the 2026, 2027 and 2028 GPUs, HBM4 vs HBM4E vs custom HBM, NVLink 6, 7 and 8, Oberon vs Kyber racks, power, what is shipping versus only reported, who should buy what, and the honest limits.
- Author: Austin Heaton
- Tags: AI, Hardware, Deep Tech

NVIDIA now ships one generation, sells the next one on slides and names the one after that before either is finished. Vera Rubin began [production shipments in August 2026](https://s201.q4cdn.com/141608511/files/content%5Ffiles/TRANSCRIPT%5F-NVIDIA-Corp-NVDA-US-Q2-2027-Earnings-Call-26-August-2026-5%5F00-PM-ET.pdf?ref=insidedeeptech.com). Rubin Ultra is pencilled in for 2027\. Feynman sits in the 2028 column of the [GTC 2026 roadmap](https://s201.q4cdn.com/141608511/files/doc%5Fevents/2026/Mar/16/GTC-2026-Keynote.pdf?ref=insidedeeptech.com) with almost no numbers attached.

If you're planning AI capacity, the useful question isn't which chip is fastest. It's which column of that roadmap is real enough to plan a building around. This comparison lines up NVIDIA Rubin vs Rubin Ultra vs Feynman using NVIDIA's own keynotes, product pages and technical blogs, labels every report as a report, and builds on Inside Deep Tech's [NVIDIA Feynman full guide](https://www.insidedeeptech.com/nvidia-feynman-rosa-kyber-nvl1152-full-guide/) and [Rubin Ultra and Kyber full guide](https://www.insidedeeptech.com/nvidia-rubin-ultra-kyber-nvl576-full-guide/).

You'll get the generation-by-generation spec table, the GPU package, memory, NVLink and rack changes, the power and cooling jump, what each one means for buyers, and the honest limits. Last updated: November 3, 2026.

## Key Takeaways

- **Rubin is the only one you can buy.** Each Rubin GPU lists [50 petaflops of NVFP4 inference and 288 GB of HBM4](https://www.nvidia.com/en-us/data-center/vera-rubin-nvl72/?ref=insidedeeptech.com), and shipments started in August 2026.
- **Rubin Ultra's only official spec is from March 2025:** four dies, [100 petaflops of FP4 and 1 TB of HBM4E](https://s201.q4cdn.com/141608511/files/doc%5Fdownloads/2025/03/GTC2025%5FKeynote.pdf?ref=insidedeeptech.com). Reports since June 2026 say it was cut to two dies.
- **Feynman has a year and a parts list, not a spec.** The [GTC 2026 slide](https://s201.q4cdn.com/141608511/files/doc%5Fevents/2026/Mar/16/GTC-2026-Keynote.pdf?ref=insidedeeptech.com) shows 2028, die stacking, custom HBM, the Rosa CPU and NVLink 8 over co-packaged optics.
- **The rack is the real dividing line.** Rubin uses the copper Oberon NVL72; Kyber NVL144 arrives with Rubin Ultra and scales to an [all-to-all NVL1152 for Feynman](https://developer.nvidia.com/blog/nvidia-vera-rubin-pod-seven-chips-five-rack-scale-systems-one-ai-supercomputer/?ref=insidedeeptech.com).
- **Inside Deep Tech's take:** plan purchases on Rubin, plan buildings for Kyber, and treat every Feynman performance number you read in 2026 as an analyst model.

---

## The short answer: three generations, three levels of certainty

NVIDIA moved to a one-year cadence in 2025\. Its GTC 2025 roadmap slide put [Rubin in 2026, Rubin Ultra in 2027 and Feynman in 2028](https://s201.q4cdn.com/141608511/files/doc%5Fdownloads/2025/03/GTC2025%5FKeynote.pdf?ref=insidedeeptech.com), and the GTC 2026 version kept the same years while filling in more parts. That rhythm is the first thing to understand. Each step is a different kind of change.

- **Rubin (2026)** is a new GPU and a new Vera CPU in the same Oberon NVL72 rack style used by Blackwell. It's shipping.
- **Rubin Ultra (2027)** is the bigger-package refresh of Rubin, paired with the same Vera CPU and introducing the Kyber rack. Its spec is the least settled of the three.
- **Feynman (2028)** is the reset: new GPU, new Rosa CPU, new LPU, new DPU and NIC, and an NVLink fabric that leaves the rack on light.

🧭

The quick rule: Rubin is a product, Rubin Ultra is a moving target, and Feynman is a direction. Plan money against the first, floor space against the second, and fiber pathways against the third.

For the generation before all three, Inside Deep Tech's [Vera Rubin NVL72 vs GB300 vs GB200 comparison](https://www.insidedeeptech.com/vera-rubin-nvl72-vs-gb300-vs-gb200-full-guide/) covers how Rubin stacks up against the Blackwell racks most buyers run today. This guide starts where that one stops.

