NVIDIA AI Explained: GPUs, CUDA, and the Software Stack That Powers Modern AI
Almost every AI system you have heard of — the chatbots, the image generators, the coding assistants, the robots learning to pick up boxes — runs, at some point, on NVIDIA. The company does not make the models. It makes the machines the models are trained on, the software the models are built with, and the systems that serve the models to users. In the quarter that ended on 25 October 2026, NVIDIA took in $115.8 billion in revenue, and $107.6 billion of it — about 93 cents in every dollar — came from its data-centre business. That is the scale of the company this guide explains: not a chipmaker that got lucky, but the company that sells the picks and shovels of the AI boom, and then sells the software that makes the picks and shovels work. Checked 20 November 2026 against NVIDIA’s own documentation, read in full, and against the company’s most recent earnings release.
The plain-language answer is this. NVIDIA dominates AI because it sells a stack, not a chip. At the bottom is the GPU, a processor designed to do thousands of small calculations at once — exactly what AI maths needs. On top of that sits CUDA, a software platform NVIDIA has been building since 2007 that lets developers program those GPUs. On top of CUDA sit specialist libraries, inference servers, enterprise software, whole rack-scale systems, and now robotics platforms. Each layer makes the layer below it more useful, and each layer makes it harder for a customer to leave. A rival can build a faster chip. It cannot, overnight, rebuild eighteen years of software, the millions of developers trained on it, and the racks, networks and tools built around it. That is the whole story in one paragraph; the rest of this guide is the map of how the stack fits together.
Two corrections belong at the top, because most explainers written even a few weeks ago get them wrong. First, the next generation of NVIDIA hardware is no longer coming — it is here. Vera Rubin systems are already operating at CoreWeave, Google Cloud, Microsoft Azure, Oracle Cloud and Nebius, NVIDIA confirmed on its earnings call of 17 November 2026. Any guide that still calls Rubin a 2026 promise is out of date. Second, the company changed leaders eight days ago. Jensen Huang, who co-founded NVIDIA in 1993 and led it for 33 years, resigned as president, chief executive and director on 12 November 2026. Chief financial officer Colette Kress is interim chief executive while the board searches for a permanent successor, and the board has said the product roadmap is on track. This guide is about the stack, not the boardroom, but a guide that still presents Huang as chief executive would be wrong on its first page.

What NVIDIA Actually Is — Not Just a Chip Company
<img src="

” alt=”A layered stack of building blocks, silicon at the base rising through software layers to a robot at the top, one company one stack. No text.”>
Call NVIDIA a chip company and you will be right but misleading, in the same way that calling a carmaker a steel company is right but misleading. The chips are the visible part. The business is the stack: hardware, software, systems and services that together let a customer go from an idea to a working AI product without leaving NVIDIA’s world. Understanding that stack is the whole point of this guide, so here it is in one picture, bottom to top.
At the bottom is the silicon: the GPUs themselves, in generations called Hopper, Blackwell, Blackwell Ultra and now Vera Rubin. Above the silicon is CUDA, the programming platform that turns a graphics processor into a general-purpose computer. Above CUDA are the specialist libraries — cuDNN, TensorRT, NCCL and others — that do the heavy lifting for AI frameworks. Above those are the serving tools that run finished models at scale: Dynamo, the renamed Triton, and the pre-built NIM containers. Above those is the enterprise layer: NVIDIA AI Enterprise, the NeMo framework for building custom models, and DGX Cloud, a managed platform. And at the top, increasingly, are the systems themselves — whole racks of GPUs wired together as a single computer, sold as DGX and HGX systems and as “AI factories” — plus a robotics platform, Isaac, for the physical AI that NVIDIA says is its next great market.
Why does the stack matter more than any single product? Because every layer reinforces the others. A developer who learns CUDA writes code that runs best on NVIDIA hardware. A company that builds its training pipeline on NVIDIA’s libraries finds it expensive to move. A cloud provider that buys NVIDIA’s racks also buys NVIDIA’s networking, because the two are designed together. And a model that is served through NVIDIA’s NIM containers runs, by default, on NVIDIA GPUs. None of this is a conspiracy; it is what happens when one company builds every layer of a technology for almost two decades while its competitors build one layer at a time.
That is also why the numbers look the way they do. NVIDIA’s data-centre revenue of $107.6 billion in a single quarter is larger than the entire annual revenue of most technology companies. The company’s other business — the Edge Computing segment, which includes the GeForce graphics cards that gamers buy — brought in $8.2 billion in the same quarter, respectable but small beside the data-centre figure. The AI stack is the company now. Gaming, the business NVIDIA was founded on, is a side line. If you want to understand modern AI, you have to understand this stack, because the stack is where most AI actually runs. If you are new to how AI works at all, our plain-language guide to how AI works is the place to start; this guide assumes you know what a model is and want to know what it runs on.
The Short Answer: Why NVIDIA Dominates AI

