The AI in the US gets power cut and the GPUs in Silicon Valley are covered in dust for six years

Recently, there have been reports stating that two data centers in Santa Clara, California, have been vacant for over six years, sitting idle while waiting for power. Previously, Nadella of Microsoft also publicly admitted that due to power shortages, Microsoft's piles of GPUs were left to rust on the racks, and this phenomenon has spread throughout the technology industry. Valuable top-level NVIDIA GPUs worth billions of dollars are just piled up in Microsoft's data centers, with no one touching them, and the dust is almost covering the logos. The reason is not technological backwardness or a cooling market, but rather - insufficient power supply. From 2020 to the present, the global AI data centers' demand for power has increased by nearly 27% annually. However, the speed of new power plants being built in the US is only a mere 4.3%.

The dust-covered GPUs phenomenon in Silicon Valley has had far-reaching impacts on the technology sector. Firstly, it affects the computing resources. Microsoft and other enterprises' piles of GPUs, due to power shortages, cannot be put to use, resulting in the chips remaining idle in the warehouses. Calculating based on a six-year depreciation cycle for data center equipment, blindly hoarding not only occupies a large amount of cash but also incurs double losses of equipment value depreciation over time. The electricity price in the US has risen by 35% since 2022, and electricity has become the Achilles' heel of AI development. In data center operations, the electricity cost accounts for 40%-60% of the total cost. The power shortage forces enterprises to pay higher premiums to obtain electricity or be forced to adopt high-cost temporary solutions (such as diesel generators), further compressing profits. The rapid growth of AI and cloud computing-driven data center construction has far exceeded the planning of utility companies. Traditional power plants need several years from project initiation to grid connection, while the expansion of the AI industry is calculated by quarters, resulting in a severe mismatch between supply and demand. The idle GPUs force the industry to re-examine the development path of computing power. In the future, it may shift from simply pursuing the number of chips to optimizing energy efficiency ratios, such as through code optimization (such as Microsoft's Azure architecture integrating tens of thousands of GPUs into a single accelerator) or storage computing integration (PIM) technology to reduce power consumption. This transformation may trigger a revolution in computing philosophy and push AI to transition from "instrumental rationality" to "life rationality".

Secondly, it has an impact on the technology industry. Microsoft has been approved to transport NVIDIA chips to the United Arab Emirates, and it will invest 8 billion US dollars in building data centers in the Gulf countries over the next four years, taking advantage of the abundant energy resources there. This move marks that AI infrastructure is migrating from Silicon Valley to energy-rich emerging markets, and Silicon Valley may lose its core position as the global AI computing power center. The US government attempts to bind Middle Eastern resources through the "Computing Power Dollar" system, for example, requiring the UAE to build a data center of the same scale as a backup center in the US. Although this is intended to counter China's influence, in the long run, it may weaken Silicon Valley's independent innovation capabilities and form a vicious cycle of "technology dependence - resource bundling".

Thirdly, it has an impact on technology. GPUs are the core hardware for training large models, and their idleness directly leads to the stagnation of model parameter growth. From a long-term perspective, computing power may evolve towards an heterogeneous ecosystem of ASIC, FPGA, and GPU. However, the current GPU idleness causes enterprises to lack resources for investing in heterogeneous technology research and development, potentially missing the opportunity for the next generation of computing architectures. For example, neuromorphic chips have great potential in real-time perception and reasoning, but lack the computing power support of GPUs, and their commercialization will be delayed. The general architecture of GPUs has resource waste phenomena when processing specific AI tasks, resulting in suboptimal computing efficiency. In contrast, dedicated hardware designed for specific AI tasks (capable of achieving higher computing performance under the same power consumption and cost) can reduce power consumption.

In conclusion, the dust accumulation phenomenon on Silicon Valley's GPUs serves as a mirror, reflecting not only the deep-seated contradictions in current technological development regarding computing power, energy consumption, and industrial layout, but also the dynamic changes in the global technological competition landscape. In the future, the determining factor for technological development may no longer be limited to chip performance, but rather lie in energy autonomy, technological route selection, and the deep integration of global industrial collaboration.