## NVIDIA Rubin vs Rubin Ultra vs Feynman at a glance

The table below uses only NVIDIA keynote slides, product pages and technical blogs. Where NVIDIA is silent, the cell says so, and reported figures are labeled.

| Item          | Rubin (Vera Rubin)            | Rubin Ultra                                      | Feynman                             |
| ------------- | ----------------------------- | ------------------------------------------------ | ----------------------------------- |
| Roadmap year  | 2026, shipping since Aug 2026 | 2027 (2H 2027 per GTC 2025)                      | 2028                                |
| GPU package   | Two reticle-limited dies      | Four dies (2025 spec); two-die redesign reported | Stacked GPU dies                    |
| Transistors   | 336 billion                   | Not disclosed                                    | Not disclosed                       |
| FP4 per GPU   | 50 PF NVFP4 inference         | 100 PF FP4 (2025 target)                         | Not disclosed                       |
| HBM per GPU   | 288 GB HBM4, 8 stacks         | 1 TB HBM4E, 16 stacks (2025 target)              | Custom HBM, capacity not disclosed  |
| Host CPU      | Vera, 88 Olympus cores        | Vera                                             | Rosa, Rigel cores                   |
| NVLink switch | NVLink 6, 3,600 GB/s          | NVLink 7, 3,600 GB/s                             | NVLink 8 CPO, no figure             |
| Scale-out     | Spectrum6 102T CPO, CX9 1600G | Spectrum6, CX9                                   | Spectrum7 204T CPO, CX10            |
| Racks         | Oberon NVL72                  | Oberon NVL72, Kyber NVL144, NVL576               | Oberon NVL72, Kyber NVL144, NVL1152 |
| LPU / DPU     | LP30 / BlueField-4            | LP35 / BlueField-4                               | LP40 / BlueField-5                  |

Source: [NVIDIA GTC 2025 keynote](https://s201.q4cdn.com/141608511/files/doc%5Fdownloads/2025/03/GTC2025%5FKeynote.pdf?ref=insidedeeptech.com) and [GTC 2026 keynote](https://s201.q4cdn.com/141608511/files/doc%5Fevents/2026/Mar/16/GTC-2026-Keynote.pdf?ref=insidedeeptech.com) roadmap slides; [Rubin GPU architecture blog](https://developer.nvidia.com/blog/inside-nvidia-rubin-gpu-architecture-powering-the-era-of-agentic-ai/?ref=insidedeeptech.com); [Vera Rubin NVL72 product page](https://www.nvidia.com/en-us/data-center/vera-rubin-nvl72/?ref=insidedeeptech.com); [NVIDIA Vera Rubin POD blog](https://developer.nvidia.com/blog/nvidia-vera-rubin-pod-seven-chips-five-rack-scale-systems-one-ai-supercomputer/?ref=insidedeeptech.com); [Rosa and Rigel announcement](https://blogs.nvidia.com/blog/nvidia-vera-max-single-threaded-cpu-at-scale/?ref=insidedeeptech.com); two-die redesign as [reported by Tom's Hardware citing SemiAnalysis](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?ref=insidedeeptech.com).

Two patterns jump out. First, the Feynman column is almost all words. Second, NVLink per-GPU bandwidth is flat at [3,600 GB/s from NVLink 6 to NVLink 7](https://s201.q4cdn.com/141608511/files/doc%5Fevents/2026/Mar/16/GTC-2026-Keynote.pdf?ref=insidedeeptech.com), so the 2027 and 2028 gains come from bigger domains and new media, not faster links.

## How sure is each column? Shipping, reported and slide-only

Spec tables flatten an important difference. A Rubin number is a datasheet value. A Rubin Ultra number is a 2025 target that may have changed. A Feynman entry is a word on a slide.

| Question             | Rubin                   | Rubin Ultra                               | Feynman                       |
| -------------------- | ----------------------- | ----------------------------------------- | ----------------------------- |
| Shipping?            | Yes, since Aug 2026     | No                                        | No                            |
| Latest official spec | Product page, 2026      | GTC 2025 slide                            | GTC 2026 slide, no numbers    |
| Reported changes     | None material           | Two dies, fewer HBM stacks, Kyber to 2028 | TSMC A16 node, 2H 2028 volume |
| NVIDIA response      | Datasheet published     | "Our roadmap is intact"                   | No comment on node            |
| OEM dates            | Shipping from many OEMs | Bull NVL72 in 1H 2028                     | None                          |

