> ## 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 Ultra and Kyber: A Full Guide to NVL144, NVL576 and the Next AI Rack - URL: https://www.insidedeeptech.com/nvidia-rubin-ultra-kyber-nvl576-full-guide/ - Published: 2026-10-22T06:03:26.000Z - Updated: 2026-10-22T06:03:25.000Z - Description: NVIDIA Rubin Ultra and Kyber full guide: the GTC 2025 and GTC 2026 specs, why NVL576 changed meaning, the vertical-blade Kyber rack, 800 VDC power, liquid cooling, HBM4E, the reported delays and what it means for buyers. - Author: Austin Heaton - Tags: AI, Hardware, Deep Tech In 19 months, Rubin Ultra has been described three different ways. At GTC 2025 it was a single [600 kW Kyber rack with 576 GPUs](https://www.datacenterdynamics.com/en/news/nvidias-rubin-ultra-nvl576-rack-expected-to-be-600kw-coming-second-half-of-2027/?ref=insidedeeptech.com). At GTC 2026 the same rack became [Kyber NVL144](https://developer.nvidia.com/blog/nvidia-vera-rubin-pod-seven-chips-five-rack-scale-systems-one-ai-supercomputer/?ref=insidedeeptech.com), and "NVL576" now meant eight racks wired together. Since June 2026, analyst reports say the [four-die GPU was cut to 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) and the [Kyber rack slipped to 2028](https://www.tomshardware.com/pc-components/gpus/nvidias-kyber-rack-for-rubin-ultra-slips-to-2028?ref=insidedeeptech.com), while NVIDIA says its roadmap is intact. That's why "what comes after Vera Rubin" is a harder question than it looks. This full guide separates what NVIDIA has officially announced from what has only been reported, then covers the Kyber rack, 800 VDC power, liquid cooling, HBM4E, the timeline and the honest limits, all as of October 20, 2026. If you're still choosing between today's racks, start with Inside Deep Tech's [Vera Rubin NVL72 vs GB300 vs GB200 NVL72 comparison](https://www.insidedeeptech.com/vera-rubin-nvl72-vs-gb300-vs-gb200-full-guide/). This guide picks up where that one stops: the generation after it. Last updated: October 20, 2026. ## Key Takeaways - **Rubin Ultra** is NVIDIA's 2027 GPU after Rubin; the GTC 2025 plan was [four reticle-sized dies, 100 PF of FP4 and 1 TB of HBM4E](https://s201.q4cdn.com/141608511/files/doc%5Fdownloads/2025/03/GTC2025%5FKeynote.pdf?ref=insidedeeptech.com). - **Kyber** is the vertical-blade rack after Oberon. NVIDIA now calls it [Kyber NVL144, with 144 GPUs per rack](https://developer.nvidia.com/blog/nvidia-vera-rubin-pod-seven-chips-five-rack-scale-systems-one-ai-supercomputer/?ref=insidedeeptech.com). - "NVL576" now means [eight 72-GPU racks in one NVLink domain](https://developer.nvidia.com/blog/nvidia-vera-rubin-pod-seven-chips-five-rack-scale-systems-one-ai-supercomputer/?ref=insidedeeptech.com), not one rack. - Kyber needs [800 VDC power](https://developer.nvidia.com/blog/nvidia-800-v-hvdc-architecture-will-power-the-next-generation-of-ai-factories/?ref=insidedeeptech.com) and full liquid cooling; NVIDIA once quoted [600 kW per rack](https://www.datacenterdynamics.com/en/news/nvidias-rubin-ultra-nvl576-rack-expected-to-be-600kw-coming-second-half-of-2027/?ref=insidedeeptech.com). - Honest limit: reports of a [two-die redesign](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), [HBM cuts](https://www.trendforce.com/presscenter/news/20260804-13166.html?ref=insidedeeptech.com) and a [2028 Kyber slip](https://www.tomshardware.com/pc-components/gpus/nvidias-kyber-rack-for-rubin-ultra-slips-to-2028?ref=insidedeeptech.com) remain unconfirmed by NVIDIA. 