The Acceleration of AI and Its Contradictions

Those following artificial intelligence news often encounter conflicting narratives: a gold rush, a speculative bubble, a threat to jobs, or an immature technology. The annual AI Index 2026 report, published by Stanford University's Institute for Human-Centered Artificial Intelligence, aims to bring clarity to this complex landscape. Despite predictions of a potential slowdown, the document highlights how leading Large Language Models (LLMs) continue to improve at a rapid pace, exceeding expectations.

AI adoption is happening faster than that of personal computers or the internet, with companies generating revenue at an unprecedented speed compared to previous technology booms. However, this exponential growth is accompanied by massive investments, in the order of hundreds of billions of dollars, allocated to data centers and chips. The benchmarks designed to measure AI progress, the policies intended to govern it, and the job market are struggling to keep up. AI is sprinting, while the rest of the world struggles to find its footing.

Infrastructural Costs and Supply Chain Fragility

This acceleration comes with significant costs, especially in terms of resources. Global AI data centers currently draw 29.6 gigawatts of power, an amount sufficient to meet the peak demand of the entire state of New York. The annual water consumption for OpenAI's GPT-4o alone may exceed the drinking water needs of 12 million people. For organizations evaluating on-premise deployment, these figures underscore the importance of a careful TCO analysis, which must include not only hardware acquisition costs but also operational expenses related to energy and cooling.

In parallel, the chip supply chain proves to be extremely fragile. The United States hosts most of the world's AI data centers, but a single Taiwanese company, TSMC, fabricates almost every leading AI chip. This critical dependence highlights a strategic vulnerability for the global AI infrastructure. Geopolitical competition between the US and China is nearly tied in AI model performance, with the US boasting more powerful models, greater capital, and ten times more data centers than any other country, while China excels in AI research publications, patents, and robotics. This dynamic makes data sovereignty and control over infrastructure even more crucial for companies and governments.

Model Evolution and Benchmark Challenges

Despite predictions of a plateau, AI models continue to improve steadily. By some measures, they now meet or exceed the performance of human experts on tests that aim to measure PhD-level science, math, and language understanding. For example, the SWE-bench Verified software engineering benchmark saw top scores jump from around 60% in 2024 to almost 100% in 2025. However, AI still exhibits what is called "jagged intelligence," struggling in areas requiring physical world experience; robots, for instance, succeed in only 12% of household tasks.

A critical aspect highlighted by the report is the inadequacy of current benchmarks. Many are outdated, poorly constructed (a popular math benchmark has a 42% error rate), or can be gamed, with models learning to score well without necessarily becoming smarter. Furthermore, AI companies are sharing less information about their training methods, parameter counts, or dataset sizes, making it difficult for independent researchers to study the safety and reliability of the models. This lack of transparency is particularly problematic for those who need to evaluate an LLM's suitability for sensitive workloads, especially in air-gapped contexts or with stringent compliance requirements.

Impact on Jobs and Regulatory Complexity

AI has been adopted by over half the world's population and 88% of organizations, with four out of five university students using it. While it is still early to measure the overall impact on the job market, some studies suggest that AI is beginning to affect younger workers in certain professions. For example, employment for software developers aged 22 to 25 has fallen by nearly 20% since 2022. Companies expect AI to shrink the workforce in sectors such as services, supply chain, and software engineering, even while boosting productivity in areas like customer service (+14%) and software development (+26%).

Public perception of AI is complex and ambivalent: 59% of people believe it will bring more benefits than drawbacks, but 52% express nervousness. There is a significant gap between experts and the public, particularly regarding AI's impact on jobs (73% of experts optimistic versus 23% of the public). Governments worldwide are struggling to regulate AI, although some progress is being made, such as the first prohibitions of the EU AI Act and new national laws in Japan, South Korea, and Italy. In the United States, while the federal government moves towards deregulation, state legislatures passed a record 150 AI-related bills in 2025. However, regulation lags behind technology, primarily because the functioning of these complex systems is not yet fully understood.