For the next NVIDIA GPU in data centers, NVIDIA's GTC roadmap of March 2025 set this sequence: Blackwell Ultra (2025), Vera Rubin (2026), Rubin Ultra (2027), then Feynman (2028), according to NVIDIA's keynote recap and Tom's Hardware's account of the roadmap. This covers NVIDIA's data center GPUs, not GeForce gaming cards. NVIDIA reaffirmed its annual cadence of new AI supercomputers on 5 January 2026, so plan a contract review before the next generation rather than treating a roadmap year as a delivery guarantee.
The buying decision is about dates and configurations. A GPU announcement, production shipment and usable cloud deployment are different milestones. Write the one you need into the contract. The NVIDIA GPU generations guide covers the broader family history; this page tracks the commitments that can change a reservation decision.
NVIDIA GPU roadmap as of 1 October 2026
The table follows NVIDIA's presentations and subsequent disclosures. Reporting of those presentations is identified where used. Memory entries describe the GPU, not the CPU. Where NVIDIA has not stated a figure, the cell says so rather than borrowing one from another generation.
| Year | GPU | CPU | GPU memory | Rack and scale-up | NVLink per GPU | Status, 1 October 2026 |
|---|---|---|---|---|---|---|
| 2025 | Blackwell Ultra (B300, GB300) | Grace in GB300 | HBM3E | GB300 NVL72: 72 GPUs, 36 Grace CPUs | 1,800 GB/s | Production shipments began in the quarter ended 27 July 2025 |
| 2026 | Rubin | Vera | HBM4, up to 288 GB | Vera Rubin NVL72, formerly NVL144: 72 GPU packages, 36 Vera CPUs | NVLink 6: 3.6 TB/s, or 3 TB/s on NVIDIA's product page | Production shipments began in August 2026; CoreWeave announced availability on its cloud on 30 September, which NVIDIA called early access |
| 2027 target | Rubin Ultra | Not stated | HBM4E, 1 TB target | NVL72, Kyber NVL144, or NVL576 across eight racks of 72 | NVLink 7, per the reported roadmap | Roadmap target; Kyber timing disputed in reporting |
| 2028 target | Feynman | Rosa | Not stated | Kyber NVL1152 across eight Kyber racks | Not stated | Forward-looking roadmap |
Sources for the table: NVIDIA's Blackwell Ultra announcement (18 March 2025) and 27 August 2025 filing; NVIDIA's Rubin announcement and technical article (5 January 2026), its 21 July 2026 Rubin architecture article and its Vera Rubin NVL72 product page; Kress on the 26 August 2026 earnings call; NVIDIA's 30 September 2026 CoreWeave post; NVIDIA's 16 March 2026 technical article and keynote recap; and Tom's Hardware's reports of the GTC 2026 roadmap and Rubin Ultra tray (17 March 2026).
NVIDIA's updated article originally dated 13 October 2025 explicitly records the Vera Rubin NVL144 to NVL72 branding change. SemiAnalysis explained on 25 February 2026 that the old name counted 144 compute dies inside 72 packages; the new name counts those packages. Keep that distinction when comparing an old presentation with a new quote.
Vera Rubin: shipments are a separate milestone
NVIDIA said Rubin was in full production on 5 January 2026 while targeting partner products for the second half of that year. CFO Colette Kress then said on 26 August that production shipments began earlier that month. CoreWeave announced availability of Vera Rubin on its cloud on 30 September and named Cognition as a production customer; NVIDIA's post that day described access as early access.
For a reservation, require a start date, region, configuration and acceptance test. A manufacturing statement does not settle any of those terms. The Vera Rubin platform guide explains the system; the Rubin versus Blackwell comparison addresses the upgrade decision.
NVIDIA's 5 January 2026 announcement paired Rubin with Vera in its rack design and also announced HGX Rubin NVL8 for x86 systems. Specify the host platform as well as the GPU in a purchase order. The Vera CPU guide covers that part of the choice.
Rubin CPX and LPX: remove the old rollout assumption
NVIDIA announced Rubin CPX for massive-context inference on 9 September 2025, targeting the end of 2026. In the GTC press Q&A published by Tom's Hardware on 23 March 2026, NVIDIA's Ian Buck said CPX had been pulled to prioritize LPU decode delivery that year. That statement changed the rollout plan without declaring permanent cancellation.
NVIDIA announced Groq 3 LPX racks on 16 March 2026 and said LPX was in full production on 24 August. Its March technical description places LPX alongside Vera Rubin as an additional inference engine. Do not assume an ordinary Rubin reservation includes LPX. Ask for it explicitly if your acceptance test needs it; the CPX and LPX guide explains their different roles.
