Think of it as Yoga for the brain, we are going to stretch a bit.
The World Is Not Enough – NVIDIA and Microsoft
The following is my take on what NVIDIA is doing, based on the NVIDIA GTC Taipei 2026 keynote by founder and CEO Jensen Huang. This is not just another technology announcement. It represents a major shift in NVIDIA’s direction, its partnership with Microsoft, and the way you should think about computers, AI, work, and your own future. The complete video is at the bottom.
NVIDIA and Microsoft Are Reinventing What You Think a Computer Is
We used to think a computer was a box. A CPU, a GPU, some
memory, a hard drive, a keyboard, a screen, and a poor human sitting in front of it trying to remember where he saved the file. Computer were something you turned on and turned off. That was the old world. Not anymore!
In his presentation Jensen Huang is describing something much bigger than GPU and AI. NVIDIA is no longer talking about a computer as one machine. He is talking about a factory that manufactures intelligence.
Not software.
Not documents.
Not spreadsheets.
Not search results.
Tokens.
Tokens are now the building blocks of useful AI. They are becoming units of revenue, units of productivity, and maybe one day units of civilization itself. Jensen’s point is simple: if AI can produce useful work, then tokens are not just computer output anymore. They are economic output.
That is why he says compute is revenue.
And once compute becomes revenue, the old computer is not enough.
The world is not enough.
NVIDIA is introducing a whole new kind of computer built for agentic AI. Not just a chatbot that answers questions, but agents that observe, reason, plan, use tools, remember things, write code, open databases, create CAD files, generate graphics, and act on your intent.
In the old computer model, you opened an application. You clicked a button. You typed a command. The software sat inside the operating system like a clerk waiting for orders.
In the new model, the “application” is an agent.
The agent has a large language model for a brain, a harness for a body, tools like Python, JavaScript, SQL, web browsers, spreadsheets, databases, CUDA libraries, and a runtime that holds the whole thing together.
That is not just a better computer.
That is a digital worker.
Jensen gave examples that make the point. A person can tell the AI to create an animation with NVIDIA green dots forming Taipei 101, morphing into the NVIDIA logo, then scattering again. The AI writes the code and creates it. A person can show it a broken remote-control battery clip and say, “Make me a CAD file.” The AI uses tools and creates a file ready for 3D printing.
That is the change.
We are moving from clicking software to commanding intelligence.
And this is where Microsoft comes in. Microsoft already understands the same future through Copilot, Codex, enterprise agents, and the larger OpenAI ecosystem. In Jensen’s presentation, Microsoft is also shown as part of this new world: one of the early companies with an operational Vera Rubin NVL72 engineering rack, and one of the names connected with adoption of NVIDIA OpenShell.
That matters.
Because Microsoft owns the office. NVIDIA owns the factory. OpenAI, Anthropic, and others are building the brains. And together they are turning the old idea of a computer inside out.
NVIDIA’s big hardware introduction is Vera Rubin.
But Vera Rubin is not just a GPU. Jensen makes that very clear. It is not one chip. It is not one board. It is not even one rack.
It is a complete agent-processing system.
Vera Rubin NVL72 does the heavy thinking: prompt processing, context understanding, reasoning, planning, and token generation. It uses the new Vera Rubin GPU, described as having six trillion transistors and more than 18,000 components on one board. It is built with three-nanometer process technology, CoWoS advanced packaging, and HBM4 memory from Micron, SK Hynix, and Samsung.
Then there is the Vera CPU rack: 256 liquid-cooled CPUs designed to orchestrate the models, move memory, and launch tools. Jensen’s argument is that old CPUs were built for humans, but agents are different. Humans wait in seconds. Agents wait in nanoseconds. If an agent is waiting for a database, a compiler, a Python process, or a tool call, then the expensive GPU is sitting there idle.
So NVIDIA built Vera as a CPU for agents.
Then comes Vera BlueField-4 STX, the storage and security system. Jensen says this is where AI keeps its memory. That matters because agents are not just answering one question at a time. They need working memory, long-term memory, context, retrieval, structured data, unstructured data, and security.
Then comes ConnectX-9, BlueField DPUs, Spectrum-X Ethernet Photonics, and co-packaged optics. In plain English, NVIDIA is not just making faster chips. It is redesigning the nervous system of the data center.
