AI dominates the headlines daily: new models, unprecedented partnerships, innovative products. As ChatGPT celebrates its third anniversary, the artificial intelligence landscape has become a maze where the same players compete on multiple fronts simultaneously. Alphabet faces off against Nvidia in the chip market, while Gemini rivals ChatGPT as a language model. But how do these different battles fit together and what is the common thread in this technological war?

CPUs have been the core of computers since the early days of modern computing. Still today, every computer runs on a CPU. It is the main processor and works as the “general-purpose brain” of the computer. Without one, a computer can’t even start. However, these chips were made for flexibility and not for repetitive calculations. Thus, in the 1990s, when video games were getting increasingly realistic, computers needed chips that could process thousands of pixels simultaneously instead of one at a time. That is what led to Graphics Processing Units (GPUs). A GPU is basically thousands of smaller, simpler cores working in parallel. While a CPU colours pixels one by one (serially), a GPU colours all of them at once (parallelly). Most computers today have both a CPU and a GPU. And the biggest producer of GPUs was Nvidia.

Then came modern AI. Training large neural networks requires mind-numbing amounts of matrix math, and that math looks almost identical to what GPUs were already designed to do for 3D graphics. Nvidia capitalised on this by adapting GPUs for datacentres, far bigger and more powerful than the ones in laptops, and gained a huge market share very quickly. GPUs, however, were still general-purpose graphics chips, not purpose-built AI chips. In 2015, Google introduced the Tensor Processing Unit (TPU), designed specifically for the operations used in neural networks. In other words, the TPU is an ASIC (application-specific integrated circuit), a chip customised for a particular use, rather than intended for general-purpose use. This specialisation enabled TPUs to offer superior cost-efficiency, energy efficiency, and high throughput at scale.

The third type of chip is the Neural Processing Unit (NPU). They work on-device—like with a smartphone—focusing on energy efficiency and real-time AI processing. It does not compete with GPUs and TPUs, as it is far from ideal when it comes to training large-scale AI models and have lower computational power. However, since they are optimised for low-power AI applications, they are primarily deployed in applications such as real-time image recognition and voice recognition.

Source: Leon Zhu & SemiVision on X

Because history locked the AI world into GPUs long before TPUs arrived.

Think of it like this: by the time Google invented the TPU in 2015, the machine-learning world had already spent a decade building everything, such as research code, early neural-network libraries, university courses, or company infrastructure on top of Nvidia GPUs. And the reason wasn’t just “GPUs are good.” It was CUDA, Nvidia’s proprietary programming platform that launched in 2007.

CUDA let developers write code that directly exploited the parallel power of GPUs. It was stable, fast, well-documented, and crucially, it worked everywhere from gaming PCs to giant datacentres. That created a flywheel, where more developers used CUDA, meaning more tools and libraries were built for CUDA, leading more companies to buy Nvidia GPUs, finally pushing even more developers toward CUDA.

By the early 2010s, the entire ecosystem was built assuming people were using Nvidia GPUs and CUDA. Universities taught it. Startups deployed on it. Cloud providers stocked warehouses full of it.

When Google dropped TPUs, they were powerful, especially for large-scale training and inference workloads. But, they were also far less flexible than GPUs in terms of the types of models and operations they could handle. Most importantly, they were late to a party Nvidia had been hosting for years, and the whole community had already built their house on CUDA. In addition, Google didn’t sell TPUs. If you wanted to use them, you had to train your model on Google Cloud. Meanwhile, Nvidia GPUs were everywhere: Amazon, Microsoft, Oracle, private datacentres, university clusters. Broad availability made GPUs the default for anyone building an LLM.

Even still, TPUs’ better performance and the wish for less reliance to a single company—in this case, Nvidia—led players to start integrating Google’s chips in their value chain. Anthropic recently signed a multi-billion-dollar deal with Google, gaining access to up to one million TPUs. Their models will now run across three different platforms: Nvidia’s GPUs, Amazon’s custom Trainium chips (also ASICs), and Google’s TPUs. Meta is reportedly also in talks with Alphabet to invest billions of dollars in Google TPUs for its data centres by 2027. This would help the company diversify their chip supply beyond Nvidia’s and AMD’s GPUs. More generally, more companies are looking to build their own ASICs. For example, OpenAI has a new deal with Broadcom, who helped build Google’s TPUs, to build its own ASICs starting in 2026.