Among the myriad of announcements include new silicon photonics networking switches, personal AI supercomputers, and plans for a quantum computing research centre in Boston.
We also got a look at the next evolution at Blackwell — Blackwell Ultra.
Here are all the major announcements from Nvidia GTC 2025:
Blackwell updates: Meet Ultra & the GB300
Nvidia’s first-gen Blackwell chips are barely out the door after a hardware hiccup but the chipmaker is doubling down, unveiling the Blackwell Ultra, the next-gen hardware aimed at boosting training and scaling inference.
Unveiled were two versions:
GB300 NVL72: A rack-scale solution with 72 Blackwell Ultra GPUs and 36 Grace CPUs, designed to act as a single massive AI GPU. This enables models to break down complex requests into multi-step solutions, improving AI reasoning capabilities.
HGX B300 NVL16: A high-performance server-class unit, boasting 11x faster inference for large language models (LLMs), 7x more compute, and 4x more memory compared to NVIDIA’s Hopper-generation GPUs.
Top cloud providers like AWS, Google Cloud, Microsoft Azure, and Oracle Cloud Infrastructure will be among the first to offer Blackwell Ultra-powered instances.
Server makers including Dell, HPE, Lenovo, and Supermicro are also set to roll out Blackwell Ultra-based AI infrastructure in late 2025.
Nvidia said the new Blackwell Ultra is its biggest leap yet into AI reasoning:
AI has made a giant leap — reasoning and agentic AI demand orders of magnitude more computing performance. We designed Blackwell Ultra for this moment — it’s a single versatile platform that can easily and efficiently do pretraining, post-training and reasoning AI inference.
The new Ultra adds to Nvidia’s annual cadence, or ‘one year rhythm’ where it plans to launch a new flagship hardware offering every year, which is currently mapped to release the next-gen Rubin line in 2026, followed by an Ultra-style updated version in 2027.
DGX Spark and DGX Station personal AI supercomputers
Nvidia is looking to bring supercomputers to the home with the firm giving us an official unveiling of DGX Spark and DGX Station, desktop supercomputers powered by Blackwell hardware.
We got our first look at the concept back at CES under the working name Project DIGITS, with the idea of a unit capable of running high-end AI models while taking up as much space as a standard desktop.
DGX Spark is powered by a GB10 Superchip, a toned-down Blackwell unit that despite its size can apparently support up to 1,000 trillion operations per second of AI compute for fine-tuning and inference.
DGX Station, meanwhile, is the more souped-up version, with Nvidia touting it as bringing “data centre-level performance to desktops”.
It features the new B300 Grace Blackwell Ultra Desktop Superchip and boasts a massive 784GB of coherent memory space to power large-scale training and inferencing workloads.
DGX Station is expected to be available from manufacturing partners like Asus, Dell, HP, and Supermicro later this year.
In terms of cost projects, nothing concrete was provided though reports from back at CES put the lower-end Spark model at a starting price of ~$3,000.
New networking switches - Nvidia now does photonics?
Arguably among the more surprising updates from GTC 2025 was that Nvidia has joined the growing photonics movement.
The firm unveiled photonics-based networking switches designed to connect millions of data centre GPUs while drastically reducing energy consumption.
The new Spectrum-X photonics switches come in multiple configurations, including 128 ports of 800Gb/s or 512 ports of 200Gb/s — with Nvidia claiming they offer 3.5x more power efficiency, 63x greater signal integrity, and 10x better network resiliency at scale compared to traditional switches.
“By integrating silicon photonics directly into switches, Nvidia is shattering the old limitations of hyperscale and enterprise networks and opening the gate to million-GPU AI factories,” Huang said.
Also unveiled were Quantum-X photonics switches, which support 144 ports of 800Gb/s InfiniBand and feature a liquid-cooled design to cool the onboard silicon photonics.
Nvidia said its new Quantum-X photonic switches offer 2x faster speeds and 5x higher scalability for AI compute fabrics compared with the previous generation units.
Open Llama Nemotron: A family of open reasoning AI models
Having already unveiled a series of telecom-focused AI models, Nvidia also revealed the open Llama Nemotron family of models, designed to let businesses build AI agents.
Built atop Meta’s series of AI models, open Llama Nemotron supposedly boast reasoning capabilities enabling them to power enterprise applications autonomously on their own or collectively as a group.
The models come in a variety of sizes, ranging from Nano, which can run on PCs and edge devices, to the mid-tier Super, and the mammoth Ultra which is designed to work across multi-GPU servers.
Among the early adopters are Microsoft, which is including the models into its Microsoft Azure AI Foundry and SAP, which is using them to enhance its SAP Business AI solutions and its AI copilot, Joule.
“These advanced reasoning models will refine and rewrite user queries, enabling our AI to better understand inquiries and deliver smarter, more efficient AI-powered experiences that drive business innovation,” said Walter Sun, global head of AI at SAP.
AI Data Platform: A new AI agent design tool
The chipmaker showcased a new customisable reference design platform that lets users build infrastructure to speed AI reasoning workloads with specialised AI query agents.
Nvidia’s new AI Data Platform is designed for storage providers, enabling them to use their vast libraries of information to fuel AI query agents.
It’s brought on board the likes of Dell Technologies, IBM, HPE, and Pure Storage, among others, to develop agentic AI systems that Nvidia suggests can reason and connect to enterprise data.
Huang said the platform will help build a “new class of enterprise infrastructure that companies need to deploy and scale agentic AI across hybrid data centres”.
Dynamo open source library for AI reasoning models
Also unveiled was Nvidia Dynamo, an open source inference software for accelerating and scaling AI reasoning models.
Designed for use across so-called ‘AI factories’ or specialised AI data centres, Dynamo is the successor to Nvidia’s Triton Inference Server.
It orchestrates and accelerates inference communication across thousands of GPUs, using disaggregated serving to “separate the processing and generation phases of large language models (LLMs) on different GPUs” — in simple terms: it speeds up AI responses by efficiently coordinating thousands of GPUs, splitting tasks so that one set of GPUs processes the data while another generates the final output.
Dynamo is fully open source, so anyone can use it (subject to licence terms) and supports PyTorch, SGLang, and Nvidia’s own TensorRTTM-LLM.
It’ll be available via enterprise platforms from the likes of AWS, Dell, Google Cloud, and Microsoft Azure among other platform providers.
A quantum computing research centre
Away from hardware updates and software showcases, Nvidia also announced plans to build a research centre to advance quantum computing research.
The Nvidia Accelerated Quantum Research Centre, or NVAQC, will be based in Boston, Massachusetts and will be tasked with helping solve some of quantum computing’s most challenging problems, from tackling qubit noise to designing experimental quantum chips.
“Quantum computing will augment AI supercomputers to tackle some of the world’s most important problems, from drug discovery to materials development,” Huang said.
NVAQC will work with Nvidia’s commercial and academic partners to use GB200 NVL72 hardware to run complex simulations of quantum systems and develop AI algorithms for use in quantum research.
The site is expected to begin operations later this year, with the EQuS group, a member of the MIT Center for Quantum Engineering, set to use NVAQC to develop techniques such as quantum error correction.
RELATED STORIES
Tech Mahindra partners with Nvidia and AWS on telco AI model
