NVIDIA Unveils Roadmap for Next-Gen AI GPUs: Blackwell Ultra and Vera Rubin

Nvidia confirms Blackwell Ultra launch, teases Vera Rubin architecture for 2026

Forward-looking: Despite facing a setback in the rollout of its Blackwell GPUs for data centers last year due to a design flaw, Nvidia has swiftly rebounded and is poised to deliver its next series of products over the next few years. CEO Jensen Huang confirmed during the company's earnings call that the next major release, dubbed Blackwell Ultra (B300-series), is on track for the second half of this year. This mid-cycle refresh of the Blackwell architecture promises significant improvements over its predecessors. The B300-series is expected to offer higher compute performance and eight stacks of 12-Hi HBM3E memory, providing up to 288GB of onboard memory. Although unofficial, there are estimates of a 50 percent performance uplift compared to the B200-series. To complement these powerful GPUs, Nvidia will introduce the Mellanox Spectrum Ultra X800 Ethernet switch, boasting a radix of 512 and support for up to 512 ports. This networking upgrade will further enhance the capabilities of AI and HPC systems built around the B300-series. Looking beyond Blackwell, Nvidia is already working on its next-generation architecture, codenamed Vera Rubin. Set to debut in 2026, the Rubin GPUs represent a significant step toward achieving artificial general intelligence (AGI). The Rubin platform will feature eight stacks of HBM4E memory, offering up to 288GB of memory, along with a Vera CPU, NVLink 6 switches operating at 3600 GB/s, CX9 network cards supporting 1,600 Gb/s, and X1600 switches. Huang has hinted at the transformative potential of the Rubin architecture, describing it as a major leap forward in terms of performance and capabilities. Nvidia has also indicated its readiness to discuss post-Rubin products at the upcoming GPU Technology Conference (GTC) in March. One potential breakthrough on the horizon is the rumored Rubin Ultra, projected for release in 2027. This product could push the boundaries of GPU design even further, potentially incorporating 12 stacks of HBM4E memory. This is a substantial increase from the 8 stacks used in previous generations, potentially offering up to 576GB of total memory. The use of HBM4E technology would provide unprecedented memory bandwidth and capacity, crucial for handling increasingly complex AI models and computations. To achieve this, Nvidia would need to master the use of 5.5-reticle-size CoWoS interposers and 100mm × 100mm substrates manufactured by TSMC. This represents a significant increase from the current 3.3-reticle-size interposers used in today's most advanced GPUs. The larger interposer size would allow for more components to be integrated onto a single package, enabling the inclusion of additional memory stacks and potentially more GPU tiles.

Nvidia confirms Blackwell Ultra and Vera Rubin GPUs are on track for 2025 and 2026 -- post-Rubin GPUs in the works

A design flaw delayed Nvidia's rollout of Blackwell GPUs for data centers, prompting the company to redesign the silicon and packaging. However, this did not affect Nvidia's work on mid-cycle refresh Blackwell 300-series (Blackwell Ultra) GPUs for AI and HPC and next-generation Vera Rubin GPUs. Nvidia can already share some details about Rubin GPUs and its post-Rubin products. "Blackwell Ultra is [due in the] second half," Jensen Huang, chief executive of Nvidia, reaffirmed analysts and investors at the company's earnings conference call. "The next train [is] Blackwell Ultra with new networking, new [12-Hi HBM3E] memory, and, of course, new processors. [...] We have already revealed and have been working very closely with all of our partners on the click after that. The click after that is called Vera Rubin and all of our partners are getting up to speed on the transition to that. [...] [With Rubin GPUs] we are going to provide a big, big, huge step up." Later this year, Nvidia plans to release its Blackwell B300-series solutions for AI and HPC (previously known as Blackwell Ultra) that will offer higher compute performance as well as eight stacks of 12-Hi HBM4E memory, thus providing up to 288 GB of memory onboard. Unofficial information indicates that performance uplift enabled by Nvidia's B300-series will be around 50% compared to the comparable B200-series products, though the company has yet to confirm this. To further improve performance, B300 will be offered with Nvidia's Mellanox Spectrum Ultra X800 Ethernet switch, which has a radix of 512 and can support up to 512 ports. Nvidia is also expected to provide additional system design freedom to its partner with its B300-series data center GPUs. Nvidia's next-generation GPU series will be based on the company's all-new, codenamed Rubin architecture, further improving AI compute capabilities as industry leaders march towards achieving artificial general intelligence (AGI). In 2026, the first iteration of Rubin GPUs for data centers will come with eight stacks of HBM4E memory (up to 288 GB). The Rubin platform will also include a Vera CPU, NVLink 6 switches at 3600 GB/s, CX9 network cards supporting 1600 Gb/s, and X1600 switches. Jensen Huang plans to talk about Rubin at the upcoming GPU Technology Conference (GTC) in March, though it remains to be seen what he plans to discuss. Surprisingly, Nvidia also intends to talk about post-Rubin products at the GTC. From what Jensen Huang announced this week, it is unclear whether the company plans to reveal details of Rubin Ultra GPUs or its GPU architecture that will come after the Rubin family. Speaking of Rubin Ultra, this could indeed be quite a breakthrough product. It is projected to come with 12 stacks of HBM4E in 2027 once Nvidia learns how to efficiently use 5.5-reticle-size CoWoS interposers and 100mm × 100mm substrates made by TSMC. "Come to GTC and I will [tell you about] Blackwell Ultra," said Huang. "There are Rubin, and then [we will] show you what is one click after that. Really, really exciting new product."

