Stanford AI Index 2026: What Developers Need to Know — 10 Key Charts

What is the Stanford AI Index and why is it trusted

Every year, dozens of organizations publish reports on the state of AI. Most of them are released by companies that themselves sell models—and this inevitably influences their conclusions.

The Stanford AI Index is structured differently. It is published by Stanford University, without commercial interest in the model race. The report has been released since 2017, relying on data from government databases, scientific journals, and independent criteria.

Chart 1. Energy footprint of models

As the capabilities of models grow, so does their impact on the environment. Training Grok 4 released approximately 72,816 tons of CO2 into the atmosphere—roughly the same amount produced by 17,000 cars over a year of driving. The power capacity of data centers for AI has grown to 29.6 gigawatts: enough to supply electricity to the entire state of New York at peak load. Annual water consumption just for GPT-4o inference could exceed the drinking water volume for 12 million people. The total load of the entire AI infrastructure is comparable to the electricity consumption of Switzerland or Austria.

The ecological footprint of a model is gradually becoming part of the technical solution—in the near future, this will begin to appear in customer requirements.

Chart 2. The US and China: the gap has almost disappeared

For years, the US led the world in all key metrics—model size, performance, and number of research papers.

China has been methodically closing the gap. In February 2025, DeepSeek-R1 briefly matched the best American model. As of March 2026, the best US model leads the best Chinese model by only 2.7%. The US still produces more top-tier models and high-impact patents, while China leads in the number of publications, citations, and installation of industrial robots.

There is no longer a monopoly on the best models. It is necessary to monitor both ecosystems.

Chart 3. The US is losing AI talent

More AI researchers and developers work in the US than in any other country. However, the flow of specialists moving there has slowed sharply. The number of AI scientists who moved to the US has decreased by 89% since 2017. Over the past year, the decline was 80%.

Competition for AI specialists is global. Working from one’s own country for foreign companies is a strategy that is becoming increasingly popular.

Chart 4. AI takes gold at the Olympics, but can’t read a clock

Model capabilities are growing unevenly—and this is fundamentally important to understand when choosing development tools. Frontier models now surpass humans on PhD-level exams in natural sciences and in competitive mathematics tasks. Gemini Deep Think won the gold medal at the 2025 International Mathematical Olympiad. The success rate of AI agents on real-world tasks grew from 20% to 77.3% over the year. In cybersecurity, agents solve tasks in 93% of cases compared to 15%.

At the same time, robots handle real household tasks—folding clothes, washing dishes—only 12% of the time. Analog clocks are read correctly by models in only half of attempts. Multi-step planning and financial analysis remain weak points.

Chart 5. AI investments doubled in a year

Global corporate investments in AI reached $581.7 billion in 2025—a 130% increase compared to the previous year. Private investments amounted to $344.7 billion, an increase of 127.5%. The US invested $285.9 billion—23 times more than China’s $12.4 billion from private sources.

It is estimated that between 2000 and 2023, approximately $912 billion was directed through state investment mechanisms into various industries, including AI.

Chart 6. Junior developers under pressure

Labor productivity in AI-related professions is growing. At the same time, employment at the entry level in these same professions is declining. The number of employed developers aged 22 to 25 has decreased by almost 20% since 2024—while employment among their senior colleagues continues to grow.

A similar picture is observed in customer support and other professions with high AI impact. Corporate surveys indicate that managers plan further staff reductions, and the pace of planned cuts exceeds those already implemented.

Chart 7. AI in science: from assistant to discovery participant

AI is moving beyond being a tool for writing articles and checking calculations—it is beginning to participate directly in scientific discoveries.

The number of AI-related publications in natural, physical, and biological sciences grew by 26–28% year over year. In 2025, AI independently conducted a full weather forecasting pipeline for the first time: from simple meteorological observations to a complete forecast of temperature, wind, and humidity. In astronomy, the first foundation model was created, automating observations across 10 telescopes simultaneously.

Chart 8. The most powerful models are the most closed

The Foundation Model Transparency Index measures how openly major AI companies disclose information about their models: training data, compute volume, parameters, risks, and usage policies. In 2026, the average score on the index fell to 40 out of 100—down from 58 last year. The pattern noted by the authors is clear: the more powerful the model, the less is publicly known about it. The largest companies are increasingly hiding technical details of their flagship products.

For developers, this is a technical risk: lack of openness complicates debugging, auditing, and compliance with regulatory requirements when choosing a model for a product.

Chart 9. Public attitude toward AI: optimism and anxiety simultaneously

A global survey recorded a rise in optimism—59% of people expressed a positive attitude toward AI, compared to 52% a year ago. At the same time, anxiety levels also rose by 2%, reaching 52%. The US stands out against the general background: only 33% of Americans expect AI to improve their jobs, compared to a global average of 40%. Trust in government regulation of AI in the US is the lowest among all countries studied, at 31%.

Users perceive the technology ambiguously. Trust in AI products must be built consciously—through openness, user control, and explainability of decisions.

Chart 10. Generative AI is spreading faster than the internet

Generative AI reached 53% of the global population in three years—faster than the personal computer and the internet did in their time. The rate of adoption varies by country and correlates with per capita GDP. Singapore shows 61%, the UAE 54%. Despite its leadership in investments and development, the US ranks only 24th with a figure of 28.3%.

The estimated value of generative AI tools for American consumers reached $172 billion per year by the beginning of 2026, and the median value per user tripled—even though most tools remain free or nearly free.

The audience has formed. The habit of using AI tools is developing quickly. While the market has not yet stabilized, there is an opportunity to create new products.