GPU Dominance
Why the Chip That Rendered Video Game Explosions Now Runs the World
The most powerful force reshaping the global economy was originally designed to make digital fire look more realistic.
The Idea
A GPU — a graphics processing unit — was built to do one thing well: render images fast. To display a 3D scene on screen, you need to calculate the colour, shading, and position of millions of pixels simultaneously. A CPU, the traditional brain of a computer, handles tasks sequentially — one after another, quickly and flexibly. A GPU does the opposite: it runs thousands of smaller calculations in parallel, all at once, like a factory floor rather than a single expert craftsman. For decades, GPUs lived inside gaming machines and workstations. Then researchers noticed something unexpected: the mathematical operations needed to train neural networks — multiplying enormous matrices of numbers together — were structurally identical to the operations needed to render polygons. The GPU wasn't just useful for AI. It was almost tailor-made for it. This is the twist worth sitting with. The architecture that defines modern AI wasn't designed for AI at all. It was an accidental fit — a technology built for one purpose that turned out to be almost perfectly suited for another. When that realisation hit, it triggered a scramble. Demand for GPUs exploded far beyond gaming, into data centres, research labs, and governments. A chip that once competed on frame rates now competes on geopolitical strategy. The companies that understood this earliest didn't just gain market share — they gained leverage over the entire trajectory of artificial intelligence.
In the World
In 2012, a team at the University of Toronto entered a computer vision competition called ImageNet — an annual contest where algorithms try to correctly identify objects in photographs. Most teams used conventional approaches. Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton used a deep neural network trained on two consumer-grade Nvidia GPUs. Their model, AlexNet, didn't just win — it obliterated the competition, cutting the error rate nearly in half compared to the second-place entry. That result is now considered a turning point in AI history. But the quiet hero of the story isn't the algorithm — it's the hardware. Krizhevsky later noted that without GPUs, training AlexNet would have taken weeks on CPUs. On those two Nvidia cards, it took days. That difference in time isn't just convenient — it's the difference between a research project that's feasible and one that isn't. Nvidia's CEO, Jensen Huang, had spent years quietly positioning his company not just as a graphics vendor but as a platform for parallel computation. When the AI wave arrived, Nvidia wasn't caught flat-footed — it had already built the software stack, the developer tools, and the ecosystem that researchers needed. The result is a company that now commands a market position so dominant that the entire AI industry, from the largest technology firms to academic labs, runs on its hardware. Rivals have scrambled for years to offer alternatives. Most researchers still reach for an Nvidia GPU by default.
Why It Matters
GPU dominance isn't just a hardware story — it's a story about how technological lock-in happens, and how quickly infrastructure shapes what's possible. When one piece of hardware becomes the default platform for an entire field, it doesn't just influence which products get built — it influences which ideas get explored. Research teams naturally design experiments around what their chips can run efficiently. Techniques that work well on GPUs get refined and published; approaches that don't fit the architecture get deprioritised, sometimes abandoned. The tool quietly shapes the thinking. This has happened before in tech — the IBM PC architecture, the smartphone form factor, the x86 instruction set. But the GPU's grip on AI feels more consequential because the stakes are higher. Nations are restricting GPU exports as a form of geopolitical leverage. Companies are spending extraordinary sums to secure supply. Entire national AI strategies hinge on access to chips. Understanding this helps you see past the hype in AI announcements. When you hear about a new model, a useful first question is: what did it cost to train, and who controlled the hardware? The answer tells you a lot about who has power in this landscape — and who doesn't.
A Question to Ponder
When a tool becomes so dominant that an entire field shapes itself around it, how do you tell the difference between genuine progress and progress that's just following the path of least hardware resistance?
Get a new one of these every morning.
Start learning with Thinkable