Why every tech company is racing to build its own AI chip

The race to power the next generation of AI has triggered a surge in new hardware. AMD recently debuted its Helios system, a giant, rack-sized unit full of specialized processors designed to train and run the most advanced AI models. Meanwhile, a startup called Etched has raised massive amounts of funding to produce its own custom-built chips. Unlike generic chips that try to do everything, Etched is designing hardware intended solely for the specific, grueling math required to run AI models.

Think of a standard computer processor as a chef who can cook a meal, organize a bookshelf, or balance a checkbook. It is versatile, but not the fastest at any one task. AI chips are like specialized assembly lines. To understand why they are special, you have to look at how an AI builds an answer. It does this in segments called tokens, which are roughly equivalent to a word or part of a word. When you ask an AI a question, it goes through two phases. First is the prefill phase, where the model reads and processes your entire prompt to understand what you want. This is memory-intensive and requires immense, immediate speed. Next is the decode phase, where the AI actually types out its answer one token at a time. This phase needs a vast, fast pool of memory so the AI can remember everything it has already written while predicting the next token in the sequence.

The demand for these systems is growing because modern AI is moving toward agentic tasks—instances where an AI doesn't just give you a single answer, but performs a long chain of actions like accessing tools, checking data, and reasoning through problems. Each step of that chain requires heavy computation, and the industry expects this demand to explode over the next few years. As this hardware race heats up, the bottleneck for AI progress might shift from the software itself to the physical limits of the factories and materials used to build these powerful, specialized engines.