Stanford HAI's latest report is out: China's models catch up with the U.S., 95% of companies see zero returns on AI investments

On April 13, U.S. time, the Stanford University Human-Centered Artificial Intelligence Institute (Stanford HAI) released the "2026 AI Index Report," which spans 423 pages. This report, published annually since 2017, has become the most comprehensive annual document tracking the development of the artificial intelligence industry, covering multiple dimensions such as the number and sources of model releases, capital inflows in the industry, changes in the labor market, energy consumption and environmental impact, as well as public attitudes. The 2026 report reveals an industry whose capabilities are rapidly breaking through, while also raising urgent questions about environmental costs, technological transparency, and who can truly benefit from this technology. Below is a summary of the latest report: 01 The Last 2.7% Game: The "Equalization Era" of Model Performance between China and the U.S. For a long time, Silicon Valley has been regarded as the sole heart of global AI, but the latest data from the Stanford report shows that this unipolar pattern is crumbling. According to data from the research organization Epoch AI, as of March 2026, the leading U.S. AI company Anthropic, led by Dario Amodei, has its most advanced model performing only 2.7 percentage points ahead of China's strongest competitor. Performance differences between top models in China and the U.S. The turning point in this competition occurred in February 2025, when the R1 model released by DeepSeek briefly matched the performance of U.S. models, after which both sides entered a high-frequency performance iteration period. This year's report editor Nestor Maslej pointed out that the repeated changes in this leading advantage signify that global top AI research and development has entered a certain degree of "technological equalization." In terms of output quantity, the U.S. released 50 notable top models in 2025, while China closely followed with 30. From 2003 to 2025, the number of AI models in the U.S., China, and Europe has shown an upward trend. Although the U.S. maintains a lead in absolute model numbers and total private investment, with $285.9 billion compared to China's $12.4 billion, the report specifically warns policymakers that the figures greatly underestimate China's actual investment. If we extend the timeline, since 2000, the funds injected by the Chinese government-guided funds into AI companies have totaled approximately $184 billion. This model of "government setting the stage, enterprises performing" has allowed China to firmly hold the global first position in the number of AI publications, paper citation shares, and patent authorizations According to data from the International Federation of Robotics, in 2024, China will install 295,000 industrial robots, Japan about 44,500, the United States 34,200, and Germany and South Korea significantly lower than China. More tangible data shows that China has also taken the lead in the installation of industrial robots, indicating that AI is becoming embodied in the physical world, and China may have a deeper engineering foundation. Some executives from American tech companies are uneasy about this narrowing gap, attributing it to China's efficiency in utilizing open-source code and engineering implementation. However, it is undeniable that when the gap in technical indicators narrows to single digits, future competition will no longer depend solely on whose model has run a few percentage points better, but on who can first convert these expensive computing powers into actual productivity gains. So far, the number of AI researchers and developers in the United States far exceeds that of any other country, but the speed at which these experts are flowing into the U.S. is sharply slowing. Since 2017, the number of AI scholars moving to the U.S. has decreased by 89%. This decline is accelerating, with an 80% drop in just the past year. The inflow of AI talent to the U.S. has significantly slowed down. 02 The U.S. has maintained its computing power throne but has shifted the pressure to the power grid. The U.S. maintains a clear lead in the data center sector. Data from Stanford HAI shows that the U.S. has 5,427 data centers, while China has 449, and Germany and the UK each have about 525. By the end of 2025, the total power capacity of AI data centers is expected to reach 29.6 gigawatts, roughly equivalent to New York State's peak electricity demand. Statistics from Epoch AI, using the computing power of NVIDIA H100e as a benchmark, show that NVIDIA GPUs account for over 60% of the global AI total computing power, with Google and Amazon ranking second and third. The expansion of data center scale comes with significant environmental costs. The Stanford HAI report estimates that training xAI's Grok 4 model generates about 72,816 tons of carbon dioxide equivalent, higher than the total carbon emissions of about 1,000 ordinary cars over their entire lifecycle. Ray Perrault, co-director of the AI Index Steering Committee, stated that this estimate is based on publicly reported data, xAI statements, and other unverifiable sources, and should be interpreted with caution. He also noted that Epoch AI independently estimates Grok 4's emissions to be about 140,000 tons of carbon dioxide. The report estimates that carbon emissions from training AI models will show an upward trend from 2012 to 2025, with a sharp increase in emissions resulting from the training of Grok 3 and Grok 4 in 2025. Model inference also generates an environmental burden. The Stanford HAI report estimates that the annual water consumption for GPT-4o inference alone exceeds the drinking water needs of 12 million people. The carbon emissions of the least efficient models are more than 10 times higher than those of the most efficient. DeepSeek's