I watched Jensen Huang sit down with Joe Rogan, and the easy version of the story is the one everybody already knows: NVIDIA made graphics cards, AI got huge, and suddenly the graphics card company became the obvious hardware name in the room.
That version is true enough to be familiar and too thin to be useful.
The more interesting story is not founder mythology. It is infrastructure. NVIDIA looked like a gaming company because gaming was the first mass market that needed what NVIDIA was really building: cheap parallel computation, real-time image generation, developer tooling, and a platform that made specialized hardware useful to normal people.
The supporting technology often attracts little attention outside its users. It looks like drivers. APIs. Compilers. Developer relations. Memory bandwidth. Thermal envelopes. Chip packaging. Render pipelines. Software libraries. Technical decisions that later applications may depend on.
AI did not make NVIDIA important out of nowhere. AI benefited from NVIDIA’s existing hardware and software work.
It was never just games
For a long time, “graphics card” sounded like a gamer purchase. It was the thing you bought so Quake, Doom, Unreal, flight simulators, and later every other visually ambitious thing could run better.
That made it easy to underestimate.
But graphics has never really been trivial. Real-time graphics is a demanding computing problem. You are trying to generate convincing reality many times per second, under hard latency constraints, on a machine somebody can fit under a desk. The system must calculate each frame within a strict time limit. Lighting, geometry, textures, shadows, simulation, physics, animation, and camera movement all have to resolve fast enough that the illusion does not fall apart.
Games created a consumer market for that computing capability.
In the Rogan conversation, Huang keeps returning to this idea: computer graphics was an early application for a different style of computation. NVIDIA’s CUDA guide describes the GPU as starting out as hardware for parallel operations in real-time 3D rendering, then becoming more programmable over time. CPUs are built around fast serial work. GPUs are built around doing huge numbers of related operations in parallel.
That distinction sounds technical because it is. It is also important to which markets the hardware can serve.
The first killer app was visual. Make the screen better. Make the game faster. Make the simulation feel real. Make the world render.
The long bet was that this kind of computation would matter far beyond the screen.
Hardware needed software
Supporting systems become more valuable as more people depend on them.
NVIDIA built a business around both hardware and the software needed to use it. The company was founded in 1993 around 3D graphics for gaming and multimedia. The visible product was the card. The business also depended on hardware, software, drivers, developer tools, game relationships, and eventually a programming model that let people treat the GPU as something more general than a graphics device.
The hardware matters, but the hardware is only part of its competitive advantage. A chip without a software stack is a component. A chip with libraries, tooling, developer habits, training material, examples, installed base, and years of weird edge cases already solved becomes infrastructure.
This is why CUDA matters. NVIDIA introduced CUDA in the mid-2000s, giving developers a way to use GPU throughput outside graphics APIs. The same basic hardware originally justified by games and graphics could now be used for scientific simulations, image classification, databases, analytics, machine learning, and eventually generative AI.
That did not look obvious at first. Near the end of the Rogan conversation, Huang says NVIDIA added CUDA before customers were asking for it, increasing cost before the market knew how to value it. That is the uncomfortable part of a long bet: the costs arrive before the additional revenue.
This is where I think the NVIDIA story is more useful when we focus on the investment rather than the founder’s personality. The point is not “one genius saw the future.” The point is that a company spent years building a technical layer before the mainstream use case arrived.
That is rarer, and more instructive.
Timing, not destiny
The clean narrative says NVIDIA built the GPU, deep learning arrived, and the plan worked.
The real narrative is messier.
Neural networks were not new. AI had already spent decades cycling between promise and disappointment. Then researchers made significant progress in computer vision. In the 2012 AlexNet paper, Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton describe training a deep convolutional network on two NVIDIA GTX 580 GPUs, a result that became one of the canonical moments in modern deep learning.
That detail matters because it is almost too perfect: the AI breakthrough that helped set off the current era ran on graphics cards.
The researchers used commercially available graphics hardware.
The tools built to make games faster turned out to be useful for teaching machines to see. The work done for rendering, pixels, shaders, and parallel math became relevant to neural networks because neural networks also thrive on large amounts of parallelizable computation.
That is timing. But it is not pure luck.
Luck is AlexNet showing up when it did. Luck is researchers finding the hardware. Luck is the world finally having enough data, compute, and technique in the same place.
The non-luck part is that NVIDIA had already spent years making the hardware programmable, available, and supported. Researchers could use that existing support when they found a new application for the hardware.
Why production people should care
From my perspective, if you come from VFX, production, post, photography, games, virtual production, or creative technology, this story should feel familiar.
You already know that supporting systems often determine what a team can deliver.
The audience sees the shot. The producer sees the schedule. The supervisor sees the pipeline. The artist sees the comp. The render wrangler sees the farm. The editor sees the storage problem. The colorist sees the monitor chain. The person responsible for delivery sees the codec, the file naming, the conforms, the missing plates, the bad metadata, the overnight render that failed at 4:12 a.m.