This is why Jensen keeps returning to the same point:
The future computer is disaggregated, distributed, and heterogeneous.
The brain may be in one place.
The memory may be in another.
The tools may run somewhere else.
The security processor watches over it.
The CPU orchestrates it.
The GPU thinks.
The network connects the whole beast together.
That is not a desktop computer. That is an industrial machine for manufacturing intelligence.
Then he introduces DSX, which is NVIDIA’s blueprint for AI factories. RTX was for graphics. DGX was for systems. DSX is for infrastructure. It includes DSX Sim, built with Omniverse, so companies can simulate an AI factory before they build it. They can test power, cooling, network design, rack layout, and integration inside a digital twin before spending tens of billions of dollars in the real world.
That is where the money gets serious.
Jensen talks about one-gigawatt AI factories costing $50 billion, $60 billion, and eventually maybe $80 billion to $100 billion per gigawatt. When the investment is that large, a wrong design is not a mistake. It is a financial crater.
So NVIDIA is selling more than chips. It is selling the architecture of the new industrial age.
Then comes the software side: the NVIDIA Agent Toolkit for Enterprise AI.
This is where the computer becomes less like Windows and more like a secure operating system for digital employees. Jensen says every company will become an agent company. Every company will need agents. And every company will ask the same question:
How do we run agents safely?
The toolkit has four parts: Models — large language models, open models, modifiable models, the smarter, faster, cheaper brains. Harnesses — systems that orchestrate the agent.
Runtime — the operating environment that keeps the agent secure, grounded, and controlled. Then he names NVIDIA OpenShell.
That one is important. OpenShell is presented as a secure enterprise harness where agents like Claude Code and Codex can run safely. It protects identity, privacy, permissions, and security policy. In other words, OpenShell is not just “run an AI.” It is “let the AI work inside your company without letting it burn down the building.”
That is the enterprise problem. A regular person may ask AI to write a poem. A company wants AI to access code, databases, financial records, customer files, engineering drawings, and internal systems. That requires identity, permissions, memory, audit trails, and security. That is what OpenShell is trying to solve.
Jensen also mentions agents and harnesses like Claude Code, Codex, OpenClaw, Hermes, and NVIDIA’s own Nemotron models. He points to a partnership with Cadence to build chip-design super agents, where Codex or Claude Code can orchestrate RTL verification, testbench creation, regression testing, and debugging.
That is not a toy. That is AI entering one of the hardest engineering jobs on earth: chip design.
And the most interesting part is this: Jensen is not saying AI replaces tools. He says the opposite. Agents will use more tools than humans ever did. CUDA-X libraries become tools for agents. cuLitho, cuOpt, cuDSS, AI-Q, Aerial, PhysicsNeMo, Parabricks — these become specialized instruments that an agent can learn to use.
So the future is not one giant AI brain floating in the cloud. The future is millions or billions of agents using millions of tools. A digital workforce. A new kind of economy.
And Microsoft wants that workforce inside Windows, Office, Azure, GitHub, and enterprise software. NVIDIA wants to build the factories that power it. OpenAI and others want to build the minds. The cloud companies want to rent it. Every business will eventually try to deploy it.
This is why the phrase “computer” is starting to feel too small. A computer used to be something you owned.
Now it may be something you command. A computer used to run programs.
Now it may run workers. A computer used to sit on a desk.
Now it may be a one-gigawatt factory full of liquid-cooled racks, optical networking, DPUs, CPUs built for agents, GPUs built for reasoning, and software that treats tokens like manufactured goods.
The old computer helped you do work. The new computer does work.
And once that happens, the world we built around the old computer is no longer enough.
The office is not enough. The desktop is not enough.
The app is not enough. The cloud is not enough.
The world is not enough.
Because NVIDIA and Microsoft are not just reinventing the computer.
They are reinventing the worker, the factory, the company, and eventually the economy itself.
And like all revolutions, most people will not notice it at first.
They will call it a chatbot. They will call it a better search engine. They will call it hype.
Then one morning they will wake up and discover that the computer no longer waits for instructions. It has become the thing that gives instructions.