V3 model consumes about 23 watts when responding to moderately long prompts, while Claude 4 Opus consumes about 5 watts. Local communities are beginning to resist the construction of data centers. According to a report from the Data Center Observers organization, in the past two years, $64 billion worth of data center projects in the U.S. have been shelved or delayed due to local opposition, with at least 142 active groups involved in 24 states. Among elected officials publicly opposing data center projects, 55% are Republicans and 45% are Democrats. In Warrenton, Virginia, every town council member who voted in favor of the Amazon data center project was subsequently defeated in the elections. Some resistance incidents have involved violence. A city council member in Indianapolis, who publicly supported the re-planning of a data center in their district, reported that someone shot at their home in early April, leaving a handwritten note on the doorstep that read "No data center." The council member and their eight-year-old son were unharmed. 03 Rapid Improvement in AI Model Performance, Significant Shortcomings in Specific Tasks Over the past decade, the performance of AI models has improved rapidly, with an accelerating trend. The speed at which multimodal large language models conquer new benchmarks is close to the speed of benchmark releases. The evolution of AI agents is particularly notable, with the OSWorld benchmark (testing autonomous computing) and the SWE-Bench Verified benchmark (testing autonomous coding) showing the steepest score curves. From 2012 to 2025, the performance benchmarks of AI in multiple tasks compared to human performance. For example, in image classification, AI surpassed human performance early on, and by the 2020s, models were approaching or exceeding human baselines in multiple tasks. The "Ultimate Human Exam" benchmark includes questions contributed by experts from various fields, representing the most difficult problems in each domain. The 2025 report shows that the highest-ranked OpenAI o1 model answered only 8.8% of the questions correctly, with accuracy later rising to 38.3%. As of April 2026, the highest-scoring models—Anthropic's Claude Opus 4.6 and Google's Gemini 3.1 Pro—have exceeded 50%. Perlo notes that benchmark testing may not necessarily reflect the real-world performance of models: "Knowing that the legal reasoning benchmark has a 75% accuracy rate does not tell us how well it adapts to activities in a law firm." Progress has been made in the application of AI in the medical field. Over the past two years, the number of publications on AI for drug discovery has more than doubled. The number of publications on multimodal biomedical AI (used for simultaneous examination of medical images and text) is 2.7 times that of two years ago. From 2018 to 2025, the number of publications on AI for drug discovery has continued to increase, with a faster growth rate in the past two years. However, AI models perform poorly on some common tasks. The ClockBench test of multimodal LLMs' ability to read simulated clocks shows that the best-performing OpenAI GPT-5.4 has an accuracy of only 50%. Most models score significantly lower. Anthropic's Claude Opus 4.6 has an accuracy of only 8.9% in correctly reading the time, while this model typically scores higher on other benchmarks. The ClockBench test shows that the accuracy of different LLMs in reading simulated clocks ranges from 8.9% to 50.60%, with overall performance being low. Pelo stated that this reflects a more general issue: "There is a line of research suggesting that when systems are asked questions that combine language with other modalities (such as images or audio tones), the language component bears the vast majority of the burden, to the extent of completely ignoring non-language information." In addition, robots still have a significant gap in handling household chores. The report indicates that they have only a 12% success rate in real household tasks such as folding clothes or washing dishes. 04 AI is Racing Ahead in Medical Research The application of artificial intelligence in the medical field has made rapid progress. The report shows that over the past two years, the number of publications on artificial intelligence for drug discovery has more than doubled. The number of publications on multimodal biomedical artificial intelligence—used for simultaneous examination of medical images and text—is 2.0 times that of two years ago. Several noteworthy projects emerged in 2025. For the first time, AI has run the complete weather forecasting process end-to-end, receiving raw real-time meteorological observation data and directly outputting final forecasts such as temperature, wind speed, and humidity. Astronomy has also established its first foundational model, achieving automated observations with one telescope. In clinical applications, tools that automatically generate clinical records from patient visits were widely adopted in 2025. In multiple hospital systems, doctors reported a reduction in note-taking time of up to 83%, and a significant decrease in feelings of professional burnout. However, the report also notes that, aside from certain tools, the value of clinical AI remains speculative. A review of over 500 clinical AI studies found that nearly half of the studies relied on exam-style questions rather than real patient data, with only 5% of studies using real clinical data Another growth area for medical artificial intelligence is digital twins, which are dynamic, data-linked computational representations of individual patients that can be updated over time and support prediction, simulation, and treatment optimization. The number of related publications has risen from nearly zero in 2015 to 372 by 2025. 