💡 Inside Deep Tech's take: Rubin Ultra will ship, but the version that ships in 2027 is likely to be smaller and more conservative than the one shown on stage in 2025\. Plan facilities for Kyber-class power and cooling, but budget compute around an NVL72-class rack until NVIDIA confirms otherwise. ## What NVIDIA Rubin Ultra actually is Rubin Ultra is the "Ultra" step of the Rubin generation, the way Blackwell Ultra (B300) followed Blackwell. NVIDIA detailed it in the GTC 2025 keynote, slotting it in for the [second half of 2027, one year after Vera Rubin](https://blogs.nvidia.com/blog/nvidia-keynote-at-gtc-2025-ai-news-live-updates/?ref=insidedeeptech.com). The official GTC 2025 slide is still the most detailed public spec. It lists each Rubin Ultra package as [four reticle-sized GPU dies with 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), paired with the same 88-core Vera CPU used in Vera Rubin. At GTC 2026, NVIDIA showed a real compute tray. Tom's Hardware reported it carried [four compute chiplets and 1 TB of HBM4E, the first AI accelerator with a terabyte of memory](https://www.tomshardware.com/pc-components/gpus/nvidia-demonstrates-rubin-ultra-tray-worlds-1st-ai-gpu-with-1tb-of-hbm4e?ref=insidedeeptech.com), and heise counted [four Rubin Ultra GPUs and two Vera CPUs per board](https://www.heise.de/en/news/Kyber-Rack-Nvidia-Shows-Rubin-Ultra-Boards-11213103.html?ref=insidedeeptech.com). So, officially, Rubin Ultra is a quad-die GPU with HBM4E that lives in a new rack. Everything beyond that is either a roadmap target or a report. ## Why NVL576 changed meaning between GTC 2025 and GTC 2026 Here's the problem. NVIDIA changed how it counts GPUs. In 2025 the "NVL" number counted GPU dies, so a rack of 144 four-die packages was called NVL576\. heise explains that the [GTC 2025 concept had 144 Rubin Ultra GPUs and 72 Vera CPUs](https://www.heise.de/en/news/Kyber-Rack-Nvidia-Shows-Rubin-Ultra-Boards-11213103.html?ref=insidedeeptech.com), and the 576 figure came from 144 times four dies. By GTC 2026, NVIDIA counted packages instead. The technical blog for the Vera Rubin platform now describes [Kyber as a rack that doubles the NVLink domain to 144 GPUs, first shipping with Vera Rubin Ultra as a standalone NVL144](https://developer.nvidia.com/blog/nvidia-vera-rubin-pod-seven-chips-five-rack-scale-systems-one-ai-supercomputer/?ref=insidedeeptech.com). Same idea, new label. Then NVIDIA reused the old name for something bigger. Vera Rubin Ultra NVL576 is now [eight separate MGX racks, each with 72 Rubin Ultra GPUs, joined into one 576-GPU NVLink domain with copper and direct optical links](https://developer.nvidia.com/blog/nvidia-vera-rubin-pod-seven-chips-five-rack-scale-systems-one-ai-supercomputer/?ref=insidedeeptech.com). NVIDIA even built a GB200-based prototype of that multirack topology, called Polyphe. That gives Rubin Ultra three official scale-up options, in NVIDIA's own words: [NVL72, NVL144 and the flagship NVL576](https://developer.nvidia.com/blog/nvidia-vera-rubin-pod-seven-chips-five-rack-scale-systems-one-ai-supercomputer/?ref=insidedeeptech.com). | Configuration | First shown | GPUs in one NVLink domain | Racks | Status, October 2026 | | -------------------------------------- | ----------- | ------------------------- | ------------- | ---------------------------------------- | | Rubin Ultra NVL576 (2025 meaning) | GTC 2025 | 576 dies (144 packages) | 1 Kyber rack | Renamed Kyber NVL144 | | Vera Rubin Ultra NVL72 | GTC 2026 | 72 packages | 1 MGX rack | Official; Bull plans 1H 2028 | | Kyber NVL144 | GTC 2026 | 144 packages | 1 Kyber rack | Official; 2028 slip reported | | Vera Rubin Ultra NVL576 (2026 meaning) | GTC 2026 | 576 packages | 8 MGX racks | Official; may slip or ship in low volume | | Kyber NVL1152 | GTC 2026 | 1,152 GPUs | 8 Kyber racks | Feynman generation | Source: [NVIDIA technical