Source: [NVIDIA Q2 fiscal 2027 earnings call](https://s201.q4cdn.com/141608511/files/content%5Ffiles/TRANSCRIPT%5F-NVIDIA-Corp-NVDA-US-Q2-2027-Earnings-Call-26-August-2026-5%5F00-PM-ET.pdf?ref=insidedeeptech.com); [Tom's Hardware on the two-die report](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?ref=insidedeeptech.com); [CNBC on the Kyber delay report and NVIDIA's response](https://www.cnbc.com/2026/07/06/nvidia-kyber-rack-system-delays-manufacturing-taiwan-rubin-chips-.html?ref=insidedeeptech.com); [Bull Vera Rubin Ultra announcement](https://www.bull.com/en/product-news/nvidia-vera-rubin-ultra?ref=insidedeeptech.com); [TechPowerUp citing DigiTimes on A16](https://www.techpowerup.com/351607/nvidia-secures-tsmc-a16-node-for-next-generation-feynman-gpus?ref=insidedeeptech.com).

NVIDIA's August 2026 call is the clearest sign of where the company's attention sits. CFO Colette Kress said Vera Rubin would account for [about 20% of data center revenue in Q3](https://s201.q4cdn.com/141608511/files/content%5Ffiles/TRANSCRIPT%5F-NVIDIA-Corp-NVDA-US-Q2-2027-Earnings-Call-26-August-2026-5%5F00-PM-ET.pdf?ref=insidedeeptech.com) of fiscal 2027\. Rubin Ultra and Feynman weren't given dates on that call.

## GPU package: two dies, then four (or two), then stacked

Let's start with silicon. Rubin is built from two reticle-limited compute dies joined by NVIDIA's NV-HBI link, with [336 billion transistors, 224 SMs and 896 Tensor Cores](https://developer.nvidia.com/blog/inside-nvidia-rubin-gpu-architecture-powering-the-era-of-agentic-ai/?ref=insidedeeptech.com). That's the same side-by-side 2.5D recipe Blackwell used, scaled up.

Rubin Ultra was meant to double it. The GTC 2025 slide showed [four reticle-sized GPU dies per package](https://s201.q4cdn.com/141608511/files/doc%5Fdownloads/2025/03/GTC2025%5FKeynote.pdf?ref=insidedeeptech.com), and a GTC 2026 demo tray carried four compute chiplets with [1 TB of HBM4E](https://www.tomshardware.com/pc-components/gpus/nvidia-demonstrates-rubin-ultra-tray-worlds-1st-ai-gpu-with-1tb-of-hbm4e?ref=insidedeeptech.com). Then, on June 30, 2026, SemiAnalysis reported that NVIDIA [cancelled the quad-die design in favor of two dies](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?ref=insidedeeptech.com). NVIDIA hasn't confirmed it or published a replacement spec.

Feynman changes direction. Instead of adding dies sideways, the GTC 2026 roadmap lists [die stacking](https://s201.q4cdn.com/141608511/files/doc%5Fevents/2026/Mar/16/GTC-2026-Keynote.pdf?ref=insidedeeptech.com), and heise reported that the Feynman package on the slide looks [much smaller](https://www.heise.de/en/news/Nvidia-Feynman-comes-with-stacked-GPU-dies-and-custom-HBM-11214678.html?ref=insidedeeptech.com) than its predecessors. Stacking shortens wires and gets around the reticle limit. It also traps heat in the lower die.

🧱

Why it matters: Rubin Ultra is the end of the 'more dies on the interposer' road, and Feynman is the start of the 'stack them' road. The second road has never carried a kilowatt-class AI GPU in volume.

## Memory: HBM4 vs HBM4E vs custom HBM

Memory is where the three generations differ most on paper. Rubin uses 12-Hi HBM4 stacks for up to [288 GB and 22 TB/s per GPU](https://developer.nvidia.com/blog/inside-nvidia-rubin-gpu-architecture-powering-the-era-of-agentic-ai/?ref=insidedeeptech.com), which NVIDIA calls a 2.8x bandwidth increase over Blackwell. The Vera Rubin NVL72 spec table lists [288 GB at 19.2 TB/s per GPU and 20.7 TB per rack](https://www.nvidia.com/en-us/data-center/vera-rubin-nvl72/?ref=insidedeeptech.com), so check which page a vendor is quoting.

Rubin Ultra's official target is [16 HBM4E stacks and 1 TB per package](https://s201.q4cdn.com/141608511/files/doc%5Fdownloads/2025/03/GTC2025%5FKeynote.pdf?ref=insidedeeptech.com), more than three times Rubin. That figure is now the most disputed number on the roadmap. Tom's Hardware, citing The Information, reported NVIDIA was testing Rubin Ultra configurations with [as little as 192 GB, some using HBM4 instead of HBM4E](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?ref=insidedeeptech.com). TrendForce expects HBM bit shipments to grow [50% to 60% in 2027, still short of demand](https://www.trendforce.com/presscenter/news/20260804-13166.html?ref=insidedeeptech.com).

