Green AI: the new chips that promise more intelligence with far less energy
The new AI processors aim to maximize performance per watt: specialized architectures, liquid cooling and renewable energy, keys to more sustainable computing.
The other cost of AI: energy
The explosion of artificial intelligence in recent years has had a little-noticed but very real side effect: electricity consumption at data centers has skyrocketed. Training and running large language models requires thousands of processors working in parallel for weeks or months, and each of those chips consumes huge amounts of energy, in addition to requiring powerful cooling systems that add even more cost.
Faced with this scenario, the semiconductor industry has focused on a goal that is as technical as it is urgent: getting the next generation of AI processors to deliver more computing power while consuming less energy. The approach is known as sustainable computing, and combines improvements in chip architecture, new materials and designs created specifically for AI workloads.
New generation of processors: performance per watt
The industry's headline metric is no longer just raw power in operations per second, but energy efficiency: how many operations a chip can perform for every watt consumed. Manufacturers such as NVIDIA, AMD, Intel and Asian giants are competing to improve this indicator with each generation, incorporating faster and more efficient memory, units specialized in matrices (the mathematical heart of AI) and increasingly miniaturized manufacturing processes.
In parallel, big technology companies are working on custom hardware for their own models. Apple, Google and Amazon have designed their own silicon optimized for specific workloads, an approach that drastically reduces wasted energy compared with using generic chips for every kind of task.
Strategies to reduce the energy footprint
Beyond the design of the processors themselves, the ecosystem is adopting complementary measures:
- Energy-smart data centers: direct liquid cooling on the chip dissipates heat far more efficiently than air, cutting the energy associated with cooling.
- Reduced-precision computing: using more compact numerical formats (FP8, INT8) reduces computing and memory effort with no practical loss of quality in many models.
- Renewable energy: hyperscalers sign renewable supply contracts and locate their new computing farms close to clean energy sources.
- Neuromorphic and photonic chips: in research, architectures inspired by the brain or based on light promise orders-of-magnitude gains in efficiency.
A market in full expansion
Economic interest pushes in the same direction. AI data centers already account for a growing share of global electricity consumption, with forecasts pointing to sustained growth through 2030. This makes efficiency not only an environmental issue but also a competitive advantage: whoever manages to train and serve models while spending less energy will be able to offer lower prices and wider margins. Sector analysts agree that the combination of efficient hardware and software optimization will be one of the decisive factors of technological leadership in artificial intelligence in the coming years.
The question is no longer just how much computing power a chip can deliver, but how much intelligence each watt consumed can generate.
Sources
El Tiempo: Artificial intelligence and sustainable technology