blog, March 2026](https://developer.nvidia.com/blog/nvidia-vera-rubin-pod-seven-chips-five-rack-scale-systems-one-ai-supercomputer/?ref=insidedeeptech.com); [heise](https://www.heise.de/en/news/Kyber-Rack-Nvidia-Shows-Rubin-Ultra-Boards-11213103.html?ref=insidedeeptech.com); [Bull, October 2026](https://www.bull.com/en/product-news/nvidia-vera-rubin-ultra?ref=insidedeeptech.com); [Tom's Hardware on the reported delays](https://www.tomshardware.com/pc-components/gpus/nvidias-kyber-rack-for-rubin-ultra-slips-to-2028?ref=insidedeeptech.com). ⚠️ When you read a Rubin Ultra spec, check the date. A 2025 figure for "NVL576" describes one rack. A 2026 figure for "NVL576" describes eight. Mixing them is the most common error in Rubin Ultra coverage. ## The GTC 2025 targets: what Rubin Ultra promised The 2025 keynote put hard numbers on the Kyber rack. Rubin Ultra NVL576 was billed at [15 exaflops of FP4 inference, 5 exaflops of FP8 training and 14x the GB300 NVL72](https://s201.q4cdn.com/141608511/files/doc%5Fdownloads/2025/03/GTC2025%5FKeynote.pdf?ref=insidedeeptech.com), with 365 TB of fast memory and 4.6 PB/s of HBM4E bandwidth. For scale, the same slide gave Vera Rubin NVL144 (now called Vera Rubin NVL72) [3.6 exaflops of FP4 inference, or 3.3x GB300 NVL72](https://s201.q4cdn.com/141608511/files/doc%5Fdownloads/2025/03/GTC2025%5FKeynote.pdf?ref=insidedeeptech.com). Rubin Ultra was meant to be roughly four times that in one rack. | Metric (NVIDIA GTC 2025 targets) | GB300 NVL72 | Vera Rubin NVL144 (now NVL72) | Rubin Ultra NVL576 (now Kyber NVL144) | | -------------------------------- | ------------ | ----------------------------- | ------------------------------------- | | Timing | 2H 2025 | 2H 2026 | 2H 2027 | | FP4 inference | 1.1 EF dense | 3.6 EF | 15 EF | | FP8 training | 0.36 EF | 1.2 EF | 5 EF | | vs GB300 NVL72 | 1x | 3.3x | 14x | | Fast memory | 40 TB | 75 TB | 365 TB | | Scale-up fabric | NVLink 5 | NVLink 6, 260 TB/s | NVLink 7, 1.5 PB/s | | Rack | Oberon | Oberon | Kyber | Source: [NVIDIA GTC 2025 keynote slides](https://s201.q4cdn.com/141608511/files/doc%5Fdownloads/2025/03/GTC2025%5FKeynote.pdf?ref=insidedeeptech.com). These were roadmap targets, not measured results, and several have since been reported to change. The keynote also listed the parts count: [576 Rubin GPUs, 12,672 Vera CPU cores, 2,304 memory chips, 144 NVLink switches and 576 ConnectX-9 NICs](https://s201.q4cdn.com/141608511/files/doc%5Fdownloads/2025/03/GTC2025%5FKeynote.pdf?ref=insidedeeptech.com). Jensen Huang summed it up on stage: ["Each rack is 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). He also explained why NVIDIA showed it two years early. Data centers have to plan land and power with engineering teams [two to three years out, because "this isn't like buying a laptop"](https://www.datacenterdynamics.com/en/news/nvidias-rubin-ultra-nvl576-rack-expected-to-be-600kw-coming-second-half-of-2027/?ref=insidedeeptech.com). ## Inside the Kyber rack: vertical blades, a midplane and NVLink 7 Kyber is the successor to Oberon, the rack behind every GB200, GB300 and Vera Rubin NVL72\. NVIDIA's own description: compute blades [rotated vertically, like books on a shelf, up to 18 per chassis, with NVLink switch blades at the back joined through a cable-free midplane](https://blogs.nvidia.com/blog/gigawatt-ai-factories-ocp-vera-rubin/?ref=insidedeeptech.com). That midplane replaces the NVL72's copper cable cartridge. ServeTheHome photographed the GTC 2025 prototype and counted [18 compute blades in one chassis and a midplane with 18 columns and four rows of connectors](https://www.servethehome.com/the-nvidia-rubin-nvl576-kyber-midplane-is-huge/?ref=insidedeeptech.com). Because