Ask why NVIDIA dominates AI and you will get three answers, and all three are true at once. The first is the hardware: NVIDIA’s GPUs really are very good at AI maths, and the company has shipped a new generation roughly every two years for a decade. The second is the software: CUDA, the libraries and the serving tools form a platform that millions of developers already know. The third is the system: NVIDIA sells not just chips but whole racks, networks and software that turn thousands of GPUs into one machine. Rivals compete on one of these at a time. NVIDIA competes on all three, at once, and that is the moat.
The hardware advantage is real but it is the weakest of the three, and NVIDIA knows it. A GPU is a processor with thousands of small cores that can all work at the same time. AI training is mostly giant matrix multiplications — the same operation repeated millions of times — and that is exactly what thousands of small cores are for. A traditional CPU, with a handful of powerful cores, does the same work far more slowly. NVIDIA did not invent this idea, but it spent fifteen years optimising for it: specialised Tensor Cores for AI maths, huge amounts of high-bandwidth memory sitting next to the processor, and interconnects that let thousands of GPUs behave as one. The result is that the largest AI training runs in the world are built on NVIDIA hardware, and the companies running them keep buying more.
The software advantage is the one that actually locks customers in. CUDA launched in 2007, when AI was a research curiosity and NVIDIA was still mostly a gaming company. Over eighteen years, universities taught it, frameworks were built on it, and an entire generation of AI engineers learned to think in it. Today, if you want to run the standard AI frameworks — PyTorch, TensorFlow, JAX — on a GPU, the path of least resistance runs through CUDA. A rival chip needs not just matching hardware but matching software, matching libraries, matching documentation and matching developer habits. AMD’s ROCm platform is the serious attempt, and it has improved a great deal, but it is still catching up to a platform that has had an eighteen-year head start.
The systems advantage is the one most people miss. The unit of purchase in AI is no longer the chip; it is the rack, and increasingly the whole data centre. NVIDIA’s GB200 NVL72 puts 72 GPUs and 36 CPUs in a single rack, wired together with NVLink so fast that the whole rack behaves like one giant GPU. The GB300 NVL72 does the same with the newer Blackwell Ultra chips, and the Vera Rubin NVL72 does it again with the Rubin generation. The networking — InfiniBand, Spectrum-X Ethernet, the ConnectX network cards, the BlueField data-processing units — is NVIDIA’s too, designed to move data between GPUs without bottlenecks. A customer buying a rival chip has to assemble the rest of the system from other vendors and make it all work together. A customer buying NVIDIA buys the whole thing, tested, from one vendor.
None of this makes NVIDIA unbeatable. The moat has real cracks, and this guide will take them seriously in the competition section: AMD’s MI400 series is a genuine rack-scale rival, Google’s TPUs and Amazon’s Trainium chips are serious inside their own clouds, and the largest AI customers — the very companies that buy the most NVIDIA hardware — are the ones most motivated to build alternatives, because they do not want to depend on a single supplier. But the moat explains the present. As of November 2026, NVIDIA still sells most of the world’s AI accelerators — one analyst estimate put its share at roughly 70% of the AI chip market in May 2026 — and its data-centre revenue doubled in a year. That is what dominance looks like, and it is built on the stack, not on any single chip.
The Hardware Layer: Hopper, Blackwell, Blackwell Ultra and Vera Rubin

NVIDIA names its data-centre GPU generations after scientists, and the names are the quickest way to read the market. Hopper, named after Grace Hopper, powered the first great wave of large-language-model training; the H100 became the most famous chip in AI. Blackwell, named after the statistician David Blackwell, succeeded it in 2024 and is the generation most AI runs on today. Blackwell Ultra, the tuned-up Blackwell, is the current flagship. Vera Rubin, named after the astronomer who discovered dark matter, is the generation after Blackwell — and as of November 2026, it is shipping.
Here is the ladder as it stands on 20 November 2026. Hopper’s H100 is still in service everywhere, but it is last generation. Blackwell’s B200 is the volume chip: NVIDIA’s DGX B200 system packs eight Blackwell GPUs into one server, with 1,440 GB of high-bandwidth memory across the eight, and delivers 144 petaflops of FP4 tensor performance with sparsity (72 without) — roughly three times the training performance and fifteen times the inference performance of the equivalent H100 system, by NVIDIA’s own measurements. Blackwell Ultra’s GB300 NVL72 is the rack-scale flagship: 72 Blackwell Ultra GPUs and 36 Grace CPUs in one liquid-cooled rack, delivering about one and a half times the AI performance of the GB200 NVL72 that came before it. And the Vera Rubin rack — announced as the NVL144 at GTC 2025 and renamed NVL72 in early 2026, when NVIDIA stopped counting the two compute dies inside each Rubin package separately — pairs 72 Rubin GPUs with 36 Vera CPUs. It carries 20.7 terabytes of HBM4 (288 GB per GPU), plus up to 54 terabytes of CPU memory, and NVIDIA’s own figures put the rack at about 3.3 times the performance of the GB300 NVL72.
The table below gives the shape of the ladder. Read it with the caution in the note underneath: these are the manufacturers’ own peak figures, measured at different precisions and with different assumptions about sparsity, and they are not directly comparable to one another — let alone to a rival’s figures. What the table does show, honestly, is direction: every generation brings more memory, more bandwidth, and more low-precision compute, because those are the three things AI workloads actually consume.
| Generation | Flagship system | GPUs per rack | Memory | Headline compute | Status, Nov 2026 |
|---|---|---|---|---|---|
| Hopper | H100 / DGX H100 | 8 per server | 80 GB HBM3 per GPU | ~2 PFLOPS FP8 (dense) per GPU | Last generation, still widely in service |
| Blackwell | B200 / GB200 NVL72 | 72 per rack | 13.4 TB HBM3e per rack | 1,440 PFLOPS FP4 (sparse) per rack | Volume generation |
| Blackwell Ultra | B300 / GB300 NVL72 | 72 per rack | ~20 TB HBM3e per rack (as listed) | ~1.5x GB200 AI performance | Current flagship, shipping |
| Vera Rubin | Rubin / Vera Rubin NVL72 | 72 per rack | 20.7 TB HBM4 per rack (+ up to 54 TB CPU memory) | ~3.3x GB300 (NVIDIA figure) | Shipping since H2 2026 |
| Rubin Ultra | Rubin Ultra NVL576 | 576 per rack system | 365 TB (NVIDIA figure) | ~14x GB300 (NVIDIA figure) | Due H2 2027 (NVIDIA roadmap) |
Two things about this ladder deserve a closer look. The first is precision. Notice how the headline numbers are quoted in FP4 — four-bit floating point — rather than the FP32 or FP16 of a few years ago. AI models have turned out to work well at very low precision: a model does not need thirty-two bits of arithmetic to decide the next word. Lower precision means more operations per second, less memory per weight, and faster answers, so each generation squeezes more work out of fewer bits. The B200’s second-generation Transformer Engine and Rubin’s NVFP4 format are the hardware expressions of that insight. The catch is that a peak FP4 number tells you almost nothing about how fast your particular model will answer; it is a ceiling, not a promise.
The second is the rack. The unit that matters is no longer the GPU. NVIDIA now sells the NVL72 rack as the product: 72 GPUs, 36 CPUs, and the NVLink fabric that joins them, acting as one machine with one enormous pool of memory. The next step, Rubin Ultra NVL576, scales that to 576 GPUs. When NVIDIA talks about an “AI factory”, this is what it means — a data centre designed as a factory floor, where the product being manufactured is intelligence, measured in tokens per second. Our shorter explainer on Blackwell and what comes next covers the Blackwell generation on its own, in less depth; this guide is the full map, and Blackwell is one stop on it.
Inside the GPU: Tensor Cores, HBM and NVLink