Feynman moves the memory controller. The GTC 2026 slide says [custom HBM](https://s201.q4cdn.com/141608511/files/doc%5Fevents/2026/Mar/16/GTC-2026-Keynote.pdf?ref=insidedeeptech.com), and in August 2026 NVIDIA described NVHBM, which puts its own controller in the HBM base die. NVIDIA claims up to [30% more memory bandwidth and 15% lower HBM power than standard HBM4E](https://blogs.nvidia.com/blog/nvlink-fusion-nvhbm-custom-high-bandwidth-memory/?ref=insidedeeptech.com). It didn't name Feynman in that announcement, only future GPUs.

| Memory item         | Rubin    | Rubin Ultra                                         | Feynman                               |
| ------------------- | -------- | --------------------------------------------------- | ------------------------------------- |
| HBM generation      | HBM4     | HBM4E (target)                                      | Custom HBM (generation not disclosed) |
| Stacks per GPU      | 8        | 16 (target); 8 reported                             | Not disclosed                         |
| Capacity per GPU    | 288 GB   | 1 TB (target); as little as 192 GB tested, reported | Not disclosed                         |
| Controller location | GPU die  | GPU die                                             | HBM base die (NVHBM approach)         |
| Status              | Shipping | Under evaluation, reported                          | Roadmap                               |

Source: [Rubin GPU architecture blog](https://developer.nvidia.com/blog/inside-nvidia-rubin-gpu-architecture-powering-the-era-of-agentic-ai/?ref=insidedeeptech.com); [GTC 2025 keynote](https://s201.q4cdn.com/141608511/files/doc%5Fdownloads/2025/03/GTC2025%5FKeynote.pdf?ref=insidedeeptech.com); [GTC 2026 keynote](https://s201.q4cdn.com/141608511/files/doc%5Fevents/2026/Mar/16/GTC-2026-Keynote.pdf?ref=insidedeeptech.com); [Tom's Hardware on reduced-memory tests](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?ref=insidedeeptech.com); [NVIDIA NVHBM blog](https://blogs.nvidia.com/blog/nvlink-fusion-nvhbm-custom-high-bandwidth-memory/?ref=insidedeeptech.com).

## Scale-up: NVL72 to NVL576 to NVL1152

The bigger story is how many GPUs share one NVLink domain. Rubin's Oberon rack ties [72 Rubin GPUs and 36 Vera CPUs](https://developer.nvidia.com/blog/nvidia-vera-rubin-pod-seven-chips-five-rack-scale-systems-one-ai-supercomputer/?ref=insidedeeptech.com) together over a copper spine, with NVLink 6 at [3.6 TB/s per GPU and 260 TB/s per rack](https://developer.nvidia.com/blog/nvidia-vera-rubin-pod-seven-chips-five-rack-scale-systems-one-ai-supercomputer/?ref=insidedeeptech.com).

Rubin Ultra adds two new options on top of NVL72\. Kyber NVL144 doubles the domain to 144 GPUs in one vertical-blade rack. Vera Rubin Ultra NVL576 joins [eight MGX racks of 72 Rubin Ultra GPUs into one 576-GPU NVLink domain](https://developer.nvidia.com/blog/nvidia-vera-rubin-pod-seven-chips-five-rack-scale-systems-one-ai-supercomputer/?ref=insidedeeptech.com) over copper and direct optical links. In 2025, 'NVL576' meant a single Kyber rack counted by dies, so be careful with older articles.

Feynman takes Kyber further. NVIDIA's technical blog says Kyber will scale to an [all-to-all NVL1152 with direct optical rack-to-rack links](https://developer.nvidia.com/blog/nvidia-vera-rubin-pod-seven-chips-five-rack-scale-systems-one-ai-supercomputer/?ref=insidedeeptech.com) and that it 'provides the foundation for the next era of extreme scale-up AI computing using NVIDIA Feynman.' Huang said the Feynman generation connects through [Kyber for both copper and co-packaged optics scale-up](https://blogs.nvidia.com/blog/gtc-2026-news/?ref=insidedeeptech.com).

| Domain                  | GPUs in one NVLink domain | Racks    | Medium                     | Generation           |
| ----------------------- | ------------------------- | -------- | -------------------------- | -------------------- |
| Vera Rubin NVL72        | 72                        | 1 Oberon | Copper spine               | Rubin                |
| Vera Rubin Ultra NVL72  | 72                        | 1 MGX    | Copper                     | Rubin Ultra          |
| Kyber NVL144            | 144                       | 1 Kyber  | Copper midplane            | Rubin Ultra, Feynman |
| Vera Rubin Ultra NVL576 | 576                       | 8 MGX    | Copper plus direct optical | Rubin Ultra          |
| Kyber NVL1152           | 1,152                     | 8 Kyber  | Direct optical, all-to-all | Feynman              |

Source: [NVIDIA Vera Rubin POD technical blog](https://developer.nvidia.com/blog/nvidia-vera-rubin-pod-seven-chips-five-rack-scale-systems-one-ai-supercomputer/?ref=insidedeeptech.com); [GTC 2026 roadmap slide](https://s201.q4cdn.com/141608511/files/doc%5Fevents/2026/Mar/16/GTC-2026-Keynote.pdf?ref=insidedeeptech.com).

