It’s difficult to ignore how rapidly the discussion surrounding artificial intelligence has moved from the chatbots themselves to the odd, humming structures that support them. When discussing AI two years ago, the majority of people wanted to discuss ChatGPT’s magic.
These days, chips, cooling systems, and electricity bills are the topics of conversation in San Francisco coffee shops and on New York earnings calls. There’s a feeling that the software was never the true story. The hardware underneath was always the problem.
| Field | Details |
|---|---|
| Subject | The Hardware Backbone of Generative AI |
| Dominant Company | NVIDIA Corporation |
| Founded | 1993, Santa Clara, California |
| Key Innovation | Graphics Processing Unit (GPU), launched 1999 |
| Parallel Architecture | CUDA, introduced 2006 |
| Market Cap (Sept 2025) | $4.3 trillion — most valuable company globally |
| Revenue Growth | $158M (1999) → $60B+ (2024) → projected to double in 2025 |
| Gaming Revenue Share | 50%+ in 2020 → roughly 8% in 2025 |
| Main Rivals | Google TPU, AMD, Intel |
| Cost to Train GPT-4 | Estimated up to $100 million |
| Global Data Centre Market | Around $243 billion, projected to double by 2032 |
| Electricity Footprint | 1.5% of global power; close to 9% in the US |
| Hotspot Region | “Data centre alley,” Northern Virginia |
The clear face of this change is NVIDIA, but at this point, referring to it as merely a chipmaker seems almost contemptuous. The company was founded in 1993 with the goal of bringing 3D graphics to the gaming industry. In its early years, it sold to teenagers who wanted shooter games with smoother frame rates. Then came the GPU in 1999, followed by CUDA in 2006. Somewhere along the line, the parallel architecture designed to render exploding helicopters proved to be just what big language models required. When training an LLM, a GPU operates at least ten times faster than a standard CPU. That is a significant benefit. That encompasses the whole industry.
When you sit with the numbers, they still seem a little surreal. $158 million in revenue in 1999. By 2020, about $11 billion. Then $60 billion in 2024, which is predicted to more than double in 2025. NVIDIA surpassed Microsoft, Apple, Amazon, and Alphabet to reach a capitalization of $4.3 trillion by September 2025. Previously accounting for over half of its revenue, gaming now makes up about 8%. To put it politely, the remaining portion comes from the AI boom itself and comes from data centers.
The surge is not solely caused by processors. When models are taking in vast amounts of data, specialized memory for AI provides more than four times the bandwidth of standard memory. Information is retained in non-volatile memory even in the event of a power outage. Once widely used in robotics and digital signal work, field programmable gate arrays are power-efficient but require specialized hardware knowledge; for LLM workloads, GPUs have largely replaced them. On some tasks, Google’s Tensor Processing Unit—an ASIC designed especially for matrix operations—can outperform GPUs. Perhaps 3% of the market is currently held by it. Intel and AMD are circling. For the time being, NVIDIA’s software ecosystem, which is larger and more developed than anything the competitors can match, is its moat.
The data centers themselves are the less glamorous part of all of this. By 2032, the $243 billion global industry may have doubled in value. Approximately 9% of US electricity is already consumed by data centers, which account for 1.5% of global electricity consumption. Of that, 25% is concentrated in Virginia, in what the locals now refer to as “data centre alley.” There are shipping container-sized chillers, miles of fencing, and signs alerting people to water restrictions outside these facilities. Equipment is corroded by salt water. In the arid, low-humidity areas where these centers are typically located, drinkable water is in short supply. Although the capital investment is severe, liquid cooling is beneficial.
It appears that investors are confident in the potential returns. Whether they will is still up in the air. A portion of this resembles the early internet, with significant infrastructure investment coming before the companies that support it. A portion of it resembles a bubble that is just waiting for a pin. In any case, the dull part is no longer the hardware story. Perhaps it’s the only thing that matters.
