Behind the AI hype

A new analysis by Paligo reveals that the global artificial intelligence (AI) industry rests on a narrow and fragile infrastructure, with critical dependencies concentrated in just a few companies and regions.

At the center of it is ASML, the only firm producing extreme ultraviolet lithography machines essential for advanced chipmaking, with about 50 units shipped annually.

These machines are used by TSMC, which produces more than 90 percent of the world’s most advanced semiconductors. Paligo’s mapping shows that any disruption, from geopolitical conflict to export bans, could cascade across the entire AI ecosystem.

Beyond supply risks, the report highlights the growing environmental footprint of AI infrastructure. Data centers are projected to consume 945 terawatt-hours of electricity annually by 2030, roughly equivalent to the energy use of a mid-sized industrialized nation. With more than 11,800 facilities globally, nearly half located in the United States, the demand for power and cooling continues to rise. Maintaining large-scale models requires thousands of GPUs operating at stable temperatures, with cooling systems alone consuming and evaporating millions of liters of water daily, underscoring the hidden resource cost behind each AI query.

The economics of AI are also shifting, with the cost of operating models now surpassing the cost of building them. Training GPT-4 exceeded $100 million, while serving newer systems like GPT-5 requires more than 200,000 GPUs. Once trained, models rely on static knowledge and must retrieve or approximate new information, increasing the risk of inaccuracies. Paligo also pointed to “content debt” as a growing issue, where fragmented and outdated internal data increases hallucination risks, turning what was once a documentation problem into a core operational liability.