The Man in the Leather Jacket
To understand the future of AI, you have to understand Jensen Huang. The founder and CEO of NVIDIA, famous for his signature black leather jacket, has spent three decades building a company that, by a stroke of strategic genius and luck, now sits at the center
of the technological universe. His company’s Graphics Processing Units (GPUs) were originally designed to render realistic video game graphics. But their ability to perform many simple calculations at once—a process called parallel computing—made them perfect for training the massive AI models that power services like ChatGPT and Google’s Gemini. Today, NVIDIA doesn’t just have a lead in the AI chip market; it has a virtual monopoly. The company’s data center division, which sells these powerful GPUs to companies like Google, Microsoft, and Amazon, generates tens of billions of dollars per quarter. This isn't just a successful business; it's a chokehold on the single most important resource for building cutting-edge AI. Huang’s decisions about production, pricing, and allocation directly dictate the pace of innovation for everyone else.
Why AI Is So Power-Hungry
Large Language Models (LLMs) like Gemini are not born smart; they are made smart through a process called “training.” Imagine forcing a student to read a library the size of the entire internet. This requires an astronomical amount of computational power. Training a model like GPT-4 reportedly cost over $100 million, largely on hardware. Running the model for users, a process called “inference,” is also incredibly resource-intensive.
This is where NVIDIA’s GPUs, specifically chips like the H100 and its successor, the B200, come in. They are the super-powered engines required for both training and inference. The demand is so high that these chips are often sold out for months, with Big Tech firms placing billion-dollar orders just to secure their spot in line. The success and capability of any new AI model is therefore directly tied to how many of these GPUs its creators can get their hands on. More GPUs mean a more powerful model, trained on more data, rolled out to more users.
Google’s Frenemy Problem
This creates a fascinating and tense dynamic for Google. On one hand, Google is one of NVIDIA’s biggest and most important customers, spending billions to power its cloud services and AI research. On the other hand, Google is desperately trying to build a moat around its own AI ecosystem to compete with Microsoft-backed OpenAI. The problem is that the moat is being built with bricks sold by a single, all-powerful supplier.
This reliance presents several strategic vulnerabilities. First, there's the cost. NVIDIA’s dominance allows it to command premium prices, making the development of next-generation AI an incredibly expensive arms race. Second, there's the issue of supply. If Google can't secure enough of the latest GPUs, its timeline for developing and launching a model like Gemini 3 could be delayed, giving rivals an opportunity to pull ahead. It’s a classic case of dependency: the kingmakers in the AI world are the ones building the models, but the king is the one selling the shovels.
The Homegrown Alternative Isn’t Enough
Google saw this coming. For years, the company has been developing its own custom-built AI chips called Tensor Processing Units (TPUs). In theory, these chips are designed specifically for Google’s software and should provide a competitive advantage and a path to independence from NVIDIA. And to some extent, they have. Google uses its TPUs extensively for its internal products, including Search and advertising.
However, the TPU strategy hasn't been a silver bullet. NVIDIA’s CUDA software platform, which is the programming language used to run its GPUs, has become the industry standard. This creates a powerful network effect; most AI researchers and developers are trained on CUDA, and a massive ecosystem of tools is built around it. Switching from NVIDIA to an internal solution at scale is difficult and risky. While Google’s TPUs are powerful, they haven't been able to displace NVIDIA’s GPUs as the go-to hardware for the most demanding, cutting-edge AI training, forcing Google to continue playing in Jensen Huang’s sandbox.
