In a few years, Nvidia’s annual GTC has transformed from a niche event into the eye of the AI storm. Expectations are high for 2025 as well. In this overview, we highlight the most important announcements from Nvidia itself and some of those from partners such as HPE, Cisco and Pure Storage.
There is no shortage of ambition over at “Team Green”, as Nvidia is still known in the gaming circles which it was once exclusively active in. Nvidia announced not only new GPUs, but also 6G innovations, collaborations with Google, T-Mobile and General Motors, among others, a security system for self-driving cars, photonic network switches and newly open-sourced GPU libraries. On top of that, the previously announced Project DIGITS will be renamed DGX Spark and DGX Station, “Personal AI Computers” (PAICs?) for developers, researchers and data scientists.
For a complete overview, visit Nvidia’s site; unfortunately, we cannot give equal attention to all the announcements while writing this evening. We will focus on the most important and, for Nvidia, most lucrative product: Blackwell Ultra. Still, it should be said that we already know (and knew) its successor: Rubin, named after American astronomer Vera Rubin. Next, it appears, will be Rubin Ultra, and then Feynman, in honor of American physicist Richard Feynman. We will no doubt hear (and write) more about these successors at a later date, but first it’s time to delve into Blackwell’s second, more powerful phase of AI GPU horsepower.
Blackwell Ultra
Blackwell Ultra represents a significant improvement over its predecessor. A single B300 GPU delivers 1.5 times faster FP4 performance, though it requires 1,400 watts instead of 1,200 watts to do so. Memory per GPU has also increased by 50% to 288 GB, thanks to a 12-layer HBM3e stack rather than eight layers. Additionally, the system employs advanced water cooling, and ConnectX 8 network cards enhance interconnectivity. The optical modules have been upgraded from 800G to 1.6T networking. For power management, the NVL72 rack configuration features a standard capacitor tray, optional Battery Backup Units (BBU), and more than 300 super capacitors per rack. Furthermore, the GB300 combines a Grace CPU with the Blackwell Ultra GPU, marking the first server implementation of LPCAMM memory.
This will clearly be an expensive offering, to say the least. Although Nvidia doesn’t mention a price, each super capacitor reportedly costs between $20,000 and $30,000. The NVL72, as mentioned, requires 300 of these. When you add the cost of Grace CPUs and massive Blackwell GPUs, you’re looking at millions of dollars. Major hyperscalers and AI companies will likely order numerous nodes, while smaller AI startups and growing companies can afford only a handful at most. Additionally, Nvidia faces a backlog of Blackwell chips to deliver, partly due to design flaws in this generation. This delay will continue to be felt when Blackwell’s successor, Rubin, arrives.
Nvidia actively encourages scaling beyond a few Blackwell Ultra nodes. Those requiring the fastest connectivity can now turn to photonic network switches: Spectrum-X Photonics. These should enable “AI factories,” or AI-focused data centers, to scale to millions of GPUs.
HPE
Nvidia’s hardware typically reaches customers through system integrators. This gives companies like HPE, Cisco, and Lenovo the opportunity to combine their solutions with the AI computing power of Blackwell and Blackwell Ultra. HPE is introducing innovations that are both simple and scalable. One notable new product is the AI ModPod. This modular “data center-in-a-box” should lead to three times faster deployment of AI infrastructure, taking months instead of years. It also brings the PUE (Power Usage Effectiveness) down to 1.1, significantly lower than the industry average of 2.0 and more ambitious targets such as Germany’s requirement to be at or below 1.2 by July 1, 2026. Since a PUE lower than 1.0 is impossible, HPE is approaching a natural limit.
AI workloads revolve around data, so storage improvements significantly impact overall performance and costs for organizations. Disaggregated storage—storage that is separated from compute and scales independently—is central to HPE Alletra Storage MP. HPE promises zero overprovisioning, 45 percent energy reduction, and 40 percent lower costs through their “unified data services architecture.” This is a critical component of the broader HPE Private Cloud AI, a combination of HPE and Nvidia solutions addressing all AI needs for organizations of any size.
Cisco
Along with HPE, Cisco has major AI announcements during GTC 2025. Their focus is on security with the Cisco Secure AI Factory with NVIDIA. A secure AI approach here encompasses more than just resistance to cyber attacks or downtime; compliance and reliability are equally important. A variety of Cisco solutions work across all AI layers, from Cisco’s Hybrid Mesh Firewall (including Hypershield) to application security with Cisco AI Defense. High-performance storage from partners such as NetApp, Hitachi Vantara, VAST Data, and Pure Storage is combined with Nvidia’s AI Enterprise suite for an optimal out-of-the-box AI platform.
The emphasis is clearly on simplicity. Ethernet networking via Nvidia’s Spectrum-X cards makes deployment in existing data centers feasible, bridging the gap to an AI factory.
Pure Storage
Pure Storage also aims to integrate seamlessly with Nvidia’s AI stack. Pure Storage FlashBlade arrays are now certified for the Nvidia AI Data Platform reference design, making this Pure Storage solution suitable for Nvidia Cloud Partner and enterprise deployments. We can assume that companies like CoreWeave and U.S. hyperscalers will now widely use Pure’s offerings, where this wasn’t previously the case.
The validated storage options achieve the same goal we mentioned with Cisco and HPE: simplicity on one hand and scalability on the other. For the largest AI workloads, this allows for plug-and-play deployment, including RAG and training on proprietary data.
“GPUs are fast becoming the driving force behind the next wave of AI innovation,” emphasized Pure CTO Rob Lee. “The integration of the NVIDIA AI Data Platform into FlashBlade provides the AI-ready storage needed for optimal performance. Additionally, our recent NVIDIA certifications confirm that Pure Storage supports the pace and scale that AI models need to effect change. With our purpose-built storage solutions, enterprises can easily and efficiently leverage the power of AI to succeed.”
HP
Looking beyond exascale data centers at the scalability of Nvidia’s updated product line, we need to consider workstations and edge solutions, including those from HP. The most conventional variants are the ZBook Fury G1i and Z2 Tower G1i, a laptop and desktop with RTX PRO Blackwell GPUs, respectively. The ZBook Fury G1i features an RTX PRO 5000 Blackwell GPU paired with Intel Arrow Lake, while the Z2 Tower G1i includes an RTX PRO 6000 Blackwell GPU. The latter has an impressive VRAM capacity of 96GB. However, this video card requires a substantial 600 watts, which is offset by 88 percent performance improvements over the previous generation RTX PRO, Ada Lovelace.
VAST Data
Not just hardware vendors are expanding their capabilities with Nvidia. VAST Data is making its own VAST InsightEngine available for Nvidia DGX- and Nvidia-certified systems. This solution combines automated data capture, large-scale vector search, event-driven orchestration and GPU-optimized inferencing. In doing so, VAST Data promises to deliver enterprise-level security, with, as we hear more often during GTC, simplicity as the primary goal.
“Enterprises need AI solutions that are not just powerful but also simple to deploy and operate,” said John Mao, Vice President, Strategic Alliances at VAST Data. “As part of the VAST Data Platform VAST InsightEngine with NVIDIA streamlines enterprise AI adoption, delivering a real-time, AI-native data platform that scales seamlessly while ensuring enterprise-grade security and cost efficiency.”
Also read: Nvidia solidifies AI lead at GTC 2024 with Blackwell GPUs