Open up an AI GPU and you will find three things that make it different from the processor in your laptop, and each one solves a different bottleneck in AI work. Together they explain why a gaming graphics card and a data-centre AI chip can share a brand name and still be different machines.
The first is the Tensor Core. A normal processor core is a generalist: it can do any calculation, one at a time, very fast. A Tensor Core is a specialist: it does one kind of calculation — matrix multiplication, the core operation of neural networks — on whole blocks of numbers at once. NVIDIA’s Blackwell chips carry fifth-generation Tensor Cores, and they are the reason the headline performance figures in the previous section are measured in petaflops rather than teraflops. The trade-off is that a Tensor Core is useless for anything that is not matrix maths, which is why your laptop does not need one and why a data-centre GPU is a poor general-purpose computer. The lesson is that AI hardware is not “faster computing” in general; it is computing specialised for one workload, and the specialisation keeps getting narrower as the precision drops from FP16 to FP8 to FP4.
The second is the memory. AI models are enormous — the largest have hundreds of billions of parameters — and the processor can only work as fast as it can fetch those parameters from memory. The memory on an AI GPU is not ordinary RAM. It is HBM, high-bandwidth memory: stacks of memory chips sitting directly beside the processor, connected by thousands of tiny wires, delivering terabytes per second. A Rubin GPU carries 288 GB of HBM4 at 22 terabytes per second of bandwidth. A Vera Rubin NVL72 rack carries 20.7 terabytes of HBM4 across its 72 GPUs, plus up to 54 terabytes of memory beside its CPUs. The practical rule is simple: in AI, memory capacity decides which models you can run, and memory bandwidth decides how fast they answer. That is why every generation’s headline is a memory figure as much as a compute figure, and why the memory shortage of 2026 pushed hardware prices up across the industry.
The third is the interconnect — the plumbing between GPUs. Training a large model is not one processor’s job; it is thousands of processors sharing the work, and they must constantly exchange partial results. If the connection between them is slow, the processors spend their time waiting instead of computing. NVLink is NVIDIA’s answer: a direct, high-speed link between GPUs, now in its fifth generation on Blackwell and its sixth on Rubin, that moves data at 130 terabytes per second inside a GB200 rack. NVSwitch extends that fabric across the whole rack, so all 72 GPUs can talk to each other as if they were one. Between racks, the job falls to networking: InfiniBand, in the Quantum-X800 platform, and Spectrum-X Ethernet, both running at 800 gigabits per second per GPU, carried by ConnectX network cards and managed by BlueField data-processing units that offload the networking work from the main processors.
Put the three together and the picture is clear. The Tensor Cores do the maths, the HBM feeds them, and NVLink and the network keep thousands of them working in step. Remove any one of the three and the system stalls. That is also why the chip alone is not the product: the value is in the balance between compute, memory and interconnect, and in the fact that NVIDIA designs all three together. A rival with faster Tensor Cores but slower memory would lose; a rival with faster memory but no NVLink equivalent would lose at rack scale. The whole stack has to move together, and it is the moving together that is hard to copy.
CUDA: The Software Nobody Sees — and the Real Moat

If you ask an AI engineer what makes NVIDIA hard to beat, they will rarely start with the chips. They will start with four letters: CUDA. CUDA is NVIDIA’s parallel-computing platform — a programming model, a set of tools, and a body of libraries that together let developers write programs that run on NVIDIA’s GPUs. It launched in 2007, when the idea of using a graphics card for general computing was still strange, and it has been extended, documented and taught ever since. The current version is CUDA Toolkit 13.5, released in November 2026, with production-ready support for the Vera Rubin architecture.
To understand what CUDA does, imagine the GPU as a factory with thousands of workers, and the CPU as a single skilled craftsperson. The craftsperson is brilliant at complex, sequential work; the factory is brilliant at doing the same simple job thousands of times at once. The problem is that programming a factory is hard: you have to decide which worker does what, in what order, and how they share results. CUDA is the management system for the factory. It gives developers a way to write code that says “run this calculation on a thousand workers at once”, and it handles the scheduling, the memory movement and the synchronisation. Without it, every developer would have to manage the factory by hand. With it, a graduate student can write a program that uses thousands of GPU cores without thinking about the hardware at all.
The deeper point is what sits on top of CUDA. The popular AI frameworks — PyTorch, TensorFlow, JAX — are all built on CUDA libraries. When a researcher writes a PyTorch model, the heavy lifting underneath is done by NVIDIA’s code. When a company fine-tunes a model, the optimiser runs on CUDA. When a cloud provider serves a model, the inference engine is compiled for CUDA. Eighteen years of this has produced a flywheel: developers learn CUDA because the jobs and the frameworks use it; frameworks target CUDA because the developers are there; hardware buyers choose NVIDIA because the software runs best on it; and NVIDIA invests the profits in more CUDA. Breaking that flywheel is the central strategic problem for every NVIDIA competitor, and nobody has broken it yet.
CUDA 13, the current major version, also shows how the platform keeps moving. The toolkit no longer bundles the GPU driver on Windows (since 13.1) or Linux (since 13.4) — the driver is installed separately. The Python package names dropped their CUDA-version suffixes, so the same package name now installs the latest CUDA 13 build. Applications built with CUDA 13 keep running on drivers from the R580 branch onward, while the newest features need the R620 branch. None of this matters to a non-developer, and all of it matters to the people who run AI infrastructure, because it is the difference between a platform that stays still and one that is still being actively built. The idea file for this guide lists “CUDA explained” as one of its three target keywords, and this section is the answer: CUDA is not a feature of NVIDIA’s GPUs. It is the reason the GPUs are worth buying.
The CUDA-X Libraries: cuDNN, TensorRT, NCCL and CUTLASS