Rubin Ultra and the NVIDIA Kyber rack
NVIDIA's 18 March 2025 presentation described Rubin Ultra with four reticle-sized GPU dies and an HBM4E memory target, according to Data Center Dynamics that day. Tom's Hardware's 6 August 2025 roadmap account counted the proposed Rubin Ultra NVL576 as 144 GPU packages containing 576 compute chiplets. Those are historical design targets.
NVIDIA's 16 March 2026 technical article gives NVL576 a different meaning: eight MGX racks, each containing 72 Rubin Ultra GPUs, connected in one NVLink domain. NVIDIA describes a two-layer all-to-all topology with copper and direct optical connections. In the same article, Kyber first appears with Rubin Ultra as a standalone NVL144 system holding 144 GPUs in one rack.
The practical consequence is substantial. A quote for NVL576 needs a dated topology diagram and a package count. Do not accept the suffix alone as a description of the space, network or power being purchased. The 2025 and 2026 designs do not describe the same physical installation.
Data Center Dynamics reported Huang's 600 kW proposed Kyber power figure on 18 March 2025, a VENDOR CLAIM for a liquid-cooled design. Tom's Hardware described the exhibit as a mockup on 19 March. That historical figure is not a published power budget for NVIDIA's later eight-rack NVL576 design. Require power and cooling acceptance criteria for the actual system ordered.
Timing is contested. On 6 July 2026, Tom's Hardware reported SemiAnalysis's projection that Kyber NVL144 would slip to 2028 because of midplane manufacturing problems, an ANALYST ESTIMATE. NVIDIA's response was: “Our roadmap is intact.” Treat the delay as a risk to negotiate around, not an established replacement date.
Memory also remains a contract detail. TrendForce reported on 4 August 2026 that Rubin Ultra's final memory configuration remained undecided, an ANALYST ESTIMATE. Preserve HBM4E as NVIDIA's announced target rather than promising a final configuration; the HBM4 guide covers the memory transition.
NVIDIA Feynman: a planning horizon
NVIDIA's GTC 2026 roadmap retained Feynman for 2028, as reported by Tom's Hardware on 17 March. NVIDIA's own 16 March keynote recap identified Rosa CPU, LP40 LPU, BlueField-5 and CX10 networking, with copper and co-packaged-optics Kyber scale-up. Its technical article that day associated the future eight-rack Kyber NVL1152 architecture with Feynman.
That is enough to schedule a review, but not to specify a purchase. Leave Feynman performance, memory capacity and delivery guarantees out of the business case. A contract that runs into this window should remain worthwhile on the hardware actually promised, even if Feynman arrives later or your workload cannot use it immediately.
Networking has its own delivery dates
NVIDIA's 5 January 2026 Rubin announcement and technical article identify NVLink 6 for scale-up and ConnectX-9 for scale-out. The GTC roadmap reported by Tom's Hardware on 17 March 2026 associates NVLink 7 with Rubin Ultra and Kyber in 2027. These are separate generations to name in a cluster specification.
NVIDIA announced Spectrum-X Photonics Ethernet and Quantum-X Photonics InfiniBand switches on 18 March 2025. That release targeted Quantum-X for later in 2025 and Spectrum-X for 2026; NVIDIA's 31 May 2026 COMPUTEX recap then described Spectrum-X Ethernet Photonics as in production. The original Quantum-X target alone does not prove delivery.
Ask which connections stay within an NVLink domain and which cross the network. CoreWeave's 16 September 2026 announcement describes its multi-rack Rubin cluster using Spectrum-X Ethernet between racks. It does not establish delivery of the future Rubin Ultra NVL576 scale-up design.
What changed between presentations
- Late December 2025: SemiAnalysis's 25 February 2026 account dates the Vera Rubin naming reversal to this period: NVL144 became NVL72, counting packages instead of dies.
- 16 March 2026: NVIDIA described NVL576 across eight MGX racks and Kyber's initial standalone NVL144 form, replacing the physical interpretation of the 2025 proposal.
- March 2026: Buck's Q&A, published on 23 March, removed CPX from the year's delivery priorities in favor of LPU decode.
- 16 March 2026: NVIDIA named Rosa for Feynman. Tom's Hardware's 6 August 2025 roadmap account had paired Feynman with Vera.
Competing timelines worth keeping in the bid
AMD's 23 July 2026 announcement launched MI400 and Helios, described Helios as in production and said OpenAI expected it online beginning in Q4 2026. That target overlaps Rubin's deployment period; it is not proof that the OpenAI deployment had finished. Use the AMD Instinct family guide to orient a separate evaluation.