Next few days expect Apples response . They can't let Microsoft have all the fun.
Software / AI / Platform Technologies
Technology
Type
Description
Agentic AI / Agents
AI computing model
The new application pattern: instead of launching apps and clicking, you give intent to an agent that observes, reasons, plans, uses tools, and acts.
Large Language Model + Harness + Tools + Runtime
Agent architecture
Jensen describes the agent as a model “brain,” a harness/body that orchestrates, tools/skills, and a runtime/workshop.
Working memory / KV caching
AI memory system
The short-term memory system for agents; Jensen says memory/retrieval will revolutionize storage.
CUDA-X Libraries
NVIDIA software libraries
NVIDIA’s CUDA libraries repackaged as tools agents can learn to use.
cuLitho
CUDA-X library
Computational lithography tool/library.
cuOpt
CUDA-X library
Decision-optimization library for planning/optimization workloads.
cuDSS
CUDA-X library
Direct sparse solver library.
AI-Q
CUDA-X / AI tool
Deep research tool across structured and unstructured documents.
Aerial
CUDA-X / telecom AI
AI-RAN platform/tool for radio access networks.
PhysicsNeMo
CUDA-X / science AI
Differentiable physics tool/library.
Parabricks
CUDA-X / genomics
Genomics acceleration library.
DOCA
NVIDIA software stack
Software stack tied to ConnectX/BlueField infrastructure, used in the Vera Rubin architecture.
Confidential computing
Security architecture
Security model where data/model are encrypted at rest, in motion, and in use.
NVIDIA DSX
AI factory infrastructure platform
Reference design/blueprint for building and operating AI factories.
DSX Sim
Simulation / planning software
Omniverse-based blueprint for designing, validating, and simulating AI factories before racks are ordered.
Omniverse
Digital twin / simulation platform
Used to build and simulate AI factory systems digitally before physical construction.
DSX OS
AI factory operating software
Provisions, operates, monitors, and remediates AI factory infrastructure.
DSX MaxLPS
Power optimization software
Lets operators deploy more GPUs inside the same power budget and balance power/cooling.
DSX Flex
Grid-interaction software
Reads real-time grid signals and adjusts AI factory power draw when the grid needs relief.
NVIDIA Agent Toolkit for Enterprise AI
Enterprise AI platform
Toolkit for building/running enterprise agents: models, harnesses, tools/skills, and runtime.
NVIDIA OpenShell
Secure agent harness/runtime
Open-source enterprise shell that protects agent identity, privacy, permissions, and security policies.
Claude Code
Coding agent
Used as an example of an agent that can generate code and operate inside the new agentic pattern.
Codex
Coding agent
Another coding agent Jensen names; also used to orchestrate Cadence chip-design workflows.
OpenClaw
Agent / harness
Named as an agentic harness that can run on-prem or anywhere.
Hermes
Agentic harness
Named as another agentic harness; also shown in the RTX Spark design demo.
Claude Sonnet
Cloud AI model
Used in the RTX Spark demo, connected through OpenShell/Hermes.
Nemotron
NVIDIA open model family
NVIDIA’s open model family for building agents.
Nemotron 3 Ultra
Open AI model
Newly announced open model; described as five times faster, 30% cheaper, and based on hybrid SSM + Mixture-of-Experts architecture.
Nemotron 4
Future model
Mentioned as currently being worked on / coming after Nemotron 3.
State Space Models + Mixture of Experts
Model architecture
The hybrid architecture behind Nemotron 3 Ultra.
Cadence Super Agents / Design Verification Agent
Chip-design AI agents
Agents for RTL generation, testbench creation, regression testing, simulation, and debugging.
Cadence Chip Stack
EDA workflow platform
Launches the RTL verification loop, powered by Nemotron and secured by OpenShell.
Cadence Xcelium
Simulation tool
Used by Chip Stack agents to run simulations.
JasperGold
Formal verification tool
Used for formal verification in the Cadence/NVIDIA chip-design workflow.
RTX Spark agent platform
PC agent platform
Local agent platform for running personal AI agents on new NVIDIA/Microsoft PCs.
Windows platform for agents
PC operating-system direction
Microsoft/NVIDIA collaboration to make Windows PCs run local/cloud agents natively.