the whole rack is liquid cooled, the board doesn't need airflow cutouts. The GTC 2026 version looked different. HPC engineer Glenn K. Lockwood, writing from the show floor, counted [two compute chassis per rack instead of four, still with 18 trays per chassis, and noted that NVIDIA did not restate the rack's power](https://blog.glennklockwood.com/2026/03/gtc-2026-recap.html?ref=insidedeeptech.com). The switch side moves to NVLink 7\. Tom's Hardware reported the [seventh-generation switch keeps the 3,600 GB/s speed of NVLink 6 but lets the domain grow](https://www.tomshardware.com/pc-components/gpus/nvidia-demonstrates-rubin-ultra-tray-worlds-1st-ai-gpu-with-1tb-of-hbm4e?ref=insidedeeptech.com), and NVIDIA plans a faster ConnectX-9 NIC for scale-out. ### Why the midplane is the hard part Every GPU-to-GPU link in a Kyber rack crosses that one board. Trade analyses cited by Tom's Hardware describe [a 78-layer stack close to a square meter, built from three 26-layer sections, replacing more than 20,000 discrete cables](https://www.tomshardware.com/pc-components/gpus/nvidias-kyber-rack-for-rubin-ultra-slips-to-2028?ref=insidedeeptech.com). That's the part SemiAnalysis blamed for the reported delay. It said [Kyber NVL144 slipped to 2028 because the PCB midplane remains challenging to manufacture](https://www.cnbc.com/2026/07/06/nvidia-kyber-rack-system-delays-manufacturing-taiwan-rubin-chips-.html?ref=insidedeeptech.com). NVIDIA's answer was one line: "Our roadmap is intact." ## Power: 600 kW racks, 800 VDC and the 1 MW roadmap Kyber is the rack NVIDIA's 800 VDC push was built around. At 54 V, NVIDIA says power shelves for [Kyber at megawatt scale would eat up to 64 U of rack space, and a single 1 MW rack would need up to 200 kg of copper busbar](https://developer.nvidia.com/blog/nvidia-800-v-hvdc-architecture-will-power-the-next-generation-of-ai-factories/?ref=insidedeeptech.com). The fix is to convert medium-voltage AC to 800 VDC at the facility edge and feed racks directly. NVIDIA says [full-scale production of 800 VDC data centers will coincide with Kyber in 2027](https://developer.nvidia.com/blog/nvidia-800-v-hvdc-architecture-will-power-the-next-generation-of-ai-factories/?ref=insidedeeptech.com), with up to 5% better end-to-end efficiency. Inside the rack, Kyber uses a [64:1 LLC converter next to the GPU to step 800 VDC down to 12 VDC, in 26% less area than multi-stage designs](https://developer.nvidia.com/blog/building-the-800-vdc-ecosystem-for-efficient-scalable-ai-factories?ref=insidedeeptech.com). At GTC 2025, NVIDIA displayed an [800 V sidecar to power the rack](https://developer.nvidia.com/blog/nvidia-800-v-hvdc-architecture-will-power-the-next-generation-of-ai-factories/?ref=insidedeeptech.com). How big a jump is that? Tom's Hardware put [Blackwell racks at about 120 kW against Kyber's 600 kW](https://www.tomshardware.com/pc-components/gpus/nvidia-shows-off-rubin-ultra-with-600-000-watt-kyber-racks-and-infrastructure-coming-in-2027?ref=insidedeeptech.com), roughly five times the load in the same footprint. Inside Deep Tech's [800 VDC power distribution guide](https://www.insidedeeptech.com/800-vdc-power-distribution-ai-data-centers-full-guide/) covers the facility side: medium-voltage conversion, busways, protection and who actually ships the gear. There's a catch for buyers. NVIDIA's October 2026 MGX 1.2 spec for Vera Rubin racks supports [deployments up to 250 kW](https://blogs.nvidia.com/blog/ocp-global-summit-vera-rubin-scale-ai-factories/?ref=insidedeeptech.com). Bull says its Vera Rubin Ultra NVL72 [keeps the rack footprint and maximum rack power of Vera Rubin NVL72 and supports both