NVIDIA CEO confirms Blackwell Ultra, Vera Rubin AI GPUs are on-track, reveal set for GTC 2025

NVIDIA has confirmed its beefed-up Blackwell Ultra and next-gen Vera Rubin AI architectures are on track, lining up with recent reports that we'll get a huge info dump on the company's new AI GPUs at GTC 2025 in a few weeks time. During the company's recent earnings call, NVIDIA CEO Jensen Huang confirmed with analysts that the recent GB200 AI server yield rates won't affect the company's annual release cadence, but the analysts also asked Jensen how NVIDIA would manage Blackwell Ultra and Rubin at similar release periods. Jensen said: "Yes. Blackwell Ultra is second half. As you know, the first Blackwell was we had a hiccup that probably cost us a couple of months. We're fully recovered, of course. The team did an amazing job recovering and all of our supply chain partners and just so many people helped us recover at the speed of light". He continued: "And so, now we've successfully ramped up production of Blackwell. But that doesn't stop the next train. The next train is on an annual rhythm and Blackwell Ultra with new networking, new memories, and of course, new processors, and all of that is coming online". Jensen talked about the company's next-gen Vera Rubin AI architecture: "And the click after that is called Vera Rubin and all of our partners are getting up to speed on the transition of that and so preparing for that transition. And again, we're going to provide a big, huge step-up". NVIDIA's next-gen Rubin AI GPUs will use the new bleeding-edge HBM4 memory standard, which SK hynix, Samsung, and Micron are all hard at work now. HBM3E is being used on B200 and GB200, while B300 and GB300 will also use HBM3E but bump up the memory capacity per GPU.

Blackwell Ultra is official but Huang is already hyping Rubin

Nvidia CEO Jensen Huang has teased the upcoming reveal of the company's flagship chip, Blackwell Ultra, slated for presentation at the Nvidia GTC 2025 event in March 2025. Huang confirmed its name during the fiscal Q4 2025 earnings call. Huang stated that Blackwell Ultra will officially launch in the second half of 2025 and will include enhancements in processors, networking, and memory, while maintaining the same system architecture as the previous Blackwell series. Both Blackwell Ultra and another product, Vera Rubin, were introduced in a company roadmap at Computex 2024, with Huang verifying their names. Nvidia initially unveiled the Blackwell series at GTC 2024, highlighting significant advancements in AI power and efficiency. The first Blackwell "superchip," the GB200, can scale from a single rack to a full data center, helping Nvidia to solidify its position in the AI competitive landscape. The Blackwell architecture features 208 billion transistors across two GPU dies, linked by a 10 TB/second chip-to-chip connection, yielding performance improvements of up to 30 times over the previous Hopper series for AI inference tasks. It is capable of delivering up to 20 petaflops of FP4 power, significantly outperforming other solutions currently available on the market. Moreover, Nvidia claims that Blackwell can reduce costs and energy consumption by up to 25 times. For instance, training a 1.8 trillion parameter model, which needed approximately 8,000 Hopper GPUs and 15 megawatts of power, can now be accomplished with only 2,000 Blackwell GPUs and just four megawatts of power. The next generation following Blackwell Ultra will be Vera Rubin, with expected releases in 2026. This will feature both CPU and GPU products, including a Vera Rubin board that integrates the GPU and CPU into a "superchip." During the earnings call, Huang reiterated, "Blackwell Ultra is [due in the] second half" of 2025, emphasizing the advancements in networking and memory with the introduction of new 12-Hi HBM3E memory and processors. The company is also preparing its partners for the transitions to Vera Rubin, stating, "With Rubin GPUs, we are going to provide a big, big, huge step up." Nvidia intends to launch the Blackwell B300-series solutions, with heightened compute performance, later this year. These will provide enhanced capabilities, featuring eight stacks of 12-Hi HBM4E memory, yielding up to 288GB of onboard memory. Although informal reports suggest a performance increase of around 50% compared to the earlier B200-series products, Nvidia has not yet officially confirmed this information. Wall Street bets big on Nvidia but is Blackwell a ticking time bomb? To further enrich performance, the B300 series will incorporate Nvidia's Mellanox Spectrum Ultra X800 Ethernet switch, supporting up to 512 ports. Nvidia plans to offer additional system design flexibility with its B300-series data center GPUs. Looking further ahead, Nvidia's roadmap extends to 2027, when the next GPU series, based on the new Rubin architecture, is expected. The first Rubin GPUs for data centers are anticipated in 2026, equipped with eight stacks of HBM4E memory. This RF platform will include a Vera CPU, NVLink 6 switches at 3600 GB/s, CX9 network cards supporting 1,600 Gb/s, and X1600 switches. Huang plans to discuss Rubin at the forthcoming GTC in March, while also addressing post-Rubin products. Uncertainty remains regarding whether Nvidia will unveil details about Rubin Ultra GPUs or its architecture that will follow the Rubin family. Projected advancements for Rubin Ultra include up to 12 stacks of HBM4E memory by 2027, contingent on Nvidia's ability to effectively utilize advanced substrate and interposer technology developed with TSMC. Nvidia reported total revenues for the financial year exceeded $130.5 billion, reflecting a 114% year-over-year increase, greatly attributed to record data center revenues of $35.6 billion, an increase of 16% from the previous quarter and 93% from the same period last year. Huang noted the incredible demand for Blackwell, stating, "Demand for Blackwell is amazing as reasoning AI adds another scaling law -- increasing compute for training makes models smarter and increasing compute for long thinking makes the answer smarter." In concluding remarks, Huang enthused, "Come to GTC and I will [tell you about] Blackwell Ultra. There are Rubin, and then [we will] show you what is one click after that. Really, really exciting new product." Nvidia successfully ramped up production of Blackwell AI supercomputers, generating billions in sales within the first quarter. Huang emphasized the rapid advancement of AI, stating, "AI is advancing at light speed as agentic AI and physical AI set the stage for the next wave of AI to revolutionize the largest industries."