CUDA is the platform. The CUDA-X libraries are the specialist tools that live on it — pre-written, heavily optimised routines for the operations AI actually performs, so that nobody has to write them from scratch. If CUDA is the factory’s management system, the libraries are the machines on the factory floor. Most AI developers never call them directly; the frameworks do. But the libraries are where much of NVIDIA’s performance advantage actually lives, and they are worth knowing by name.
cuDNN is the oldest and most important. It is NVIDIA’s library of deep-learning primitives: the optimised routines for convolutions, attention, normalisation and the other building blocks of neural networks. When PyTorch runs a model on an NVIDIA GPU, cuDNN is very often doing the actual arithmetic underneath. It is tuned, generation by generation, for each new GPU architecture, which is one reason a new NVIDIA chip is fast on day one: the library already knows how to use it.
TensorRT is the inference specialist. Training a model and running it are different jobs, and TensorRT is built for the second one. It takes a trained model and optimises it for serving: fusing operations together, choosing the fastest kernels, and quantising the weights to lower precision — FP8, FP4 — where the accuracy cost is negligible. TensorRT-LLM is the large-language-model version of the same idea, with specialised attention kernels and in-flight batching, which groups incoming requests together so the GPU is never idle. If you have ever marvelled at how quickly a chatbot answers, TensorRT or something like it is part of the reason.
NCCL is the teamwork library. When a model is too big for one GPU, or a training run spans hundreds of them, the GPUs must constantly exchange gradients and activations. NCCL — the NVIDIA Collective Communications Library — is the optimised code that does that exchanging, over NVLink inside a server and over InfiniBand between servers. It is invisible to the end user and decisive for training speed: a training run is only as fast as its slowest communication step, and NCCL is what keeps that step fast.
CUTLASS is the odd one out, and the most revealing. It is a collection of CUDA templates for writing your own high-performance matrix operations — essentially, the toolkit NVIDIA’s own engineers use to build the other libraries, published openly for anyone to use. Its existence tells you something about NVIDIA’s strategy: the company is confident enough to publish the tools it uses to stay ahead, because the advantage is not in any single library but in the whole stack working together, and in the pace at which new versions arrive. The libraries are updated with every GPU generation and every CUDA release; a competitor that matches today’s cuDNN still has to match tomorrow’s.
From Training to Inference: Dynamo, Triton and NIMs

Training a model and serving it are two different businesses, and the shift from one to the other is the biggest change in NVIDIA’s software story since CUDA itself. Training is a one-off, enormously expensive job: weeks or months of compute on thousands of GPUs, producing a single set of weights. Inference is the everyday job: answering millions of user prompts, one at a time, as cheaply and as quickly as possible. As AI moved from research labs into products, inference became the larger market — and NVIDIA rebuilt its serving software around it.
The centrepiece is Dynamo. Announced at GTC 2025, Dynamo is NVIDIA’s open-source inference-serving platform, and the company calls it the “operating system of the AI factory”. Its job is to manage the whole serving problem at data-centre scale: splitting a model’s work across GPUs, balancing requests, keeping the expensive hardware busy, and making sure a reasoning model that “thinks” for a long time before answering does not block everyone else. Dynamo is the successor to the Triton Inference Server, NVIDIA’s earlier serving software — and the rename matters, because it is easy to get confused by the old name still appearing in tutorials. As of March 2025, Triton Inference Server is now Dynamo-Triton, the serving component inside the Dynamo platform. The current Dynamo release is version 1.6.0, its nineteenth feature release, with engines including TensorRT-LLM 1.3.0, vLLM 0.31.0 and SGLang 0.5.21.
Below Dynamo, in the stack, sit the NIMs. A NIM — an NVIDIA Inference Microservice — is a pre-built container that packages a single AI model with everything needed to serve it: the optimised inference engine (TensorRT-LLM or vLLM), the Triton serving layer, and an OpenAI-compatible API on the front, so any application written against the OpenAI API can talk to it. The point of a NIM is that a developer does not have to build the serving stack at all. Download the container, point it at a GPU, and the model is live behind a standard API. NVIDIA curates NIMs for the popular open models — Llama, Mistral, DeepSeek, Qwen, Gemma and others — and enterprises use them as the fastest route from “we have a model” to “we have a service”. The reasoning-model boom is what made this layer so important: when models spend far more compute thinking before they answer, as DeepSeek’s reasoning models showed the world in early 2025, the serving layer becomes the bottleneck, and Dynamo and the Blackwell Ultra generation were both designed with that bottleneck in mind.
For a reader who will never serve a model, the practical takeaway is simpler. The reason AI products feel fast and cheap is not only that the models got better; it is that the serving layer got enormously more efficient. Quantisation, batching, kernel fusion, disaggregated serving — the techniques in TensorRT, Dynamo and the NIMs — are what turn a model that costs a fortune to run into a product that costs fractions of a cent per query. That efficiency is a large part of what NVIDIA sells, and it is a large part of why the company’s data-centre revenue keeps growing even as the price of a single answer keeps falling.
The Enterprise Layer: NVIDIA AI Enterprise, NeMo and DGX Cloud

