A tweet arguing that today’s artificial intelligence boom owes its existence to 1990s first-person shooters has caught the attention of Elon Musk, who publicly agreed with the theory.
What Was Said
The post, from X user Dan (@KettlebellDan), argued that 3D games were effectively the starting point for modern AI. He wrote that it’s “crazy to think that without 3D games we wouldn’t have AI,” adding that if humanity ever extends consciousness to the stars, “DOOM will have been in a way the precursor technology.” The post was illustrated with a screenshot from id Software’s 1993 classic Doom.
Musk, who has founded or backed multiple AI ventures including xAI and OpenAI, replied with a simple, two-word confirmation: “True.”
It’s a short exchange, but it points to a real and often underappreciated chapter in computing history. The chip that renders a health bar and a shotgun in a first-person shooter turned out to be the same basic piece of hardware that trains today’s large language models. Here’s how that connection actually formed.
The Problem Doom Created
When Doom shipped in 1993, there was no such thing as a “graphics card” in the modern sense. Home PCs rendered 3D-ish environments using the CPU alone, through clever programming tricks rather than dedicated hardware. Doom, and the wave of 3D shooters that followed it, made the limits of that approach obvious. Gamers wanted faster frame rates, bigger levels, and more realistic lighting and texturing than a CPU could comfortably deliver.
That consumer demand created a market opportunity, and the industry responded quickly. Companies like 3dfx released the Voodoo graphics accelerator in 1996, built specifically to offload 3D rendering — texture mapping, polygon fills, shading — from the CPU. A new company founded in 1993, NVIDIA, entered the same race and would go on to define the category. In 1999, NVIDIA marketed its GeForce 256 as the world’s first “GPU,” or graphics processing unit, a chip purpose-built to do enormous numbers of simple math operations, like calculating pixel colors and vertex positions, at the same time.
That last detail — doing huge numbers of simple calculations in parallel — is the thread that connects Doom to modern AI.
Why Gaming Hardware Turned Out to Be AI Hardware
A CPU is designed to do a relatively small number of complex tasks quickly, one after another. A GPU is built the opposite way: it has thousands of simpler cores designed to do the same basic operation on massive amounts of data simultaneously. That’s exactly what rendering a 3D game requires, since every pixel and polygon on screen needs similar calculations performed at once.
It turns out that training a neural network requires almost the identical kind of math. Deep learning is built on matrix multiplication — repeatedly multiplying and adding huge grids of numbers together to adjust the weights of a neural network. Researchers in the mid-2000s began to notice that the parallel architecture built to push polygons in video games was, by coincidence, extremely well suited to this kind of computation too.
The turning point came in 2006, when NVIDIA released CUDA, a programming platform that let developers write general-purpose software for GPUs rather than only graphics code. For the first time, researchers outside the gaming and graphics world could tap into the parallel horsepower of a gaming card for their own computing problems, AI included.
The Moment It Clicked: AlexNet, 2012
The clearest proof of this connection arrived in 2012, at a competition called ImageNet, where researchers submitted algorithms to identify objects in photographs. A University of Toronto team led by Alex Krizhevsky, working with Ilya Sutskever and Geoffrey Hinton, entered a deep neural network called AlexNet, running it across two consumer NVIDIA GTX 580 gaming cards with the workload split between them.
The result wasn’t close. AlexNet’s error rate came in at roughly 15%, beating the runner-up by close to eleven percentage points, a gap that stunned a field used to incremental, hard-fought progress. Those gaming GPUs let the team train a neural network hundreds of times faster than a CPU-based approach would have allowed, which in turn meant they could train a much larger model on far more images than had previously been practical.
AlexNet is now widely regarded as the launch point of the modern deep learning era. It proved, concretely, that consumer gaming chips — refined over nearly two decades to render increasingly complex 3D worlds — were also the most efficient hardware on the planet for training neural networks.
From Gaming Chips to the AI Industry
That discovery reshaped NVIDIA’s business almost entirely. A company that built its fortune selling GPUs to gamers became the primary hardware supplier for the AI industry, with its chips now powering the training of nearly every major large language model, including the systems behind ChatGPT, Gemini, and Grok. The data-center GPUs used for that training today are direct architectural descendants of the same parallel design first mass-produced to render games like Doom, Quake, and their successors.
Seen through that lens, the joke in the original tweet has a real technical backbone. The gaming industry’s decades-long demand for faster, more realistic 3D graphics funded the development of massively parallel processors. Those processors, originally meant to draw imps and shotguns on a screen, turned out to be exactly the hardware neural networks needed to become powerful enough to drive the current AI boom.
Whether or not consciousness ever makes it to the stars, it’s fair to say that the graphics card born from Doom-era gaming demand is one of the more overlooked building blocks of the AI systems in use today.