The standard for measuring large compute deals has shifted from number of GPUs to gigawatts of power. This provides a normalized, apples-to-apples comparison across different chip generations and manufacturers, acknowledging that energy is the primary bottleneck for building AI data centers.

The performance gains from Nvidia's Hopper to Blackwell GPUs come from increased size and power, not efficiency. This signals a potential scaling limit, creating an opportunity for radically new hardware primitives and neural network architectures beyond today's matrix-multiplication-centric models.

Digital computing, the standard for 80 years, is too power-hungry for scalable AI. Unconventional AI's Naveen Rao is betting on analog computing, which uses physics to perform calculations, as a more energy-efficient substrate for the unique demands of intelligent, stochastic workloads.

When power (watts) is the primary constraint for data centers, the total cost of compute becomes secondary. The crucial metric is performance-per-watt. This gives a massive pricing advantage to the most efficient chipmakers, as customers will pay anything for hardware that maximizes output from their limited power budget.

To contextualize the energy cost of AI inference, a single query to a large language model uses roughly the same amount of electricity as running a standard microwave for just one second.

Pat Gelsinger contends that the true constraint on AI's expansion is energy availability. He frames the issue starkly: every gigawatt of power required by a new data center is equivalent to building a new nuclear reactor, a massive physical infrastructure challenge that will limit growth more than chips or capital.

The infrastructure demands of AI have caused an exponential increase in data center scale. Two years ago, a 1-megawatt facility was considered a good size. Today, a large AI data center is a 1-gigawatt facility—a 1000-fold increase. This rapid escalation underscores the immense and expensive capital investment required to power AI.

Efficiency gains in new chips like NVIDIA's H200 don't lower overall energy use. Instead, developers leverage the added performance to build larger, more complex models. This "ambition creep" negates chip-level savings by increasing training times and data movement, ultimately driving total system power consumption higher.

The Fetus GPT experiment reveals that while its model struggles with just 15MB of text, a human child learns language and complex concepts from a similarly small dataset. This highlights the incredible data and energy efficiency of the human brain compared to large language models.