So far this guide has described tools that are mostly free: CUDA, the libraries, Dynamo, the NIM containers. The enterprise layer is where NVIDIA charges for the packaging, the support and the management — and it is where a large business actually buys. Understanding it matters because it explains how NVIDIA turns open software into a recurring, enterprise-grade business, and because it is the layer a business buyer will meet first.
NVIDIA AI Enterprise is the paid platform: a supported, security-patched, enterprise-licensed bundle of the whole software stack — NIMs, NeMo, RAPIDS for data science, Triton/Dynamo for serving, and the CUDA libraries underneath — with NVIDIA support behind it. The pitch is not new technology but risk reduction: a bank or a hospital cannot run its AI on unsupported open-source software and hope for the best. It needs a vendor who will answer the phone, patch a vulnerability, and certify the stack for its compliance regime. AI Enterprise is that vendor relationship, and NVIDIA recommends it on every DGX system it sells.
NeMo is the framework inside that bundle for building and customising models. Where a NIM is a pre-built model ready to serve, NeMo is the workshop for making a model your own: data curation, fine-tuning, alignment, guardrails and evaluation, aimed at teams that need a model trained on their own data rather than a generic one off the shelf. It is the part of the stack that speaks most directly to the current moment in enterprise AI, in which the frontier models are general-purpose and the value is in adapting them — safely, measurably — to a specific business.
DGX Cloud is the managed end of the same idea: NVIDIA’s AI platform as a service, running on the major clouds. A customer rents access to NVIDIA’s systems and software, fully managed, with the company’s AI expertise included, rather than buying hardware and hiring the team to run it. It is the answer for a company that wants NVIDIA’s stack without NVIDIA’s data centre, and it is also the channel through which NVIDIA sells to customers who will never own a GPU. The model makers themselves are the biggest users of this kind of capacity — OpenAI’s models and Meta’s Llama family both run, in large part, on NVIDIA hardware, whether owned or rented — which is why the enterprise layer and the model layer are really one market: everyone needs somewhere to run the model, and NVIDIA sells the somewhere.
One more piece belongs here: NGC, NVIDIA’s catalogue of containers, models and software. It is the app store of the stack — the place where the NIMs, the framework containers, the Helm charts and the pre-trained models live, ready to pull and run. For a developer, NGC is the practical starting point: it is where “download the container and run it” actually happens. For NVIDIA, it is the distribution channel that keeps the whole stack one download away from any developer with a GPU.
The Systems Ladder: DGX, HGX and the AI Factory

NVIDIA sells chips, but it delivers systems. The ladder runs from a single server to a whole data centre, and each rung is a product with a name. Knowing the names is the fastest way to read any NVIDIA announcement, because the announcement is almost always about a rung on this ladder.
The DGX line is the server rung. A DGX is a complete, ready-to-run AI server: eight GPUs, the CPUs, the memory, the networking and the software, tested as one machine. The DGX B200 is the current Blackwell version — eight Blackwell GPUs, 1,440 GB of memory, and, by NVIDIA’s figures, three times the training performance and fifteen times the inference performance of the DGX H100 it replaces. The DGX H200 is the Hopper equivalent, still sold. DGX systems are the workhorses of enterprise AI: a company buys one or a few, puts them in its own data centre, and has a serious AI capability without building a supercomputer.
The HGX line is the baseboard rung: the eight-GPU board that server makers build into their own machines. If DGX is NVIDIA’s own server, HGX is the board inside everyone else’s. The HGX Rubin NVL8 is the current flagship here — eight Rubin GPUs with a single Vera CPU, 400 petaflops of NVFP4 inference compute, 2 terabytes of HBM4, and, by NVIDIA’s figures, ten times the token throughput of the HGX B200 it replaces. The distinction matters because most of the world’s AI capacity is not bought as DGX servers; it is bought as HGX boards inside servers from Dell, HPE, Supermicro, Lenovo and the other system makers.
Above the server is the rack. The NVL72 systems — GB200, GB300, and now Vera Rubin NVL72 — put 72 GPUs and 36 CPUs in a single liquid-cooled rack, joined by NVLink into what NVIDIA calls a single massive GPU. The rack is the unit that hyperscalers buy by the hundred, and it is the unit that defines the economics: a rack is a fixed quantity of compute, memory and power, and a data centre is so many racks. The Vera Rubin NVL72 is the rack that is shipping now, to the clouds and the neoclouds — CoreWeave, Google Cloud, Microsoft Azure, Oracle Cloud and Nebius were all named on NVIDIA’s earnings call as already running Rubin systems.
Above the rack is the AI factory. This is NVIDIA’s term for a data centre designed from the ground up to manufacture intelligence: racks of NVL72 systems, connected by InfiniBand or Spectrum-X Ethernet, managed by software — Dynamo for serving, Mission Control for operations — that treats the whole building as one computer. The factory is also where the networking matters most: at this scale, the network is the computer’s nervous system, and NVIDIA’s ownership of InfiniBand, Spectrum-X, ConnectX and BlueField means it sells the nervous system as well as the brain. Amazon’s Bedrock and the other cloud AI platforms are, from the hardware’s point of view, tenants in these factories — they rent the capacity and sell it on as an API. The factory is the product; the cloud is the shopfront.
AI on Your Desk: DGX Spark, DGX Station and GeForce RTX

Not every AI workload needs a data centre. NVIDIA’s desk-scale line, announced at GTC 2025, is the company’s answer to the developer who wants to prototype on a desk and deploy in a rack — using the same software stack at both ends.
The DGX Spark is the small one: a palm-sized box, about 150 millimetres square and 1.2 kilograms, built around the GB10 Grace Blackwell Superchip. Its CPU and GPU share a single pool of 128 GB of memory at 273 gigabytes per second, and NVIDIA quotes up to one petaflop of FP4 compute. The practical claim is that one Spark can run models of up to 200 billion parameters — which covers most open models a developer would actually use. A 64 GB configuration, sold by NVIDIA’s partners from $4,999, was added in October 2026; the 128 GB version is $6,950. Prices move with the memory market, so treat those as October 2026 figures and check a live quote before buying. The Spark runs the full NVIDIA AI software stack, including the NIMs, so a prototype built on a Spark moves to a DGX server without being rewritten.
The DGX Station is the step up: a deskside tower built on the GB300 Grace Blackwell Ultra chip, pairing a Blackwell Ultra GPU with a 72-core Grace CPU. The headline figure is the memory — 748 GB of coherent memory in total, made of 252 GB of HBM3e next to the GPU (at 7.1 terabytes per second) and 496 GB of LPDDR5X next to the CPU. NVIDIA quotes up to 20 petaflops of AI compute, and says a single Station can run models up to a trillion parameters. It is a serious machine with serious requirements: 1,600 watts of power and a 20-amp circuit, which is to say, a machine for an office that has been wired for it. NVIDIA does not publish a list price and does not sell its own version; partners — ASUS, Dell, HP, MSI, Supermicro and others — sell it, with public listings from around $92,550 in October 2026.
And then there is the product most readers will actually own: GeForce. The RTX 50 series, launched in January 2025, brings the Blackwell architecture to consumer graphics cards, and the same CUDA stack runs on them. A gaming PC with an RTX card is, with the right software, a small AI workstation — good for running local models, for image generation, for learning CUDA. The consumer business is small beside the data centre now — the Edge Computing segment, which includes GeForce, did $8.2 billion in the last quarter against the data centre’s $107.6 billion — but it matters strategically. Every student who learns CUDA on a GeForce card is a future data-centre customer, and every developer who prototypes on a Spark is a future DGX buyer. The desk is the top of the funnel, and NVIDIA has always understood that.
Physical AI: Isaac, GR00T, Omniverse and Cosmos

