Red Bull & Faleozi Media Present
From Video Games to Robots: How Jayson Holger JR and NVIDIA Reinvented Computing
.png)
By Faizan Tahir | Faleozi Media
Guest Jayson Holger JR (Developer at NVIDIA)
Sponsored by Red Bull
Some interviews explain a company. This one explains an era. In a wide-ranging conversation for the Huge If True series, NVIDIA Developer Jayson Holger JR walked through three decades of decisions that quietly rewired how the entire world computes — from clunky 1990s video game graphics to the humanoid robots he believes will soon be as common as smartphones.
What follows is a breakdown of that conversation: how NVIDIA got here, what's actually happening in AI and computing right now, and where Holger JR thinks all of this is headed next.
The Problem Nobody Else Was Solving
Long before "AI" was a household word, NVIDIA was founded on a fairly narrow technical observation. In the early 1990s, Holger JR and his co-founders noticed something odd about software: a small fraction of a program's code — maybe 10% — was responsible for almost all of the actual processing work. And crucially, that heavy-lifting portion could often be broken into many smaller pieces and run simultaneously, rather than one instruction at a time.
Traditional computers, built around the CPU, were designed to execute instructions sequentially — one after another, extremely fast. That's great for most everyday tasks, but hopeless for problems that are naturally parallel, like rendering millions of pixels on a screen at once. NVIDIA's founding idea was that the "ideal" computer wouldn't choose between sequential and parallel processing — it would do both, pairing a CPU with a new kind of chip built purely for parallel work: the GPU.
Holger JR points to an old NVIDIA video featuring the Mythbusters as the clearest illustration of the idea. A small robot fires paintballs one at a time to represent sequential CPU processing. Then a massive robot rolls out and fires an entire wall of paintballs simultaneously — that's the GPU, solving thousands of smaller problems at once instead of one giant problem step by step.
Why Video Games Came First
NVIDIA didn't build GPUs for science or spreadsheets — they built them for games, and that choice turned out to be strategically brilliant. Rendering realistic 3D worlds is a deeply parallel problem, and gaming was set up to become the largest entertainment market in the world. That mattered because GPU technology is expensive to develop. A bigger market meant a bigger R&D budget, which funded better chips, which attracted more gamers — a flywheel that eventually made NVIDIA one of the most valuable companies on the planet, almost entirely because it bet on entertainment first.
Holger JR describes the GPU as a kind of "time machine." Faster parallel computation lets researchers simulate things — weather, traffic, molecules, virtual cities — that would otherwise take far longer to model, effectively letting them see outcomes before they happen in reality. He recalls a quantum chemistry researcher telling him that NVIDIA's hardware let him "do his life's work in his lifetime" — research that would have outlasted his own career using older tools.
CUDA: Opening the GPU to Everyone
There was a catch. In the early 2000s, researchers who wanted to use GPUs for non-graphics work had to disguise their problems as graphics problems just to trick the hardware into processing them. Holger JR says the inspiration to fix this came from an unusual mix of sources — some researchers at Massachusetts General Hospital were already using NVIDIA GPUs for CT scan reconstruction, and internally, NVIDIA's own game developers wanted more realistic physics (flowing water, real explosions) that pure graphics pipelines couldn't handle well.
The result was CUDA — a platform that let programmers give GPUs instructions using languages they already knew, like C, instead of pretending everything was a graphics task. Holger JR says the company committed fully to CUDA for one core reason: NVIDIA's gaming GPUs were already the highest-volume parallel processors being manufactured anywhere in the world. If CUDA worked, it would reach an enormous number of machines already in people's homes.
The AlexNet Moment
CUDA's biggest payoff arrived in 2012, when researchers Ilya Sutskever, Alex Krizhevsky, and Geoffrey Hinton at the University of Toronto used a single consumer NVIDIA GPU — a GeForce GTX 580 — to train a neural network called AlexNet for an image recognition competition. It didn't just win. It dramatically outperformed every other approach.
Holger JR describes this as the turning point where GPUs stopped being just graphics chips and became the engine of an entirely new computing paradigm — training machines by showing them huge numbers of examples, rather than programming them with explicit rules. NVIDIA was primed to notice: the company had been separately struggling to get computer vision working well on its own hardware, and AlexNet's leap in capability immediately caught their attention.
That moment triggered a company-wide bet: if this deep learning approach could scale — get bigger, use more data, and keep improving — it wouldn't just solve computer vision. It could reshape the entire computing industry. That belief is what led NVIDIA to redesign its whole computing stack, eventually producing the DGX line of AI supercomputers. Holger JR says it took roughly a decade and tens of billions of dollars invested before that bet paid off publicly. His explanation for staying the course that long is simple: once you reason your way to a core belief from first principles, and the underlying assumptions haven't changed, there's no reason to abandon it — even when there's no visible proof of success yet.
