Summary
Highlights
In 2025, the world's largest companies spent approximately $400 billion on capital expenditures for AI development, a sum comparable to nine Manhattan Projects or two Apollo programs. This spending is primarily on data center infrastructure, excluding operational costs and non-public company investments. Despite record-breaking spending, no AI company (except hardware suppliers like Nvidia) has turned a profit. Interestingly, while companies announce record new spending, over half of the predicted data center sites for this year have been delayed or cancelled, creating a logical paradox within the industry.
Nvidia holds a critical position in the AI boom due to its market cap, profit, and reinvestment. Three major questions surround Nvidia's role: where are all the chips going, power limitations, and chip longevity. Jensen Huang claimed Nvidia shipped around 10 GW of GPUs in 2025. However, estimates from Goldman Sachs suggest only 7.7 GW of AI data centers are currently operational globally. Investigations reveal that much of the announced data center capacity is not under active construction, leading to questions about the actual demand and where Nvidia's production is being utilized.
A significant bottleneck for new data centers is not the advanced computer chips themselves, but the electrical infrastructure required to support them. Power has become the real-world constraint, with essential components like transformers more than doubling in price and facing supply struggles. This has led companies to buy components as soon as they can, even if they don't have immediate use for them, contributing to a 'bullwhip effect' in industrial planning. This situation benefits Nvidia, as they can readily sell their chips, but it's a risky game that could lead to massive oversupply with a small demand correction. Nvidia's inventory has more than doubled, suggesting challenges in moving chips or supply chain issues.
Beyond power, energy costs are a major concern. Fully operational data centers consume vast amounts of power, and rising energy prices (exacerbated by global events) significantly impact their viability and cash burn rates. Many data centers rely on local grids, facing higher costs and longer waits for capacity, while others using natural gas generators contend with doubled gas prices. Furthermore, the industry standard for depreciating GPUs over six years is unrealistic, as newer models render existing hardware obsolete much faster (within three years). This accounting practice makes profits appear better than they are and could reduce investor hype and demand for future GPUs, directly affecting Nvidia. The combination of supply bottlenecks and high energy prices makes this depreciation problem even worse, potentially turning expensive hardware into e-waste if operating costs outweigh rental value.