NVIDIA’s next bet is that the largest AI market after chatbots is not another chatbot. It is robots — what the company calls physical AI: machines that perceive the world, reason about it, and act in it. The robotics stack is the newest part of the NVIDIA platform, and the least mature, but it is where the company is placing its longest bet, and it follows the same pattern as everything else in this guide: hardware, software, simulation and systems, sold as a stack.
Isaac is the robotics platform: the libraries, simulation tools and runtime that let developers build robots on NVIDIA hardware, from the Jetson computers that go inside the robot to the data-centre systems that train it. The headline product is Isaac GR00T N1, announced at GTC in March 2025 and described by NVIDIA as the world’s first open, fully customisable foundation model for humanoid reasoning and skills. Its architecture copies a well-known idea from human cognition: two systems, one slow and one fast. The slow system perceives, reasons and plans; the fast system executes precise, continuous actions in real time. The model is open — the weights and a subset of the training data are available on Hugging Face and GitHub — and it is designed to be customised for different robots and tasks, which is why robotics companies from Unitree to Figure AI to Boston Dynamics are building on the platform.
Training a robot is a data problem, and that is where Omniverse and Cosmos come in. Real-world robot data is scarce and expensive: you cannot learn to pick up a box by watching a million videos of boxes being picked up, because the videos do not exist. NVIDIA’s answer is to generate the data. Omniverse is the simulation platform — a physically accurate digital twin of the world, where a robot can practise a task millions of times in simulation before it ever touches a real object. Cosmos is the world foundation model platform: generative models that create realistic, controllable synthetic video and environments, turning a small amount of real data into a large amount of training data. The two work together — Omniverse for physical accuracy, Cosmos for photorealism and variety — and the GR00T Blueprint is the recipe that turns the combination into training data for manipulation tasks. Newton, a physics engine developed with Google DeepMind and Disney Research, supplies the underlying physics.
It is worth being honest about where this stands. Humanoid robots are not yet a mass market, and GR00T N1 is a first-generation foundation model, not a finished product. The robots that use it are still, for the most part, in laboratories and pilot deployments in warehouses and factories. But the strategic logic is clear: if robots become the next great AI workload, the company that sells the training stack, the simulation stack and the computer inside the robot wins the same way it won the data centre. The Jetson Thor computer, the robot’s onboard brain, is the desk-scale end of that bet — the Spark and Station of the physical world. Watch this space: it is the part of NVIDIA’s story most likely to look different in two years’ time.
The Business of NVIDIA: $115.8 Billion, a $5 Trillion Peak — and a New CEO

The technology story and the business story are the same story, told in dollars. On 17 November 2026, NVIDIA reported its third-quarter results for fiscal 2027, covering the three months to 25 October. Revenue was $115.8 billion — more than double the $57 billion of the same quarter a year earlier, and up 20% on the previous quarter. The data-centre business brought in $107.6 billion, up 110% year on year, and accounted for about 93% of all revenue. Inside that figure, hyperscale customers — the big cloud providers — contributed $60.1 billion, while AI clouds, industrial and enterprise customers contributed $47.5 billion, nearly triple the year before. Net income was $65.3 billion. For the current quarter, the company guided revenue of $130 billion, plus or minus 3%.
One line in that guidance deserves attention, because it is the clearest statement of how geopolitics touches this company. The forecast assumes no data-centre revenue from China. Export controls bar NVIDIA from selling its most advanced AI chips to China, and the company is planning as if that remains the case. In the quarter just reported, it still sold $2.4 billion of older-generation Hopper products to China. The China question — what NVIDIA can sell, to whom, and under what licence — is now a permanent feature of the company’s outlook, and the board underlined the point on the earnings call by creating a new national-security and export-control oversight committee.
The market value tells the other half of the story, and it has been a volatile month. In early October 2026, NVIDIA became the first company in history to pass a $5 trillion market capitalisation. By 16–17 November, after the leadership change described below, it had fallen to roughly $3.75 trillion — third in the world, behind Apple at about $5.1 trillion and Microsoft at about $3.85 trillion. The figures move daily and will be stale by the time you read this; what matters is the shape. A company whose revenue doubles in a year can still lose a quarter of its market value in a month, because the market is pricing not just the present but the pace of AI spending, the durability of the moat, and the risk that the build-out slows.
The leadership change is the reason for the sharpest part of that fall, and it belongs in this guide because any reader meeting NVIDIA for the first time needs to know who runs it. Jensen Huang, who co-founded the company in 1993 and led it for 33 years, resigned as president, chief executive and director on 12 November 2026, effective immediately. The board named Colette Kress — chief financial officer since September 2013 — as interim chief executive, keeping her CFO role, and appointed director Harvey C. Jones as the company’s first-ever chair of the board. A permanent-CEO search is under way, considering internal and external candidates. On the earnings call of 17 November, Huang said he would be honoured to return as chief executive if the board allowed it; the board unanimously declined, with Jones expressing full confidence in Kress, who told investors the roadmap is on track and that “we’ve got this”. No successor has been named. For the purposes of this guide, the fact that matters is the one the board itself stated: operations, customer commitments and the product roadmap continue without interruption, and the Rubin ramp — the generation that defines 2026 — is under way on schedule.
The Competition: AMD, Google TPUs, Trainium — and Why Switching Is Hard