AC and 800 VDC](https://www.bull.com/en/product-news/nvidia-vera-rubin-ultra?ref=insidedeeptech.com). In other words, the first Rubin Ultra you can buy may not need Kyber power at all. ## Cooling: 100% liquid, no fans Kyber has no air-cooled option. Tom's Hardware describes it as using [vertical trays with liquid cooling by default](https://www.tomshardware.com/pc-components/gpus/nvidia-demonstrates-rubin-ultra-tray-worlds-1st-ai-gpu-with-1tb-of-hbm4e?ref=insidedeeptech.com), and the GTC 2025 slide labeled it a [liquid-cooled Kyber rack](https://s201.q4cdn.com/141608511/files/doc%5Fdownloads/2025/03/GTC2025%5FKeynote.pdf?ref=insidedeeptech.com). [ServeTheHome](https://www.servethehome.com/the-nvidia-rubin-nvl576-kyber-midplane-is-huge/?ref=insidedeeptech.com) notes that fans and power supplies move out of the rack to free space for compute. That changes the facility, not just the server. A rack drawing hundreds of kilowatts needs coolant distribution units, facility water loops and leak detection sized for it. Inside Deep Tech's [data center liquid cooling guide](https://www.insidedeeptech.com/data-center-liquid-cooling-ai-full-guide/) walks through cold plates, CDUs and the ASHRAE water classes that set what a hall can support. NVIDIA hasn't published Kyber's coolant flow rate, inlet temperature or heat-rejection split. Until it does, treat any vendor's "Kyber-ready" cooling claim as a design target, not a qualified spec. ## Memory: HBM4E, 1 TB per GPU and the reported cuts Memory is where Rubin Ultra's spec is least settled. The official target is [1 TB of HBM4E per package](https://s201.q4cdn.com/141608511/files/doc%5Fdownloads/2025/03/GTC2025%5FKeynote.pdf?ref=insidedeeptech.com), from 16 stacks according to [heise](https://www.heise.de/en/news/Kyber-Rack-Nvidia-Shows-Rubin-Ultra-Boards-11213103.html?ref=insidedeeptech.com). That's more than three times the [288 GB of HBM4 on each Vera Rubin GPU](https://s201.q4cdn.com/141608511/files/doc%5Fdownloads/2025/03/GTC2025%5FKeynote.pdf?ref=insidedeeptech.com). Since mid-2026, three separate reports point the other way. None is confirmed by NVIDIA. | Item | NVIDIA announced | Reported since June 2026 (unconfirmed) | Reported by | | ------------------------ | -------------------- | ------------------------------------------------------------ | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | Compute dies per package | 4 reticle-sized dies | 2 dies, about half the performance | [SemiAnalysis via Tom's Hardware](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) | | HBM stacks | 16 (1 TB) | 8 HBM4E stacks | [SemiAnalysis via Tom's Hardware](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) | | Memory per GPU | 1 TB HBM4E | Tests at 192 GB and 256 GB, some on HBM4 | [The Information via Tom's Hardware](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) | | Stack height | Not specified | 12-Hi HBM4E baseline; 8-Hi and HBM4 options under evaluation | [TrendForce](https://www.trendforce.com/presscenter/news/20260804-13166.html?ref=insidedeeptech.com) | | Kyber NVL144 timing | 2027 | 2028 | [SemiAnalysis via CNBC](https://www.cnbc.com/2026/07/06/nvidia-kyber-rack-system-delays-manufacturing-taiwan-rubin-chips-.html?ref=insidedeeptech.com) | Source: [NVIDIA GTC 2025 keynote](https://s201.q4cdn.com/141608511/files/doc%5Fdownloads/2025/03/GTC2025%5FKeynote.pdf?ref=insidedeeptech.com) for the announced column; reported column as linked in each row. TrendForce's reasoning is supply, not demand. It 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), and says NVIDIA's priority for Rubin Ultra is raising I/O speed rather than shipping more GPUs. NVIDIA's only comment, given