Here's why that matters. A whole Vera Rubin POD today already holds [1,152 Rubin GPUs across 40 racks](https://developer.nvidia.com/blog/nvidia-vera-rubin-pod-seven-chips-five-rack-scale-systems-one-ai-supercomputer/?ref=insidedeeptech.com), but as separate 72-GPU domains. Feynman's pitch is to make that same count behave like one machine. Inside Deep Tech's [Feynman full guide](https://www.insidedeeptech.com/nvidia-feynman-rosa-kyber-nvl1152-full-guide/) covers NVLink 8 CPO and the optical supply chain in detail.

## Racks, power and cooling: Oberon vs Kyber

Rubin runs in a familiar envelope. The Vera Rubin NVL72 is built on the third-generation MGX NVL72 rack and cooled with [45°C inlet liquid](https://www.nvidia.com/en-us/data-center/vera-rubin-nvl72/?ref=insidedeeptech.com). It's a big rack, but data centers built for GB200 and GB300 can plan for it.

Kyber is a different building. At GTC 2025, Huang described the Rubin Ultra rack as ['600kW, 2.5 million parts'](https://www.datacenterdynamics.com/en/news/nvidias-rubin-ultra-nvl576-rack-expected-to-be-600kw-coming-second-half-of-2027/?ref=insidedeeptech.com), and NVIDIA says [800 VDC power distribution will reach full-scale production alongside Kyber](https://developer.nvidia.com/blog/nvidia-800-v-hvdc-architecture-will-power-the-next-generation-of-ai-factories/?ref=insidedeeptech.com), because 54 V busbars can't carry megawatt-class loads efficiently.

There's a twist. Bull says its Vera Rubin Ultra NVL72 [keeps the rack footprint and maximum rack power of Vera Rubin NVL72](https://www.bull.com/en/product-news/nvidia-vera-rubin-ultra?ref=insidedeeptech.com) and supports both AC and 800 VDC, with availability planned for the first half of 2028\. So the first Rubin Ultra many buyers see may not need Kyber power at all.

Feynman inherits Kyber. NVIDIA hasn't published a Feynman TDP or rack power, but heise estimates a stacked Feynman GPU could exceed [2,000 watts](https://www.heise.de/en/news/Nvidia-Feynman-comes-with-stacked-GPU-dies-and-custom-HBM-11214678.html?ref=insidedeeptech.com). That's an outside estimate, not an NVIDIA figure. If it's close, cooling moves from the rack level down to the package.

⚠️

Watch this: SemiAnalysis blamed the reported Kyber slip on the PCB midplane, which [remains challenging to manufacture](https://www.cnbc.com/2026/07/06/nvidia-kyber-rack-system-delays-manufacturing-taiwan-rubin-chips-.html?ref=insidedeeptech.com). Every Rubin Ultra Kyber date and every Feynman NVL1152 date runs through that one board.

## CPU, LPU and networking: what changes around the GPU

Rubin and Rubin Ultra share the Vera CPU. NVIDIA says Vera's Olympus core delivers [50% higher instructions per cycle than Grace](https://blogs.nvidia.com/blog/nvidia-vera-max-single-threaded-cpu-at-scale/?ref=insidedeeptech.com), and each Vera Rubin superchip pairs one Vera with two Rubin GPUs and up to [1.5 TB of LPDDR5X](https://www.nvidia.com/en-us/data-center/vera-rubin-nvl72/?ref=insidedeeptech.com).

Feynman brings the two-year CPU reset. Rosa, [named for Rosalind Franklin](https://blogs.nvidia.com/blog/gtc-2026-news/?ref=insidedeeptech.com), uses the Rigel Arm core, which NVIDIA says offers [higher per-core performance than Olympus in the same silicon footprint](https://blogs.nvidia.com/blog/nvidia-vera-max-single-threaded-cpu-at-scale/?ref=insidedeeptech.com). NVIDIA hasn't given a Rosa core count.