NVIDIA's CEO Jensen Huang Confirms Blackwell Ultra & Vera Rubin AI Architectures To Be On-Track; Ready To Unleash Their Might On GTC 2025

NVIDIA has confirmed that its upcoming Blackwell Ultra and Rubin AI architectures are on track, saying that the next "AI train" will be broader than ever. Well, Team Green's next-gen AI lineups are said to be much more disruptive than existing ones, and it seems like NVIDIA isn't putting the pedal off the accelerator. In a statement to analysts, NVIDIA's CEO Jensen Huang confirmed that Blackwell Ultra and Vera Rubin AI lineups are on launch schedule and that the recent "GB200" yield rates issue won't affect Team Green's annual release cadence. Upon being asked about how NVIDIA will manage Blackwell Ultra and Vera Rubin at nearly similar periods by JPMorgan's Harlan Sur, here's what Jensen had to say (via earnings call): Yes. Blackwell Ultra is second half. As you know, the first Blackwell was we had a hiccup that probably cost us a couple of months. We're fully recovered, of course. The team did an amazing job recovering and all of our supply chain partners and just so many people helped us recover at the speed of light. And so, now we've successfully ramped up production of Blackwell. But that doesn't stop the next train. The next train is on an annual rhythm and Blackwell Ultra with new networking, new memories, and of course, new processors, and all of that is coming online. - NVIDIA's CEO For a quick rundown on the Blackwell Ultra "B300" series, it is rumored that NVIDIA is planning to go big on power figures this time, as the GB300 AI server is supposed to feature up to 1400W of TDP, which is a massive rise when compared against the Blackwell GB200, which stands at 1000W. With the architectural upgrades, we are looking at around 1.4 times higher FP4 performance when compared to the previous generation and higher memory capacity from 192 GB to 288 GB by effectively utilizing 12-Hi stacks of HBM3E technology. Blackwell Ultra is said to act as a bridge between the original Blackwell series and the next-gen Vera Rubin, and Team Green's intention with this lineup is clear: maximize the options available to its mainstream clients. With the ongoing Blackwell yield rate issues, it is supposed that B300 will act as a "hard reset" for NVIDIA, allowing them to show the true power of their lineup. Following Blackwell Ultra will indeed be with Vera Rubin, which will be revolutionary for the AI market. And the click after that is called Vera Rubin and all of our partners are getting up to speed on the transition of that and so preparing for that transition. And again, we're going to provide a big, huge step-up. - NVIDIA's CEO While we won't go much into the details of Rubin architecture, the one big thing with it is the use of HBM4, which is said to play a pivotal role in how computing markets will move in the future. The interesting part is that both Blackwell Ultra and Rubin are expected to be unveiled in March at GTC 2025, with availability being pretty close as well, given that the industry has already transitioned into developing the essentials needed for the respective AI lineups.