to Tom's Hardware, was that it is [optimizing Rubin Ultra across compute, networking and memory so it can build more GPUs and deploy more systems](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). That's not a denial. For background on why stack height, base dies and packaging set HBM supply, see Inside Deep Tech's [high-bandwidth memory guide](https://www.insidedeeptech.com/high-bandwidth-memory-hbm-full-guide/). ## Timeline: what NVIDIA says vs what the supply chain reports - **March 2025:** GTC 2025 keynote shows [Rubin Ultra NVL576 in a Kyber rack for 2H 2027](https://s201.q4cdn.com/141608511/files/doc%5Fdownloads/2025/03/GTC2025%5FKeynote.pdf?ref=insidedeeptech.com). - **May 2025:** NVIDIA says [800 VDC data centers will reach full-scale production alongside Kyber in 2027](https://developer.nvidia.com/blog/nvidia-800-v-hvdc-architecture-will-power-the-next-generation-of-ai-factories/?ref=insidedeeptech.com). - **October 2025:** at OCP, NVIDIA says Kyber will house [576 Rubin Ultra GPUs by 2027](https://blogs.nvidia.com/blog/gigawatt-ai-factories-ocp-vera-rubin/?ref=insidedeeptech.com). - **March 2026:** GTC 2026 shows a Kyber prototype; CNBC reports it is [expected to ship in 2027](https://www.cnbc.com/2026/03/16/nvidia-gtc-2026-ceo-jensen-huang-keynote-blackwell-vera-rubin.html?ref=insidedeeptech.com). NVIDIA renames the rack [Kyber NVL144](https://developer.nvidia.com/blog/nvidia-vera-rubin-pod-seven-chips-five-rack-scale-systems-one-ai-supercomputer/?ref=insidedeeptech.com). - **June 30, 2026:** SemiAnalysis reports the [quad-die Rubin Ultra was cancelled for a dual-die design](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). - **July 6, 2026:** SemiAnalysis reports [Kyber NVL144 delayed by more than 12 months to 2028](https://www.tomshardware.com/pc-components/gpus/nvidias-kyber-rack-for-rubin-ultra-slips-to-2028?ref=insidedeeptech.com). NVIDIA: "Our roadmap is intact." - **Mid-July 2026:** Huang calls Vera Rubin delay reports ["not true" but doesn't address Kyber](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?ref=insidedeeptech.com). - **August 2026:** TrendForce and The Information report [HBM configuration cuts under evaluation](https://www.trendforce.com/presscenter/news/20260804-13166.html?ref=insidedeeptech.com). - **October 13, 2026:** Bull plans [Vera Rubin Ultra NVL72 for the first half of 2028](https://www.bull.com/en/product-news/nvidia-vera-rubin-ultra?ref=insidedeeptech.com) on its European line. Notice what's missing. NVIDIA's August 2026 earnings call gave [no Rubin Ultra or Kyber ship date](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 its October 2026 OCP announcements centered on [MGX 1.2 for Vera Rubin](https://blogs.nvidia.com/blog/ocp-global-summit-vera-rubin-scale-ai-factories/?ref=insidedeeptech.com) rather than Kyber. ## Rubin Ultra vs Vera Rubin: who should buy now and who should wait Vera Rubin is shipping. NVIDIA said on its August 2026 call that it [commenced production shipments of Vera Rubin that month](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 Inside Deep Tech's [Vera Rubin NVL72 full guide](https://www.insidedeeptech.com/nvidia-vera-rubin-nvl72-full-guide/) covers the rack's specs and the HBM4 supply that sets its ramp. For most buyers, that settles it. If you need capacity in 2027, buy Vera Rubin NVL72 and treat Rubin Ultra as a 2028 refresh. NVIDIA itself 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), so waiting rarely gets you a better delivery slot. Wait for Rubin Ultra only if three things are true: you're building a new hall with 800 VDC and high-density liquid cooling anyway, your models need a scale-up domain larger than 72 GPUs, and