Agentic runtime on PC
PC software model
Replaces the old app model with local agents connected to local or cloud models.
MCP server
Agent-tool interface
Adobe Photoshop/Premiere are described as becoming agent-friendly through an MCP server.
Adobe Photoshop / Premiere re-engineered for RTX Spark
Creative software integration
Adobe is reworking the core architecture for RTX Spark, making it faster and agent-interactive.
Rhino
CAD/design tool
Used by a local RTX Spark agent to model a house/site.
Blender
3D/rendering tool
Used in the RTX Spark design workflow to render the house.
Flux 2
Generative image model
Used to make Blender house renders photorealistic in the RTX Spark demo.
Cosmos 3
Physical AI foundation model
New open frontier model for physical AI; can understand, reason, generate, simulate, and act as policy.
Mixture of Transformers
Model architecture
Cosmos architecture combining autoregressive transformer reasoning with diffusion transformer generation.
VLM / world model / simulator / action model
Physical AI roles
Cosmos is described as a vision-language model, world model, simulator, and post-trainable action model.
NVIDIA OmniDreams
Action-conditioned world model
Built on Cosmos; predicts future frames for physical AI.
Alpamayo 2 Super
Autonomous vehicle model
Open model for self-driving cars; described as the world’s first reasoning autonomous vehicle.
NVIDIA DRIVE Hyperion runtime
Autonomous vehicle runtime
Runtime for deploying Alpamayo 2 Super / NVIDIA driving stack in cars.
Halos operating system
Vehicle operating system
Operating system used with DRIVE Hyperion runtime for autonomous driving deployment.
NVIDIA Isaac GR00T software stack
Humanoid robotics platform
Humanoid stack including model, data generation, simulation, runtime, and OS.
Isaac Lab
Robotics simulation/training
Simulation environment for GR00T humanoid robotics work.
Isaac Teleoperation
Robotics data capture
Captures demonstrations from real or simulated robots.
Isaac Lab Arena
Robotics evaluation
Used to train and evaluate robot policies.
Isaac ROS
Robotics runtime/deployment
Used to deploy robot policies on Jetson Thor.
Hardware / PC Equipment / AI Factory Equipment
Equipment
Type
Description
Vera Rubin
Full AI supercomputer system
NVIDIA’s new disaggregated, distributed agent-processing system; not just one GPU.
Vera Rubin NVL72
Rack-scale AI system
Handles prompt/context understanding, reasoning, and planning for agentic AI.
Vera Rubin GPU
GPU
Described as six trillion transistors with over 18,000 components on one board.
Vera CPU
CPU
CPU built for agents instead of humans; optimized for low latency, bandwidth, and efficiency.
Vera CPU Rack
CPU rack
256 liquid-cooled CPUs in one rack for model orchestration, memory shuffling, and tool launching.
NVIDIA Olympus Core
CPU core
Custom data-center CPU core inside Vera; used for modern data-center/agent workloads.
LPDDR5 / LPDDR5X memory
CPU memory
High-bandwidth memory used by Vera CPU; transcript says Vera uses LPDDR5X while correcting multiple errors.
PCIe Gen 6
I/O bus
Vera is described as the first CPU to use PCIe Gen 6.
NVLink chip-to-chip
Interconnect
Connects GPUs directly to CPU and scales Vera across multiple sockets.
ConnectX-9
Network adapter / NIC
Part of Vera Rubin infrastructure and storage/network connectivity.
BlueField / BlueField-4 DPU
Security / data processing unit
Security processor for isolation and confidential computing; BlueField-4 DPUs appear in the Vera Rubin system.
Vera BlueField-4 STX
Storage/security system
“Where AI keeps its memory”; accelerates storage processing and connects memory, storage, and in-silicon security.
SuperNICs
Networking hardware
Used with ConnectX-9 and BlueField-4 DPUs for AI factory scaling.
NVLink switch trays
Rack interconnect hardware
Nine hot-swappable NVLink switch trays are part of the Vera Rubin NVL72 rack design.
Spectrum-X Ethernet Photonics
Optical networking switch
Described as the world’s first Ethernet switch with 200Gb co-packaged optics.