- **LPU:** the roadmap runs LP30 with Rubin, LP35 with Rubin Ultra and [LP40 with Feynman](https://s201.q4cdn.com/141608511/files/doc%5Fevents/2026/Mar/16/GTC-2026-Keynote.pdf?ref=insidedeeptech.com), the Groq-derived low-latency inference line.
- **DPU and NIC:** BlueField-4 and ConnectX-9 serve both Rubin generations; Feynman moves to [BlueField-5 and CX10](https://blogs.nvidia.com/blog/gtc-2026-news/?ref=insidedeeptech.com).
- **Scale-out:** Spectrum6 at 102T with CPO serves Rubin, and Feynman moves to [Spectrum7 at 204T with CPO](https://s201.q4cdn.com/141608511/files/doc%5Fevents/2026/Mar/16/GTC-2026-Keynote.pdf?ref=insidedeeptech.com).

One Rubin-generation part dropped off the roadmap entirely: Rubin CPX, the GDDR7 context GPU. Inside Deep Tech's [Rubin CPX full guide](https://www.insidedeeptech.com/nvidia-rubin-cpx-context-gpu-full-guide/) explains why it was pulled and what replaced it in the rack plan.

## Performance: what NVIDIA has published and what it hasn't

NVIDIA's performance numbers come in three grades. Rubin has datasheet values. Rubin Ultra has 2025 targets. Feynman has none.

| Metric                 | Vera Rubin NVL72         | Rubin Ultra NVL576 (2025 target) | Feynman       |
| ---------------------- | ------------------------ | -------------------------------- | ------------- |
| FP4 inference          | 3.6 EF NVFP4             | 15 EF                            | Not disclosed |
| FP8 training           | 1.26 EF                  | 5 EF                             | Not disclosed |
| vs GB300 NVL72         | 3.3x (2025 slide)        | 14x                              | Not disclosed |
| Fast memory            | Up to 75 TB (2025 slide) | 365 TB                           | Not disclosed |
| HBM bandwidth per rack | 1.6 PB/s (GTC 2026)      | 4.6 PB/s                         | Not disclosed |

Source: [Vera Rubin NVL72 product page](https://www.nvidia.com/en-us/data-center/vera-rubin-nvl72/?ref=insidedeeptech.com); [GTC 2025 keynote](https://s201.q4cdn.com/141608511/files/doc%5Fdownloads/2025/03/GTC2025%5FKeynote.pdf?ref=insidedeeptech.com); [GTC 2026 keynote](https://s201.q4cdn.com/141608511/files/doc%5Fevents/2026/Mar/16/GTC-2026-Keynote.pdf?ref=insidedeeptech.com). The Rubin Ultra column counts 144 four-die packages, the 2025 meaning of NVL576.

Read the multipliers carefully. The 14x figure compares a 2025 Rubin Ultra target on NVIDIA-chosen workloads with a shipping GB300 rack. If the two-die report is right, per-package FP4 could land closer to Rubin's [50 petaflops](https://www.nvidia.com/en-us/data-center/vera-rubin-nvl72/?ref=insidedeeptech.com) times two than to the original 100, though NVIDIA hasn't said so.

NVIDIA also publishes system-level claims for Rubin that are easier to use. It says Vera Rubin NVL72 trains MoE models with [one-fourth the GPUs and runs inference at one-tenth the cost per million tokens](https://www.nvidia.com/en-us/data-center/vera-rubin-nvl72/?ref=insidedeeptech.com) of GB200 NVL72\. Those are vendor benchmarks, but they're measured on a product that exists.

## The money: cost per gigawatt keeps rising

NVIDIA frames each generation by its revenue opportunity per gigawatt of data center capacity. On the August 2026 call, Huang put it at about [$18 billion for Hopper, $25 billion for Grace Blackwell and $40 billion for Vera Rubin](https://s201.q4cdn.com/141608511/files/content%5Ffiles/TRANSCRIPT%5F-NVIDIA-Corp-NVDA-US-Q2-2027-Earnings-Call-26-August-2026-5%5F00-PM-ET.pdf?ref=insidedeeptech.com), and said 'after that, it's going to be higher.'

That's the honest pricing signal for Rubin Ultra and Feynman. Each generation is meant to do more work per watt and cost more per megawatt. NVIDIA also said it expects supply to [remain a bottleneck at least through the end of fiscal 2028](https://s201.q4cdn.com/141608511/files/content%5Ffiles/TRANSCRIPT%5F-NVIDIA-Corp-NVDA-US-Q2-2027-Earnings-Call-26-August-2026-5%5F00-PM-ET.pdf?ref=insidedeeptech.com), which ends in January 2028.

The supply chain is being built for the later columns. In October 2026 NVIDIA announced partner deals including a [more than 300% increase in Amkor interposer capacity by the end of 2028](https://nvidianews.nvidia.com/news/nvidia-and-supply-chain-partners-expand-capacity-to-accelerate-ai-infrastructure-buildout?ref=insidedeeptech.com) and a 15x increase in SPIL co-packaged optics engine capacity by mid-2027\. NVIDIA didn't tie them to a named generation, but the dates line up with Rubin Ultra and Feynman.