NVIDIA Blackwell Ultra Coming This year and Rubin Architecture in 2026

The Rubin architecture is set to roll out potentially in 2026 and is an important part of the company's new GPU roadmap. This design will pack eight stacks of HBM4(E) memory along with a new kind of "Vera" CPU. It also features NVLink 6 switches that promise to deliver around 3600 GB/s in bandwidth and CX9 network cards that can handle speeds up to 1600 Gb/s. With the addition of X1600 switches, the architecture is built to support the demanding needs of AI and high-performance computing. The Blackwell B300 series, or "Blackwell Ultra" as it's known internally, is scheduled for release in the latter half of 2025. This series is engineered to offer a noticeable performance boost over the previous B200 series. It includes eight stacks of 12-Hi HBM3E memory, giving it a total memory capacity of up to 288 gigabytes. The system also comes with a Mellanox Spectrum Ultra X800 Ethernet switch that provides 512 ports to boost data connectivity and throughput. Early estimates indicate that with a thermal design power of about 1,400 watts, this chip might deliver up to 50% more performance thanks to improvements in core count and memory bandwidth. Meanwhile, the team is also working on fine-tuning semiconductor packaging and overcoming supply chain challenges. Looking further ahead, more details about future GPU designs are expected at the upcoming GPU Technology Conference in March. One of the potential announcements is for Rubin Ultra, which might launch in 2027. This design could feature 12 stacks of HBM4E memory and introduce new packaging techniques, including 5.5-reticle-size CoWoS interposers and TSMC substrates measuring 100 mm by 100 mm. Although final performance details haven't been confirmed yet, these technical upgrades suggest a significant leap in memory bandwidth and compute power. Source: TomsHardware

Nvidia CEO Jensen Huang teases Blackwell Ultra reveal, along with "the one click after that"

Nvidia sees overall revenues grow 114% year-on-year as AI demand booms Nvidia CEO Jensen Huang has teased that the company's next flagship chip will be shown off within the next few weeks - as well as apparently officially confirming its name. Speaking to analysts on the company's fiscal Q4 2025 earnings call, Huang revealed the new hardware will bear the moniker of Blackwell Ultra, and will be unveiled at its Nvidia GTC 2025 event in March 2025. "Come to GTC and I'll talk to you about Blackwell Ultra, Vera Rubin, and then show you the one click after that," Huang declared. Huang went on to add Blackwell Ultra will be officially launched in the second half of 2025, and offer upgrades in processors, networking, and memory - but will be built on the same system architecture as Blackwell. Both Blackwell Ultra and Vera Rubin were teased on a company roadmap document at Computex 2024, but Huang has now seemingly confirmed the naming of both. Nvidia unveiled Blackwell at GTC 2024, promising a major step forward in terms of AI power and efficiency. It has since seen a number of releases, including the first Blackwell "superchip", the GB200, which has the ability to scale up from a single rack all the way to an entire data center, as Nvidia looks to push on with its leadership in the AI race. Blackwell contains 208 billion transistors (up from 80 billion in Hopper) across its two GPU dies, which are connected by 10 TB/second chip-to-chip link into a single, unified GPU, making it up to 30x faster than Hopper when it comes to AI inference tasks, offering up to 20 petaflops of FP4 power, far ahead of anything else on the market today. Despite this, Nvidia says Blackwell can reduce cost and energy consumption by up to 25x, giving the example of training a 1.8 trillion parameter model - which would previously have taken 8,000 Hopper GPUs and 15 megawatts of power - but can now be done by just 2,000 Blackwell GPUs consuming just four megawatts. Vera Rubin is the *next* step forward for Nvidia after Blackwell Ultra, with an expected release in 2026, with CPU and GPU products anticpated, including a Vera Rubin board combining the GPU and CPU into a "superchip". Huang's tease came as Nvidia revealed its most recent set of financial results, with the company once again seeing record returns as it benefits from the huge surge in AI demand. Overall Nvidia revenues more than doubled year-on-year, reaching $130.5 billion for the financial year, a 114% rise, largely helped by record data center revenues of $35.6 billion, up 16% from the previous quarter, and up 93% from a year ago. "Demand for Blackwell is amazing as reasoning AI adds another scaling law -- increasing compute for training makes models smarter and increasing compute for long thinking makes the answer smarter," Huang said of the results. "We've successfully ramped up the massive-scale production of Blackwell AI supercomputers, achieving billions of dollars in sales in its first quarter. AI is advancing at light speed as agentic AI and physical AI set the stage for the next wave of AI to revolutionize the largest industries."