NVIDIA is dominant, not unchallenged. The competition falls into three groups, and each one attacks a different part of the moat. Understanding them is the honest test of everything this guide has claimed: if the moat is real, the challengers should be struggling in specific, explicable ways. They are — but they are also real, funded, and improving.
AMD is the only other company selling top-end AI GPUs to anyone who wants to buy them. Its Instinct MI400 series, built on the CDNA 5 architecture, is its first serious rack-scale answer to NVIDIA: the MI455X carries 432 GB of HBM4 per GPU — more memory per chip than Rubin’s 288 GB — and the Helios rack packs 72 of them with 18 EPYC CPUs. AMD announced further MI400 details, including the MI450X, on 12 November 2026, and it has real customers: OpenAI and Meta have each committed gigawatts of AMD capacity, and Anthropic has agreed up to 2 gigawatts of MI450-class capacity. AMD’s software, ROCm, is the open platform that challenges CUDA, and it has improved steadily. The honest assessment is that AMD now competes on hardware at the top end; what it has not yet matched is the depth of the software ecosystem, the breadth of the library and serving stack, and the eighteen years of developer habit. That is a gap that closes slowly, but it is closing.
Google and Amazon attack from inside their own clouds. Google’s Ironwood, its seventh-generation TPU, became generally available on 31 March 2026: a chip designed for inference, with 192 GB of HBM per chip and pods of 9,216 chips, available only on Google Cloud. Google’s own Gemini models run on it, and so do a growing number of Google Cloud customers. Amazon’s Trainium3, generally available since December 2025, is the equivalent inside AWS: a chip you cannot buy, only rent, on EC2. Microsoft’s Maia 200 and Meta’s MTIA play the same role inside their owners’ fleets. The pattern is the same in each case: the hyperscaler designs a chip for its own workloads, runs it at enormous scale, and sells the capacity. The chips are good — Ironwood and Trainium3 are serious hardware — but they are not for sale, and they do not come with CUDA. They are a bet that the cloud, not the chip, is the product.
And then there is the most interesting challenger of all: NVIDIA’s own customers. The largest buyers of NVIDIA hardware — the cloud providers and the model makers — are the companies with the strongest incentive to build alternatives, because dependence on a single supplier is a strategic risk and a negotiating weakness. Anthropic runs Claude on more than a million AWS Trainium chips and has committed to AMD and Google capacity alongside NVIDIA. OpenAI has signed gigawatt-scale AMD agreements. This diversification is the single biggest structural threat to NVIDIA’s dominance, and it is already under way. The counter-argument — the one NVIDIA would make — is that diversification is additive, not substitutive: customers buy AMD and TPUs for some workloads and keep buying NVIDIA for the workloads where the software stack matters most, and the total market keeps growing fast enough for everyone. The evidence so far supports the counter-argument: NVIDIA’s data-centre revenue doubled in the year its customers diversified most aggressively. But it is a race, and the honest verdict is that the moat is wide, not unbreachable.
One final caution about comparing the rivals. Vendor peak-performance figures are not comparable across companies: they use different precisions, different assumptions about sparsity, and different definitions of the system being measured. A Rubin GPU quoted at 50 petaflops of NVFP4 inference and an MI455X quoted at roughly 40 petaflops of MXFP4 are not two numbers on the same scale. Memory capacity and bandwidth are the more honest first comparison, and even those do not tell you how fast your model will answer. The right question is never “which chip is fastest on paper”. It is “which stack runs my workload, at my scale, with my team, at a price I can afford” — and that question has a different answer for every buyer.
Verdict: What NVIDIA Means for You — and What to Watch Next

