Jensen Huang’s GTC keynote big takeaway:

We are witnessing another paradigm shift — an Apollo moment for technology.
AI is no longer a tool that aids work; it has become the work itself, solving problems directly rather than enabling humans to do so.

With the end of Moore’s Law, traditional CPU scaling has plateaued. The answer is accelerated computing, powered by GPUs. This gives rise to a virtuous cycle of AI:
more GPUs → more CUDA libraries → more AI breakthroughs → greater demand for GPUs.
Each element reinforces the other, creating compounding progress.

Groundbreaking GPUs like Rubin and Blackwell now train trillion-parameter models in record time. Partnerships with firms such as Nokia could enable AI at the wireless network edge — specialized GPUs operating within RAN base stations, bringing intelligence to telecom infrastructure. This vision connects directly to the broader idea of 6G and AI-powered IoT networks.

At the frontier of quantum computing, specialized GPUs may assist quantum processors (QPUs) in handling one of the field’s hardest problems: error correction. Collaborations with national labs like the DOE could make hybrid quantum–GPU computing practical.

Another innovation addresses AI’s memory limitation — GPUs designed to manage vast context windows, unlocking models that can reason across longer narratives and data sequences.

Beyond computing, the Hyperion platform represents a holistic chassis for autonomous mobility. By partnering with automakers, it aims to enable full self-driving and robo-taxi solutions, directly competing with Tesla.

Similarly, specialized GPUs are emerging for physical AI — robots and humanoids that learn from both real-world and simulated data. The Omniverse enables the creation of entire AI factories, where digital twins of manufacturing systems are simulated before being built in the physical world, from Arizona to Virginia.

Underpinning all of this is a new concept: Co-Design. At the CUDA level — the foundation of GPU programming — NVIDIA is working with partners like Palantir, ServiceNow, Nokia, and Foxconn to co-develop domain-specific CUDA libraries. These libraries don’t just support software; they perform the work itself, redefining productivity and computation.

Accelerated computing, quantum computing, AI on RAN, 6G, IoT, Hyperion, robo-taxis, humanoids, AI factories, and co-designed CUDA systems together mark the next era — one where AI becomes the engine of creation itself.

Theme/Idea	Description	Context/Significance
The AI Industrial Revolution	AI is declared the new industrial revolution, essential infrastructure that every company and nation will use.	Framed as a civilizational leap—America’s next “Apollo moment.”
End of Moore’s Law	Dinard scaling has stopped, slowing the performance and power gains of transistors, necessitating a new computing model.	Explains why Accelerated Computing is now a requirement, not an option.
Accelerated Computing Inflection Point	The 30-year effort to develop accelerated computing using the GPU and CUDA has reached its critical tipping point.	The market is now rapidly adopting GPU-based systems to overcome physical limits.
AI as “Work,” not a “Tool”	AI is not just software (like Excel) but an agent/worker that performs tasks and uses tools, allowing technology to augment labor.	Enables AI to address the $100 trillion global economy for the first time, driving massive productivity gains.
The Virtuous Cycle of AI	A self-sustaining loop where high AI intelligence leads to commercial viability, funding more compute, which makes the AI even smarter.	Ensures the continuous, exponential growth required for the industry to advance.
Full-Stack, Full-Scale Integration	The strategy involves integrating technology across the entire vertical stack (chips, systems, software, cloud services) and horizontally across all industries (telecom, robotics, factories).	Ensures a cohesive, optimized ecosystem that accelerates development everywhere.

