NVIDIA GTC Washington Keynote

Earlier today in Washington the story of technology’s great American arc was retold: from the Macintosh’s arrival to the dawn of the internet, from datacentres rising across the plains to the promise of a future still to come. On stage, Jensen Huang opened with that sweep of history - Apple, the Mac, the bold entrepreneurial leaps of Silicon Valley - and then made this claim: “Artificial Intelligence: the new industrial revolution”. He elaborated with grand ambition: “We will soon power it all with unlimited clean energy”, “extend humanity’s reach to the stars”, “America’s next Apollo moment”. The language was large; the intent, surgical.

From proclamation to plumbing: the new computing model

Huang’s central claim was not merely thematic: it was architectural. The problems ahead - massive models, near-real-time context processing, robotics that must perceive and act - are not a better use case for yesterday’s general-purpose processors. They require a fundamentally different programming and system model: accelerated computing. In Huang’s telling, the GPU is not simply a vendor component; it is the heart of a new system architecture. But silicon needs software to become useful. That software, he said, is the company’s “treasure”, the CUDA-X ecosystem of libraries and microservices. NVIDIA’s own pages now describe hundreds of GPU-accelerated libraries (NVIDIA currently cites 400+ CUDA-X libraries) that span domains from data processing and inference to quantum workflows, the scaffolding that turns raw FLOPS into production answers.

Huang signalled that this is not incremental tuning; it is a shift to a stack-first industry where the hardware and software are designed as a single, deployable system.

Telecom reborn

If you expected announcements only about chips and racks, Huang had other plans. He announced a sweeping partnership with Nokia, framed as a move to pioneer the next era in telecommunications, a market Huang characterised in the keynote as worth trillions, and introduced NVIDIA ARC (Aerial RAN Computer): an accelerated RAN architecture that couples NVIDIA compute (Grace and Blackwell CPUs/GPUs) with Mellanox networking and CUDA-X software so RAN workloads and AI inference can run together on the same platform. In simple terms: base stations become local AI nodes that can sense, infer and act at the edge rather than simply passing bits back to a distant cloud.

Why this matters

It reframes telecoms as compute infrastructure, not just transport. Operators planning 6G or edge compute need to think about software lifecycle, racks in radio sites and the economics of on-site inference.

Quantum - now with classical scaffolding (NVQLink)

Huang made a clear, strategic bet: quantum’s early promise will only be realised if quantum processors live alongside classical accelerators that handle control, calibration and error correction. To that end he announced NVQLink, a high-speed interconnect and open architecture to tightly couple quantum processors to GPU-based supercomputers. The pitch is practical: quantum devices generate noisy outputs and require low-latency classical processing for error correction and control loops. NVQLink is being rolled out with a group of quantum builders and US national labs, and it was presented as the connective tissue for accelerated quantum-classical workflows.

If you’re in scientific computing or chemistry modelling, NVQLink suggests a pathway to hybrid workflows where quantum experiments and GPU-based simulations feed each other in real time.

The AI factory

Scale was the keynote’s metronome. Huang framed a new unit of compute not as the server but the rack: enter the NVL72 “Thinking Machine”, a rack-scale system that unifies 72 Blackwell GPUs and an array of Grace CPUs into a single, liquid-cooled NVLink domain. Huang touted step-change performance, Grace/Blackwell pairings that deliver substantially faster inference and lower cost per token than prior generations, and positioned the NVL72 as the architectural “atom” of the AI factory.

But scale isn’t only about compute density; it’s about design and operation. Huang unveiled Omniverse DSX — a blueprint for gigascale AI factories — a digital-twin platform for modelling, simulating and optimising gigawatt-scale facilities before a single breaker is installed. The DSX blueprint folds power, cooling, rack layout and workload orchestration into a single simulation environment so operators can balance performance, resilience and sustainability at design time.

Huang argued - convincingly - that the industry has “turned a corner.” Models are now both more capable and more widely used; they have crossed a threshold of usefulness that people and businesses will pay for. That creates a demand-side imperative: more users, more contexts, more queries → far higher compute and infrastructure needs. His economics slide emphasised token cost: the Grace-Blackwell NVL72 family, Huang said, generates the lowest cost per token for large-scale inference, which is central to the unit economics of real-time AI services.

NVIDIA Rubin Platform: The Next AI Era

NVIDIA unveiled the Rubin Platform at GTC Washington, marking the successor to Blackwell. Expected in 2026, Rubin promises a massive performance leap, delivering over 3x the AI acceleration of current systems, especially for complex, massive-context applications like large language models.  
A key innovation is in its physical build: future Rubin systems, like the CPX variants, leverage midplane architecture and advanced liquid cooling. This minimises  external cable clutter by designing components that fit more directly and densely into the server chassis, essentially creating a more cable-less, integrated board-like deployment. This shift maximises compute density and efficiency for the coming generation of AI factories.

The Department of Energy, seven supercomputers and national scale

Huang closed the loop between product and policy as NVIDIA announced partnerships with the US Department of Energy and national labs, including plans for multiple new systems - seven AI supercomputers - intended to accelerate national research and to anchor “AI factory” capacity in US labs and institutions. The message was explicit: industrial-scale AI is strategic, and scale will be built with public-private collaboration.

Get in touch with one of Boston’s experts

As an NVIDIA Elite Partner, Boston Limited is already engaged with customers on a number of these vectors. Get in contact with one of the experts at Boston to discover how we can help you unlock your full potential with Boston and NVIDIA.

Tags: nvidia, gtc, washington, keynote, jensen huang, nvidia gtc, keynote announcements