Co-packaged optics / CPO
Optical network packaging
Used in Spectrum-X Ethernet Photonics, with high-powered laser dies.
Vera LPX Rack / Groq LPX Rack
Low-latency inference rack
Uses 256 Groq LPUs across 16 trays for ultra-low-latency token generation.
Groq LPUs
AI inference processors
Low-latency processors used in the LPX rack.
Third-generation MGX Rack
Rack architecture
Vera Rubin rack with 1.3 million components, compute trays, NVLink switches, and liquid-cooled bus bars.
Modular compute tray
Server tray
New tray design with PCB midplane and no-cable maintenance access.
PCB midplane
Rack interconnect hardware
Replaces large cable complexity in Vera Rubin racks.
Liquid-cooled bus bars
Power/cooling hardware
High-efficiency bus bars carrying over 5,000 amps inside the rack.
TSMC 3nm process
Chip manufacturing process
Used for Vera Rubin and RTX Spark chips.
CoWoS advanced packaging
Chip packaging
Advanced packaging used for Vera Rubin chips.
HBM4 memory
GPU memory
Memory from Micron, SK Hynix, and Samsung used in Vera Rubin.
Grace Blackwell NVLink 72
Prior rack-scale AI system
Used as the previous-generation rack example for LLM thinking/inference.
Hopper / Ampere / Pascal
Prior GPU generations
Mentioned as earlier NVIDIA architectures in the evolution toward Vera Rubin.
DGX / DGX-1
AI system line
DGX is referenced as NVIDIA’s systems line; DGX-1 as the first AI supercomputer.
RTX Spark
New PC chip/platform
New NVIDIA/Microsoft PC platform for local agents. Includes Blackwell RTX GPU, Grace CPU, NVLink, unified memory, and Windows agent platform.
RTX Spark laptops
New PC device class
Jensen shows RTX Spark laptops as part of the new PC reinvention.
RTX Spark desktop
New PC device class
Transcript mentions an RTX Spark desktop, including an MSI example.
RTX Spark workstation
New PC device class
Part of the three-machine Windows lineup: desktop, laptop, and workstation.
Blackwell RTX GPU
PC GPU
RTX Spark includes a Blackwell RTX GPU with 6,144 Tensor Cores.
6,144 Tensor Cores
AI acceleration cores
Tensor cores in the Blackwell RTX GPU inside RTX Spark.
Custom 20-core Grace CPU
PC CPU
Built with MediaTek and fused with GPU by NVLink in RTX Spark.
128 GB unified memory
PC memory architecture
Unified memory in RTX Spark.
70 billion transistors
Chip scale
RTX Spark chip transistor count given in the transcript.
NVIDIA AI Tensor Core PCs
PC platform category
New Windows machines are described as 100% CUDA and 100% NVIDIA AI Tensor Core.
GeForce with Tensor Cores
Consumer GPU hardware
Mentioned as another platform where agentic AI can run.
NVIDIA DRIVE Hyperion systems / cars
Autonomous vehicle hardware platform
Vehicle platform for running Alpamayo 2 Super and NVIDIA driving stack.
Jetson Thor
Robotics computer
New robot computer used by Isaac GR00T reference humanoid robot and Isaac ROS deployment.
NVIDIA Isaac GR00T reference humanoid robot
Humanoid robot hardware platform
Fully integrated reference robot: 25 degrees of freedom per hand, 31 robot degrees of freedom, 6 feet, 150 pounds.
Sharpa robotic hands
Robot hardware component
Hands used in the Isaac GR00T reference humanoid robot.
Humanoid robotics computers / self-driving car computers / satellites / base stations
Edge/physical AI equipment
Jensen says the agentic computing pattern will run across robots, cars, satellites, base stations, factories, agriculture, manufacturing, and heavy industry.
Clean takeaway
The core stack he is pushing is:
Vera Rubin for AI factories, Vera CPU for agent orchestration, NVIDIA Agent Toolkit + OpenShell for enterprise agents, Nemotron 3 Ultra for open agent models, RTX Spark for the reinvented PC, Cosmos 3 for physical AI, Alpamayo 2 Super for self-driving cars,
and Isaac GR00T + Jetson Thor for humanoid robots.