## Which generation should you plan for?

### If you need capacity in 2026 or 2027

Buy Rubin. It's the only one with a datasheet and a ship date, and waiting doesn't get you a better slot when NVIDIA itself expects supply to stay tight. Inside Deep Tech's [Vera Rubin NVL72 full guide](https://www.insidedeeptech.com/nvidia-vera-rubin-nvl72-full-guide/) covers the rack, HBM4 suppliers and the ramp.

### If you're building a new hall for 2028

Design for Kyber even if you buy Oberon first. That means 800 VDC distribution, dense liquid cooling and fiber pathways between racks. NVIDIA tied [800 VDC production to Kyber](https://developer.nvidia.com/blog/nvidia-800-v-hvdc-architecture-will-power-the-next-generation-of-ai-factories/?ref=insidedeeptech.com), and Feynman's NVL1152 assumes all three.

### If you're waiting for Feynman

Don't, unless your workload truly needs a scale-up domain far beyond 72 GPUs and you can absorb a late-2028 date. TechPowerUp reported Feynman [mass production in the second half of 2028](https://www.techpowerup.com/351607/nvidia-secures-tsmc-a16-node-for-next-generation-feynman-gpus?ref=insidedeeptech.com), citing DigiTimes, and NVIDIA hasn't published a single Feynman performance figure.

## Inside Deep Tech's take

📌

Inside Deep Tech's take: Rubin Ultra is the risky column, not Feynman. Feynman's lack of numbers is normal for a 2028 part. Rubin Ultra's problem is that its only official spec is 19 months old while credible reports describe a smaller chip on a later rack. Until NVIDIA restates it, assume the Rubin Ultra you can buy in 2027 looks like a better Rubin NVL72, and the Kyber era really starts in 2028.

That view rests on public reporting that NVIDIA hasn't confirmed, and NVIDIA has said its roadmap is intact. If GTC 2027 restates four dies and 1 TB with a 2027 Kyber ship date, this take is wrong, and buyers waiting for Rubin Ultra will have been right.

## Honest limits of this comparison

- **Stale official specs.** The newest NVIDIA Rubin Ultra numbers date from [March 2025](https://s201.q4cdn.com/141608511/files/doc%5Fdownloads/2025/03/GTC2025%5FKeynote.pdf?ref=insidedeeptech.com). Comparing them with shipping Rubin numbers mixes a target with a datasheet.
- **NVIDIA's own pages differ.** The Rubin architecture blog says [22 TB/s per GPU](https://developer.nvidia.com/blog/inside-nvidia-rubin-gpu-architecture-powering-the-era-of-agentic-ai/?ref=insidedeeptech.com), while the product spec table lists [19.2 TB/s per GPU and 216 TB/s of NVLink per rack](https://www.nvidia.com/en-us/data-center/vera-rubin-nvl72/?ref=insidedeeptech.com).
- **Reports aren't confirmations.** The two-die Rubin Ultra, reduced HBM and [Kyber slip to 2028](https://www.tomshardware.com/pc-components/gpus/nvidias-kyber-rack-for-rubin-ultra-slips-to-2028?ref=insidedeeptech.com) all come from analysts or press, not NVIDIA.
- **Feynman is unquantified.** No FLOPS, memory capacity, power or node are confirmed; [TSMC A16 is reported](https://www.techpowerup.com/351607/nvidia-secures-tsmc-a16-node-for-next-generation-feynman-gpus?ref=insidedeeptech.com), not announced.
- **Power is unpublished.** NVIDIA hasn't restated a Kyber rack power for the 2026 design or given any Feynman TDP.
- **Vendor benchmarks.** Every multiplier in this guide comes from NVIDIA-chosen workloads, not independent tests.

## What to watch next

- **NVIDIA's Q3 fiscal 2027 results** in November: Vera Rubin's share of data center revenue against the [about 20% guide](https://s201.q4cdn.com/141608511/files/content%5Ffiles/TRANSCRIPT%5F-NVIDIA-Corp-NVDA-US-Q2-2027-Earnings-Call-26-August-2026-5%5F00-PM-ET.pdf?ref=insidedeeptech.com), and any Rubin Ultra date.
- **GTC Washington, D.C.**, [November 30 to December 3, 2026](https://www.nvidia.com/gtc/dc/?ref=insidedeeptech.com), with Huang's keynote: any roadmap slide at all.
- **GTC 2027 in San Jose**, where NVIDIA has updated the roadmap in [2025](https://s201.q4cdn.com/141608511/files/doc%5Fdownloads/2025/03/GTC2025%5FKeynote.pdf?ref=insidedeeptech.com) and [2026](https://s201.q4cdn.com/141608511/files/doc%5Fevents/2026/Mar/16/GTC-2026-Keynote.pdf?ref=insidedeeptech.com): a restated Rubin Ultra spec and the first Feynman numbers.
- **Bull and other OEM dates** for Vera Rubin Ultra NVL72, starting with Bull's [first-half 2028 plan](https://www.bull.com/en/product-news/nvidia-vera-rubin-ultra?ref=insidedeeptech.com).
- **Kyber midplane and CPO capacity:** SPIL's optical-engine ramp and Amkor's interposer build in the [October 2026 supply deals](https://nvidianews.nvidia.com/news/nvidia-and-supply-chain-partners-expand-capacity-to-accelerate-ai-infrastructure-buildout?ref=insidedeeptech.com).