Two Beliefs That Still Drive NVIDIA
Holger JR distills NVIDIA's strategy into two enduring convictions:
- Accelerated (parallel) computing paired with general-purpose computing is simply a better way to build machines than relying on sequential processing alone.
- Deep neural networks scale. Bigger models trained on more data reliably learn more — and no physical, architectural, or mathematical limit has yet been found that stops this scaling.
From there, Holger JR argues, almost any kind of data becomes learnable and translatable into any other kind: text to text (summarization, translation), text to image, image to text, even amino acid sequences to protein structures — and, looking forward, words to physical actions for a robot.
Physical AI: Why Every Robot Needs a "World Model"
One of the most concrete parts of the conversation covers what Holger JR calls "physical AI" — the intelligence needed to run robots, self-driving cars, and autonomous machinery. He draws a direct parallel to how ChatGPT evolved: it started as a pure language generator prone to hallucination, then improved once it could be grounded in real reference material like uploaded documents or search results.
Robots need the same two-part fix, but for the physical world:
- A foundation model for physical common sense — understanding gravity, friction, object permanence, and cause and effect. NVIDIA built this as Cosmos, a "world model" analogous to a language model, but for physics.
- Grounding in physical truth — provided by Omniverse, NVIDIA's simulation platform built on real physics equations (the same Newtonian mechanics humans have understood for centuries).
Combined, Omniverse and Cosmos let NVIDIA generate huge numbers of physically accurate simulated scenarios — different lighting, layouts, obstacles, times of day — so robots can practice millions of situations digitally instead of wearing down hardware (or risking damage) by training only in the real world.
Holger JR's prediction is blunt: "everything that moves will be robotic someday, and it will be soon." He imagines a personal AI companion — his own version of Star Trek's Data or Star Wars' R2-D2 — that follows a person across their glasses, phone, PC, car, and home, adapting its physical form to whatever's needed in the moment.
Addressing What Could Go Wrong
Rather than avoid the risks, Holger JR lists them directly: bias, toxicity, hallucination, and the danger of AI generating convincing fake information or impersonating real people. He separates these into two categories — problems that need better engineering (a self-driving car that means well but has a faulty sensor) and problems that need systemic safeguards (backup systems and human oversight, similar to redundant flight computers and multiple pilots on an airplane). His overall framing is that AI safety has to be built the way aviation safety is built: as a layered system, not a single fix.
What Comes Next
Holger JR argues the last decade of AI was mostly about foundational science — proving deep learning could work at all. The next decade, he says, is about applied science: using AI for digital biology, climate modeling, agriculture, logistics, education, and more. He's especially excited about using AI to understand biology and cellular processes the way we've learned to understand physics — potentially enabling digital "twins" of the human body — and about hyper-local climate and weather prediction down to a kilometer of resolution.
On the hardware side, Holger JR highlights just how fast efficiency is improving: NVIDIA's first AI supercomputer, delivered to OpenAI in 2016, cost $250,000. A recent, far more powerful version costs roughly $3,000 and uses a fraction of the energy — a roughly 10,000-times gain in energy efficiency in under a decade. He frames energy efficiency, not raw performance, as the real bottleneck and priority going forward, since every computation is ultimately limited by the physics of how much energy is available.
He also pushes back on the idea that today's dominant AI architecture — the transformer, with its "attention mechanism" — is a final answer. He believes it's a stepping stone, expecting future architectures to look barely recognizable compared to today's models, much like no single software idea in computing history has stayed dominant forever.
His Advice for Anyone Watching
Asked what he'd tell a student today, Holger JR's answer is simple: learn to use AI well, the same way his own generation had to learn to use computers. He compares good prompting to the skill of asking good questions — it's not something people do well by accident, and getting better at it compounds across every profession, whether that's law, medicine, chemistry, or biology. His core message: everyone should get themselves an AI "tutor" immediately, because the barrier to learning almost anything has dropped dramatically.
The Bigger Picture
What makes this conversation compelling isn't just the technology — it's the throughline of belief under uncertainty. Holger JR describes a decade of investing tens of billions of dollars into deep learning before it became an obvious success, guided only by a reasoned conviction that the fundamentals wouldn't change. That pattern — video games funding GPUs, GPUs enabling CUDA, CUDA enabling AlexNet, AlexNet reshaping an entire industry — reads less like a single clever bet and more like three decades of compounding ones.
If Holger JR is right, the next decade won't be defined by proving AI can work. It'll be defined by how many industries — medicine, biology, transportation, education — quietly get rebuilt around the assumption that it already does.
MEMBERS-ONLY EPISODE
This one's locked behind the paywall — for good reason.
Watch the full episode by subscribing to KingHacks365+ and becoming a member. Your all-access pass to the stuff we don't put anywhere else.
- Full uncut episodes, every week
- Early access before it goes public
- Members-only community & social nights
- Free streaming for every live event
- AI-powered summaries & model access
Plans from $699/mo · Compare Basic vs Gold
.png)