New Technologies and Architectures

Category	Concept/Platform Name	Description
AI Factory Infrastructure	The AI Factory	A dedicated computing infrastructure, unlike a general-purpose data center, whose sole purpose is to produce high-value, high-speed tokens cost-effectively.
AI Factory Infrastructure	Omniverse DSX	A digital twin blueprint for designing, simulating, and operating gigascale AI Factories. It allows co-design of the building, power, and compute stack for maximum efficiency.
GPU/System	Grace Blackwell (GB200) NVLink 72	The most extreme co-designed system, delivering 10x generational performance and producing tokens at the lowest cost in the world, validating the co-design approach.
System Future	Vera Rubin (Rubin)	Nvidia’s next-generation, completely cableless, 100% liquid-cooled AI supercomputer rack, offering 100x the performance of the 2016 DGX-1.
Design Methodology	Extreme Co-Design	The method of designing the entire computing stack simultaneously (chip, system, software, model, application) to achieve compounding exponential performance gains.

New Processors and Specialized Platforms

Category	Concept/Platform Name	Description
Networking/6G	Nvidia Arc	A new product line for telecommunications, enabling a software-defined, AI-enabled 6G base station in partnership with Nokia.
Quantum Computing	NVQLink & CUDAQ	NVQLink is an interconnect that links quantum processors (QPUs) and GPUs. CUDAQ is the software platform enabling the two to work together for error correction and hybrid simulations.
Processor for Context	Context Processor (CPX)	A dedicated processor added to the Rubin node to handle the increasing load of massive context processing (ingesting documents, video) before an AI can generate a response.
Processor for Memory	Bluefield 4 (BF4)	A revolutionary data processor designed to manage and accelerate KV Caching (the retrieval of an AI’s long-term conversational memory) to prevent slowdowns.
Robotics Platform	NVIDIA Drive Hyperion	A standardized, sensor-rich computing chassis architecture that serves as the computing platform on wheels for the entire autonomous vehicle industry (e.g., robo-taxis).
Physical AI System	Jetson Thor	The dedicated, high-performance computing system that runs the trained Physical AI model in real-time inside the robot or autonomous vehicle.

New Concepts and Market Shifts

Concept/Idea	Description	Impact
Three New Scaling Laws	Pre-training, Post-training, and real-time Thinking/Inference—all require extraordinary computation and are driving demand.	The shift in model complexity is driving the “two exponentials” of compute demand.
AI on RAN / AI for RAN	The 6G network will host two major AI roles: improving radio efficiency (AI for RAN) and building a new edge industrial robotics cloud (AI on RAN).	Transforms the telecommunications network into an intelligent, edge computing platform.
Physical AI	AI that understands the physical world, including laws of physics, causality, and permanence, required to train and operate embodied agents like robots.	Enables a new market in humanoid robotics and automated factories.
Agentic SAS	The future state of SaaS where Nvidia libraries and AI models are integrated to transform major enterprise workflows into systems where AI agents perform work.	Deeply embeds AI into the world’s largest enterprise software (e.g., SAP, ServiceNow).

Co-Design: Jensen Huang identified Extreme Co-Design as the only answer to the convergence of two major challenges: the end of Moore’s Law and the exponential compute demands of modern AI. Nvidia’s foundational architecture (Extreme Co-Design) is now being scaled and applied to critical, multi-trillion-dollar sectors via deep, vertical partnerships.

Palantir: Accelerating Operational Intelligence

Partnership Focus	Detail	Strategic Value
The Challenge	Extracting deep, timely business insight from massive volumes of complex structured and unstructured data, often for government and mission-critical operations.	Accelerating human judgment and turning it into dynamic decision intelligence at an extraordinary scale.
The Goal	To accelerate everything Palantir does in its Ontology framework and AIP (Artificial Intelligence Platform).	Provides the advanced, context-aware reasoning necessary for complex operational AI.
The Integration	Integrating Palantir’s Ontology (the core data framework) with NVIDIA’s accelerated computing, CUDA-X libraries, and open-source Nemotron models.	Enables faster data processing, analytics, and the creation of long-thinking, reasoning agents within the Palantir platform.
Outcome	Creating a next-generation integrated technology stack to accelerate the end-to-end AI pipeline for large-scale enterprise and government customers (e.g., logistics for Lowe’s).	Ensures that data processing is done at “speed of light” and “extraordinary scale,” from data ingestion to production AI.