Google's 31 March 2026 release note marked Ironwood-family TPU7x generally available. AWS announced general availability of Trainium3 EC2 Trn3 UltraServers on 2 December 2025. These dated launches justify evaluating alternatives during a renewal, without assuming equivalent software support or performance. The AI accelerator comparison helps scope that work. Put porting effort into the evaluation budget before using a competing roadmap as negotiating pressure.
Match the commitment to the annual cadence
The rule: prefer commitments of a year or less, with a review before renewal. Choose a longer term only when the contracted hardware pays back without an assumed upgrade. For purchases, use the same test: the workload should justify the asset without a promised resale value after the next launch.
For a stable production workload, negotiate predictable capacity and an annual price review. For a changing model, prioritize migration rights and a smaller committed baseline. If you reserve Rubin Ultra or Kyber, tie deposits and acceptance to the specified topology, power envelope and delivery milestone. A roadmap year is too broad to be an acceptance condition.
Plan for generations to coexist. AWS's 26 August 2026 announcement included Blackwell Ultra, Rubin and Rubin Ultra in its planned deployments across 2027 and 2028, rather than replacing the older family outright. Treat previous-generation discounting as a planning hypothesis, not a guaranteed saving. Older hardware can remain useful for your workload; a new announcement alone does not establish a lower rental bill. Check the GPU price index at renewal and compare completed work, migration effort and contract restrictions before moving.
The live table below shows current H100, H200, B200, B300 and GB300 offers for that comparison.
| GPU | Cheapest $/GPU-hr | Provider | Providers in stock |
|---|---|---|---|
| H100 | $2.50 | Hyperstack | 9 |
| H200 | $3.43 | QuantaCloud | 6 |
| B200 | $6.79 | RunPod | 3 |
| B300 | $7.89 | RunPod | 3 |
| GB300 | none in stock | ||
Whatever NVIDIA ships next, buy enough certainty to finish the job and preserve a review point before the next annual step. Sign a longer commitment only if its economics work on today's contracted configuration and its migration or exit terms are written down.
Sources
- NVIDIA: Computex 2024 keynote and first Rubin roadmap, 2 June 2024.
- NVIDIA: Blackwell Ultra announcement, 18 March 2025.
- NVIDIA: quarterly filing (PDF), 27 August 2025.
- Tom's Hardware: NVIDIA's roadmap through Feynman, 6 August 2025.
- Data Center Dynamics: Rubin Ultra NVL576 and the 600 kW Kyber rack, 18 March 2025.
- NVIDIA: Rubin platform announcement, 5 January 2026.
- NVIDIA: Inside the Rubin platform, 5 January 2026, updated 16 March 2026.
- Tom's Hardware: Rubin Ultra and the Kyber mockup at GTC 2025, 19 March 2025.
- NVIDIA: Rubin Ultra NVL576 and Kyber topology, 16 March 2026.
- Tom's Hardware: NVIDIA's GTC 2026 data center roadmap, 17 March 2026.
- NVIDIA: GTC 2026 keynote recap, 16 March 2026.
- NVIDIA: Spectrum-X and Quantum-X Photonics switches, 18 March 2025.
- NVIDIA: Computex 2026 recap, 31 May 2026.
- CoreWeave: multi-rack Vera Rubin NVL72 cluster, 16 September 2026.
- AMD: MI400 series and Helios launch, 23 July 2026.
- Google Cloud: TPU release notes, 31 March 2026.
- AWS: Trn3 UltraServers generally available, 2 December 2025.
- Tom's Hardware: report of Kyber slipping to 2028, 6 July 2026.
- TrendForce: Rubin Ultra memory options, 4 August 2026.
- NVIDIA: Q2 fiscal 2027 earnings call transcript (PDF), 26 August 2026.
- NVIDIA: NVL144 to NVL72 naming note, 13 October 2025.
- SemiAnalysis: Vera Rubin design and naming, 25 February 2026.
- NVIDIA: Rubin GPU architecture, 21 July 2026.
- CoreWeave: Vera Rubin NVL72 availability, starting with Cognition, 30 September 2026.
- NVIDIA: CoreWeave early access to Vera Rubin, 30 September 2026.
- NVIDIA: Rubin CPX announcement, 9 September 2025.
- Tom's Hardware: GTC 2026 press Q&A transcript, 23 March 2026.
- NVIDIA: Vera Rubin platform with Groq 3 LPX, 16 March 2026.
- NVIDIA: Groq 3 LPX architecture, 16 March 2026.
- NVIDIA: Groq 3 LPX in full production, 24 August 2026.
- Tom's Hardware: Rubin Ultra tray with 1 TB of HBM4E, 17 March 2026.
- NVIDIA: GTC 2025 keynote recap, 25 March 2025.
- AWS: multigeneration NVIDIA GPU deployment plan, 26 August 2026.
- NVIDIA: Vera Rubin NVL72 specifications, as of 1 October 2026.