Problems with storage, rendering, or delivery can prevent a creative plan from being completed.
That is why NVIDIA’s story is not just a Wall Street story or an AI story. It is a production story.
Modern production is already a stack of specialized compute. VFX depends on rendering, simulation, compositing, color pipelines, asset management, and increasingly real-time engines. Virtual production workflows moved game-engine technology into previs, pitchvis, techvis, postvis, live compositing, in-camera VFX, virtual scouting, and earlier design review. AI is now sliding into the same stack: search the archive, generate references, build temp boards, rough out shots, denoise plates, assist roto, synthesize voices, summarize footage, organize notes, test looks, and speed up routine coordination and file handling.
Across these uses, more production tasks can be represented and manipulated in software.
First the image became computable. Then the scene became computable. Then the camera, lighting, physics, materials, motion, and edit decisions became increasingly computable. Now language, reference, planning, coding, search, and visual ideation are becoming part of the same computational workspace.
The team still needs to judge which results are useful.
When the tool makes rough output cheap, the valuable person is not the one who can generate the most options. The valuable person is the one who understands which options can be made and will serve the finished production.
VFX people should understand NVIDIA because they have been living inside this transition longer than most people. They know that a faster render does not automatically make a better shot. A real-time preview does not eliminate judgment. A beautiful synthetic frame can still be useless if it cannot be approved, tracked, licensed, matched, rendered, versioned, or delivered.
The tool is never the whole system.
The platform was the product
One of the least interesting ways to talk about NVIDIA is to obsess over any single chip as if the story is just a comparison of benchmarks.
The better question is what becomes possible once enough people build around the same toolchain. That is what happened with games. Developers optimized for the hardware. Engines absorbed the assumptions. Players bought the cards. The cards justified better games. Better games justified better cards. A market became a loop. Huang describes this as turning a technology business into a platform business. That is the important sentence.
The same pattern shows up again in AI. Researchers write against the libraries. Frameworks support the hardware. Cloud providers buy the systems. Startups assume the stack. Engineers learn the tools. Models are designed around the available compute.
At that point, infrastructure is not passive anymore. It starts shaping the field.
This is why shared technical dependencies influence production and business decisions. The API influences what gets built. The engine influences how a shot can be planned. The render budget limits its complexity. The available compute influences the model. The energy constraint influences the geography.
People like to imagine technology arriving as an independent invention. More often, it arrives as a dependency.
Physical requirements
The deeper NVIDIA goes into AI, the less the story is about a chip in isolation.
It becomes about fabs, packaging, memory, interconnects, data centers, cooling, power, supply chains, and export controls. It also becomes about customers trying to turn investment into working systems before model requirements change.
That is a very different story from “graphics card company gets hot.”
AI at scale is industrial infrastructure. It needs land, power, water, network capacity, specialized labor, long-range planning, and giant purchasing commitments. The International Energy Agency estimates data centers used about 415 TWh of electricity in 2024 and projects that demand could roughly double by 2030, with AI accelerating deployment of high-performance servers. The broader numbers back that up. The limiting factor is not just whether somebody can imagine a bigger model. It is whether somebody can power, cool, connect, finance, and operate the machine that trains or serves it.
Again, production people should recognize this.
The creative question is always attached to the physical one. Can we make the shot? Can we render it by Friday? Can we store the plates? Can we move the data? Can the stage handle it? Can the crew turn it around? Can the budget absorb the change?
AI did not escape those questions. It made them larger.
Less myth, more pattern
Huang is an effective advocate for the company, but his personality explains only part of its development.
The founder version of the NVIDIA story is tempting because it gives the narrative a face. A founder’s personality is easier to describe than a software ecosystem or a hardware constraint.
But the more useful pattern is bigger than one person.
Some technologies look narrow because the first market is narrow. Some tools look like toys because the first users are playing. Not every niche tool becomes world historical. Most stay niche. Some disappear. Occasionally, a tool developed for a narrow market proves useful much more broadly.
Graphics looked like games. Games required real-time worlds. Real-time worlds required parallel computation. Those habits turned out to be useful for simulation, rendering, scientific computing, machine learning, and AI.
The line from Quake to ChatGPT is not straight. It is not destiny. But it is real enough to make you reconsider what “just a tool” means.
What to watch next
I am watching the tools that look too specialized for mainstream people to care about: real-time engines used earlier in design, local AI tools for private footage search, simulation systems for synthetic training data, render pipelines that combine previs and review, and agents that handle routine coordination among files, calendars, notes, transcripts, shot lists, asset databases, and delivery specs.
Not all of these tools will prove useful. I’m interested in those that solve a specific problem reliably enough for teams to depend on them.
That is what makes the graphics-to-AI history relevant to production. A tool can become useful beyond its original market when other people can program it, support it, and incorporate it into their work.