you can absorb a 2028 date. Cost points the same way. Huang said NVIDIA's revenue opportunity rose from [about $25 billion per gigawatt with Grace Blackwell to $40 billion with Vera Rubin, and "after that, it's going to be higher"](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). Expect Rubin Ultra to cost more per megawatt, not less. ## Honest limits The real limits as of October 20, 2026: - **Spec drift:** the only official Rubin Ultra spec dates from [March 2025](https://s201.q4cdn.com/141608511/files/doc%5Fdownloads/2025/03/GTC2025%5FKeynote.pdf?ref=insidedeeptech.com). Reports since then say the die count and memory have been [cut](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), and NVIDIA hasn't published a replacement. - **Power density:** a Kyber rack was quoted at [600 kW](https://www.datacenterdynamics.com/en/news/nvidias-rubin-ultra-nvl576-rack-expected-to-be-600kw-coming-second-half-of-2027/?ref=insidedeeptech.com), about five times a [Blackwell rack](https://www.tomshardware.com/pc-components/gpus/nvidia-shows-off-rubin-ultra-with-600-000-watt-kyber-racks-and-infrastructure-coming-in-2027?ref=insidedeeptech.com). NVIDIA hasn't restated the figure for the 2026 design. - **800 VDC dependency:** Kyber is designed around [800 VDC input](https://developer.nvidia.com/blog/building-the-800-vdc-ecosystem-for-efficient-scalable-ai-factories?ref=insidedeeptech.com). The switchgear, protection and safety standards are still maturing, and few halls are built for it. - **Liquid cooling:** there's no air option, and coolant specs are unpublished. - **HBM4E supply:** [TrendForce expects DRAM to stay tight through 2027](https://www.trendforce.com/presscenter/news/20260804-13166.html?ref=insidedeeptech.com), which caps how many 1 TB packages can exist. - **Timeline slippage:** a [2028 Kyber date](https://www.tomshardware.com/pc-components/gpus/nvidias-kyber-rack-for-rubin-ultra-slips-to-2028?ref=insidedeeptech.com) is reported, the NVL72x2 stopgap was reportedly cancelled, and the eight-rack NVL576 depends on optics that may ship in low volume. - **Performance claims:** the [14x](https://s201.q4cdn.com/141608511/files/doc%5Fdownloads/2025/03/GTC2025%5FKeynote.pdf?ref=insidedeeptech.com) multiplier is a 2025 vendor target on vendor-chosen workloads, not a benchmark. What Inside Deep Tech chose to leave out: per-GPU TDP, rack price and exact Kyber coolant specs. NVIDIA hasn't published them, and the figures circulating are estimates. ## Inside Deep Tech's take 🧭 Inside Deep Tech's take: Kyber is the right idea arriving faster than the buildings that need to house it. The rack itself, vertical blades, a passive midplane and 800 VDC, is where AI hardware has to go. But the 2027 product is likely to be Rubin Ultra in an NVL72-class rack, with Kyber NVL144 as a 2028 upgrade. Inside Deep Tech's take: the safest planning assumption for the next 24 months is that the scale-up domain you can actually buy stays at 72 GPU packages through 2027\. Build new halls for 800 VDC and dense liquid cooling now, because Feynman's [Kyber NVL1152](https://developer.nvidia.com/blog/nvidia-vera-rubin-pod-seven-chips-five-rack-scale-systems-one-ai-supercomputer/?ref=insidedeeptech.com) will need them. Just don't pay a premium today for Kyber density that isn't confirmed. ## What to watch next - GTC 2027: does NVIDIA restate a [Rubin Ultra spec and a Kyber date](https://developer.nvidia.com/blog/nvidia-vera-rubin-pod-seven-chips-five-rack-scale-systems-one-ai-supercomputer/?ref=insidedeeptech.com)? - Any NVIDIA statement confirming or denying the [dual-die](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) and [HBM](https://www.trendforce.com/presscenter/news/20260804-13166.html?ref=insidedeeptech.com) reports. - OEM ship 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). - The first 800 VDC halls, such as [Foxconn's 40 MW Kaohsiung-1 site](https://blogs.nvidia.com/blog/gigawatt-ai-factories-ocp-vera-rubin/?ref=insidedeeptech.com), entering service. - HBM4E qualification at SK hynix, Samsung and Micron, which sets [Rubin Ultra's memory speed](https://www.trendforce.com/presscenter/news/20260804-13166.html?ref=insidedeeptech.com). --- ## FAQ #### What is NVIDIA Rubin Ultra? Rubin Ultra is NVIDIA's next data center GPU after Rubin. The official GTC 2025 spec was [four reticle-sized 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), due in the second half of 2027. #### What is the Kyber rack? Kyber is NVIDIA's next rack architecture after Oberon. It mounts compute blades [vertically, up to 18 per chassis, with NVLink switches behind a cable-free midplane](https://blogs.nvidia.com/blog/gigawatt-ai-factories-ocp-vera-rubin/?ref=insidedeeptech.com), and runs on 800 VDC with full liquid cooling. #### Is Rubin Ultra NVL576 one rack or eight? It depends on the date. In 2025, NVL576 meant one Kyber rack counted by GPU dies. Since GTC 2026, [NVL576 means eight 72-GPU racks joined into one NVLink domain](https://developer.nvidia.com/blog/nvidia-vera-rubin-pod-seven-chips-five-rack-scale-systems-one-ai-supercomputer/?ref=insidedeeptech.com), and the single rack is Kyber NVL144. #### How much power does a Kyber rack use? NVIDIA's CEO quoted [600 kW per rack](https://www.datacenterdynamics.com/en/news/nvidias-rubin-ultra-nvl576-rack-expected-to-be-600kw-coming-second-half-of-2027/?ref=insidedeeptech.com) for the 2025 design. NVIDIA hasn't restated a figure for the 2026 Kyber design, and Vera Rubin Ultra NVL72 [keeps Vera Rubin NVL72's rack power](https://www.bull.com/en/product-news/nvidia-vera-rubin-ultra?ref=insidedeeptech.com). #### Has Rubin Ultra been delayed? NVIDIA says no and that its roadmap is intact. SemiAnalysis reported that [Kyber NVL144 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). #### Should you wait for Rubin Ultra or buy Vera Rubin now? Most buyers should buy Vera Rubin. It's [shipping now](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), while Rubin Ultra's final spec and Kyber's date are unconfirmed. Inside Deep Tech's [Vera Rubin NVL72 guide](https://www.insidedeeptech.com/nvidia-vera-rubin-nvl72-full-guide/) covers the current rack. ## The bottom line Rubin Ultra and Kyber are NVIDIA's answer to the question every AI data center planner is asking: how big can one rack get? The 2025 answer was [576 dies at 600 kW](https://www.datacenterdynamics.com/en/news/nvidias-rubin-ultra-nvl576-rack-expected-to-be-600kw-coming-second-half-of-2027/?ref=insidedeeptech.com). The 2026 answer is more cautious, and the 2027 product may be more cautious still. Plan your next hall for 800 VDC and dense liquid cooling, and buy compute against what NVIDIA has confirmed. To see how today's racks stack up before Rubin Ultra arrives, read Inside Deep Tech's [Vera Rubin NVL72 vs GB300 vs GB200 NVL72 comparison](https://www.insidedeeptech.com/vera-rubin-nvl72-vs-gb300-vs-gb200-full-guide/) next. [Vera Rubin NVL72 vs GB300 vs GB200 NVL72: A Full Comparison Guide to NVIDIA's AI RacksMemory, NVLink, FLOPS, power and cooling compared line by line across NVIDIA's three NVL72 rack generations.![](https://www.insidedeeptech.com/favicon.ico)Inside Deep Tech](https://www.insidedeeptech.com/vera-rubin-nvl72-vs-gb300-vs-gb200-full-guide/) [Read the Vera Rubin vs GB300 vs GB200 comparison](https://www.insidedeeptech.com/vera-rubin-nvl72-vs-gb300-vs-gb200-full-guide/)