[Read the NVIDIA Feynman Full Guide](https://www.insidedeeptech.com/nvidia-feynman-rosa-kyber-nvl1152-full-guide/)

[NVIDIA Feynman: A Full Guide to the 2028 GPU, Rosa CPU, Kyber NVL1152, Custom HBM and LimitsWhat NVIDIA has confirmed about Feynman, what is only reported, the supply chain behind it and the honest limits.![](https://www.insidedeeptech.com/favicon.ico)Inside Deep Tech](https://www.insidedeeptech.com/nvidia-feynman-rosa-kyber-nvl1152-full-guide/)

---

## FAQ

#### What is the difference between NVIDIA Rubin and Rubin Ultra?

Rubin is the 2026 GPU with two dies and [288 GB of HBM4](https://www.nvidia.com/en-us/data-center/vera-rubin-nvl72/?ref=insidedeeptech.com), shipping in NVL72 racks. Rubin Ultra is the 2027 step, officially four dies and [1 TB of HBM4E](https://s201.q4cdn.com/141608511/files/doc%5Fdownloads/2025/03/GTC2025%5FKeynote.pdf?ref=insidedeeptech.com), and the first GPU for the Kyber rack.

#### Is Feynman the next GPU after Rubin Ultra?

Yes. NVIDIA's [GTC 2026 roadmap](https://s201.q4cdn.com/141608511/files/doc%5Fevents/2026/Mar/16/GTC-2026-Keynote.pdf?ref=insidedeeptech.com) puts Feynman in 2028, after Rubin in 2026 and Rubin Ultra in 2027\. It brings stacked GPU dies, custom HBM and the Rosa CPU.

#### Is Rubin Ultra delayed?

NVIDIA says its roadmap is intact. SemiAnalysis reported the Kyber NVL144 rack [slipped to 2028](https://www.cnbc.com/2026/07/06/nvidia-kyber-rack-system-delays-manufacturing-taiwan-rubin-chips-.html?ref=insidedeeptech.com), and Bull plans its Vera Rubin Ultra NVL72 for the [first half of 2028](https://www.bull.com/en/product-news/nvidia-vera-rubin-ultra?ref=insidedeeptech.com).

#### How fast is Feynman compared with Rubin?

Nobody outside NVIDIA knows. NVIDIA hasn't published any Feynman FLOPS, memory or power figures, so any multiplier you see is an estimate. The [Feynman full guide](https://www.insidedeeptech.com/nvidia-feynman-rosa-kyber-nvl1152-full-guide/) tracks what is confirmed.

#### Should buyers choose Vera Rubin now or wait for Rubin Ultra?

Most buyers should buy Vera Rubin now. It's shipping, while Rubin Ultra's final spec is unconfirmed and NVIDIA expects supply to stay tight [through fiscal 2028](https://s201.q4cdn.com/141608511/files/content%5Ffiles/TRANSCRIPT%5F-NVIDIA-Corp-NVDA-US-Q2-2027-Earnings-Call-26-August-2026-5%5F00-PM-ET.pdf?ref=insidedeeptech.com).

#### What rack does each generation use?

Rubin uses the Oberon NVL72\. Rubin Ultra adds Kyber NVL144 and an eight-rack NVL576, and Feynman adds an [all-to-all Kyber NVL1152](https://developer.nvidia.com/blog/nvidia-vera-rubin-pod-seven-chips-five-rack-scale-systems-one-ai-supercomputer/?ref=insidedeeptech.com) over optical links.

## The bottom line

Rubin vs Rubin Ultra vs Feynman isn't really a speed contest yet. It's a certainty ladder: one shipping product, one 2025 target under revision, and one 2028 architecture defined by ideas rather than numbers. The safest move is to buy Rubin, design new halls for Kyber's power, cooling and fiber, and wait for GTC 2027 before pricing anything later.

For the deepest look at the far end of that ladder, read Inside Deep Tech's [NVIDIA Feynman full guide](https://www.insidedeeptech.com/nvidia-feynman-rosa-kyber-nvl1152-full-guide/).