So what does all of this mean, for you? It depends on who you are — and the stack looks different from each vantage point.
If you are a developer, NVIDIA is the platform you will meet first, and CUDA is the skill that pays. Learn it, and learn the layer above it: the frameworks, the NIMs, Dynamo. The hardware generations will keep turning over — Rubin now, Rubin Ultra in 2027, Feynman in 2028 — but the software is the durable part of the skill. If you want to run models locally, start with a GeForce card or a DGX Spark and the NIM containers; the stack is the same from the desk to the data centre, and that is the point.
If you are a business buyer, the decision is not “NVIDIA or AMD” but “which stack, at which scale, for which workload”. For most teams, the answer will involve NVIDIA somewhere — if only because the software your engineers already know runs best on it. The questions to ask are about the serving layer and the enterprise layer: what does inference cost per query, what does NVIDIA AI Enterprise buy you over the open stack, and how much of your spend is locked into a single vendor’s roadmap. Rent before you buy; DGX Cloud and the clouds exist precisely so you do not have to build a data centre to find out.
If you are a student of AI, NVIDIA is the case study in how a platform wins: not with one product but with a stack, not with one generation but with a cadence, not with hardware alone but with software, systems and a developer community built over eighteen years. The company’s history — a gaming chipmaker that bet on parallel computing in 2007, years before AI needed it — is the best illustration in technology of what it means to be early and then to be ready.
If you are watching the market, five things decide what happens next. First, the Rubin ramp: NVIDIA says Rubin systems are already operating at the major clouds; the speed and smoothness of that ramp is the single most important operational fact of 2027. Second, the permanent chief executive: who the board chooses, and whether the choice is an insider who knows the stack or an outsider who brings something new. Third, the challengers: whether AMD’s MI400 series and the hyperscalers’ custom silicon move from diversification to substitution. Fourth, export rules: the China line in the guidance is now a standing assumption, and any change in it moves the numbers materially. And fifth, the shift from training to inference: as reasoning models spend more compute thinking, the serving layer — Dynamo, the NIMs, the Blackwell Ultra and Rubin generations — becomes the market, and the company that owns serving owns the next decade. xAI’s Grok and the other frontier models are already being served at a scale that would have been unimaginable three years ago; the infrastructure that serves them is the story.
The final verdict is simple. NVIDIA is not a chip company that got lucky with AI. It is a full-stack AI company that spent eighteen years building the stack before the market arrived, and then spent the last four years scaling it faster than anyone else. The chips are excellent, the software is the moat, and the systems are the product. The leadership has changed, the market value has swung by trillions, and the competition is better funded than ever — and none of that has changed the stack. That is the point of this guide, and the reason it is worth 7,000 words: to understand modern AI, you have to understand what it runs on. Now you do. Checked 20 November 2026.
Sources
- NVIDIA Investor Relations — “NVIDIA Blackwell Ultra AI Factory Platform Paves Way for Age of AI Reasoning” (18 March 2025): GB300 NVL72, HGX B300, Spectrum-X and Quantum-X800 at 800 Gb/s per GPU, Dynamo announcement.
- NVIDIA — DGX B200 product page (nvidia.com): 8x Blackwell GPUs, 1,440 GB total HBM3e, FP4 144 PFLOPS sparse / 72 dense, 3x training / 15x inference vs DGX H100.
- NVIDIA — HGX platform page (nvidia.com): HGX Rubin NVL8 (400 PFLOPS NVFP4, 2 TB HBM4, 130 TB/s, 24 TB/s NVLink Switch), Rubin GPU (288 GB HBM4, 22 TB/s, 50 PFLOPS NVFP4 inference), HGX B300/B200 figures.
- NVIDIA — GB200 NVL72 product page (nvidia.com): 36 Grace CPUs + 72 Blackwell GPUs, 13.4 TB HBM3e, NVLink 130 TB/s, NVFP4 1,440 PFLOPS sparse.
- NVIDIA CUDA documentation, Release 13.5 (docs.nvidia.com, November 2026): current toolkit, Vera Rubin support, driver unbundling (Windows from 13.1, Linux from 13.4), PyPI suffix removal, R620 driver branch.
- NVIDIA Dynamo documentation, v1.6.0 release notes (docs.nvidia.com): 19th feature release; engines TensorRT-LLM 1.3.0, vLLM 0.31.0, SGLang 0.5.21.
- NVIDIA Developer — Dynamo-Triton pages (developer.nvidia.com): Triton Inference Server renamed Dynamo-Triton, March 2025; Dynamo as successor to Triton for LLM inference.
- NVIDIA Newsroom — Dynamo launch announcement (nvidianews.nvidia.com): Dynamo as “the operating system of the AI factory”.
- NVIDIA Investor Relations — “NVIDIA Announces CEO Transition” (12 November 2026): Huang resignation; Colette Kress interim CEO; Harvey C. Jones first board chair; permanent-CEO search.
- CNN — “Nvidia rules out Jensen Huang’s return as CEO, tells investors ‘we’ve got this'” (17 November 2026): board unanimously declined Huang’s return offer; national-security/export-control oversight committee created.
- NVIDIA Q3 fiscal 2027 results (17 November 2026), as reported by TechRepublic, Investor’s Business Daily and TechMonitor from NVIDIA’s own release and CFO commentary: revenue $115.8B (+103% YoY), Data Center $107.6B (+110%), Edge Computing $8.2B, GAAP net income $65.3B, Q4 guidance $130B ±3% with no China data-centre revenue, Vera Rubin systems operating at CoreWeave, Google Cloud, Microsoft Azure, Oracle Cloud Infrastructure and Nebius, $2.4B Hopper sold to China in Q3.
- GTC 2025 coverage — CRN, eWeek, GIGAZINE (18–19 March 2025): Vera Rubin NVL144 (H2 2026, 3.3x GB300), Rubin Ultra NVL576 (H2 2027, 365 TB, 14x GB300), Feynman (2028), DGX Spark and DGX Station announcements, Isaac GR00T N1.
- Vera Rubin NVL72 rack specifications and the NVL144 → NVL72 rename — CUDO Compute, Inside Deep Tech, Spheron, QuantaCloud (November 2026 guides, cross-checked against NVIDIA’s product pages): 72 Rubin GPUs × 288 GB HBM4 = 20.7 TB HBM4 per rack, up to 54 TB LPDDR5X CPU memory, NVLink 6 at 260 TB/s per rack; the NVL144 name counted the two compute dies per Rubin package and was dropped in early 2026.
- DGX Spark and DGX Station specifications — NVIDIA product pages via ReviewRamble (1 November 2026), pi3g, vuncloud, Docker blog: Spark GB10 128 GB LPDDR5x at 273 GB/s, 1 PFLOP FP4 sparse, up to 200B-parameter models, $4,999 (64 GB) / $6,950 (128 GB) as of October 2026; Station GB300 748 GB coherent memory, 20 PFLOPS FP4 sparse, 1,600 W on a 20 A circuit, partner-sold from ~$92,550.
- Isaac GR00T N1 and Cosmos — Hackster, Geeky Gadgets, Interesting Engineering (March 2025): GR00T N1 as the first open foundation model for humanoid reasoning and skills, dual-system architecture, Omniverse digital twins, Cosmos world foundation models, Newton physics engine with Google DeepMind and Disney Research.
- Competitor landscape — gedolab (14 November 2026), techietory (18 October 2026), digitalmatters, blog.prompt20.com: AMD Instinct MI400 series and MI450X (12 November 2026), Helios rack, ROCm; Google Ironwood TPU v7 generally available 31 March 2026; AWS Trainium3 generally available December 2025; Microsoft Maia 200; Meta MTIA; AI chip market-share estimate of ~70% (Tom’s Hardware, May 2026).
- Market capitalisation — financecharts (16 November 2026 close: $3.748T), Cointelegraph / KuCoin (17 November 2026: Apple ~$5.13T, Microsoft ~$3.85T, NVIDIA ~$3.76T), techfastforward (30 October 2026: Apple $5.03T passed NVIDIA $4.93T; NVIDIA first past $5T earlier in October 2026).


0 Comments