05 Employee Efficiency Improvement, Yet Companies Remain Weak Focusing on specific tasks, AI brings measurable efficiency improvements. Customer support agents solve nearly 15% more issues per hour, software developers using GitHub Copilot complete 26% more pull requests, and marketing teams using AI for ad creation see a 50% increase in per capita output. When expanded to the overall U.S. economy, the productivity growth rate is projected to be 2.7% in 2025, about twice the average of the previous decade. However, the report cites the Wharton budget model from the University of Pennsylvania, estimating that AI's actual contribution to total factor productivity is only 0.01 percentage points, close to zero. The report also notes that for tasks requiring deeper reasoning, AI tools actually reduce human efficiency. The speed of AI-assisted open-source developers decreased by 19%. Engineers relying on AI for learning did not show speed improvements and instead experienced what researchers call a "learning penalty," which may slow their professional development over time. Labor data shows clear generational differences. By September 2025, the number of software developers aged 22 to 25 in the U.S. is expected to decline by nearly 20% from the peak in 2022, while the number of older developers continues to grow. A similar trend is observed in the customer support agent field. From 2021 to 2025, the trend in the allocation of software developers and customer support personnel by age shows a significant decrease in the number of early-career workers, while the number of mid- to late-career workers remains stable or increases. These changes are difficult to completely separate from macro trends. The report points out that unemployment rates across various occupations are rising, and contrary to expectations, the unemployment rate for workers with the lowest exposure to AI is increasing more than for those with the highest exposure. One-third of surveyed companies expect to reduce their workforce in the coming year due to AI impacts. An independent study from the Massachusetts Institute of Technology found that 95% of companies received zero returns on approximately $35 to $40 billion in AI investments, with only 5% of companies successfully achieving large-scale deployment of tools. 06 Global AI Investment Grows Significantly, U.S. Leads but China is Undervalued According to data from AI analytics firm Quid, global AI investment is set to reach a record high of over $581 billion in 2025, more than double the $253 billion in 2024 and surpassing the previous record of $360 billion in 2021. Unlike 2021, which was dominated by mergers and acquisitions, the record investment in 2025 is led by private investments. Most of the funding flows to the U.S., with AI investment in the U.S. exceeding $344 billion in 2025 Quid data shows global corporate AI investment by activity type from 2013 to 2025. Investment rose in 2021, declined from 2022 to 2024, and then surged again in 2025. However, the report emphasizes that comparisons based solely on private investment may underestimate the amount of funding China has allocated to artificial intelligence. The report estimates that from 2000 to 2023, $912 billion of government-guided funds have been deployed across various industries, including artificial intelligence. In terms of computing power, according to statistics from Epoch AI based on the computing power of NVIDIA H100e, NVIDIA GPUs account for over 60% of the total global AI computing power, with Google and Amazon ranking second and third, respectively. Global AI computing power has grown 3.3 times annually since 2022, and total computing power has increased 30 times since 2021 (the first year tracked). 07 AI Adoption Rate Hits Record High, Surpassing the Internet The adoption rate of generative artificial intelligence continues to accelerate. The report notes that generative AI reached a 53% adoption rate among the population within three years, faster than the adoption rates of personal computers or the internet. However, the adoption rate varies by country and is closely related to per capita GDP. Some countries have adoption rates higher than expected, with Singapore reaching 61% and the UAE at 54%. The United States ranks 24th with a 28.3% adoption rate. By early 2026, the estimated annual value of generative AI tools for American consumers is projected to reach $172 billion, with the average value per user doubling between 2025 and 2026. In the education sector, four-fifths of American high school and college students use AI for school-related tasks. However, only half of middle and high schools have established AI policies, and only 6% of teachers report that these policies are clear. Formal education is significantly lagging behind the use of AI. In terms of software development platforms, by 2025, the number of AI-related projects on GitHub is expected to rise to 5.58 million, nearly five times the number in 2020, and a 23.7% increase from 2024. The number of projects with at least 10 stars and the total number of stars received by AI projects have also grown at a similar pace. The open-source agent-based AI software OpenClaw has received 352,000 stars. From 2011 to 2025, the number of AI projects on GitHub grew from nearly zero to 5.58 million, with a significant acceleration in recent years. Pelo stated, "The intensity of GitHub usage is likely highly correlated with the intensity of AI usage." However, according to data from the activity tracking site "Agents in the Wild," most GitHub activities are still performed by humans In the past decade, the number of computer science publications related to artificial intelligence has increased from 102,000 to 258,000, more than doubling. As of 2024, over 68% of publications come from academia, with government and industry contributing approximately 11.5% and 12.5%, respectively. The growth is primarily driven by the fields of machine learning, computer vision, and generative artificial intelligence. 08 Americans Are Most Cautious About AI According to Ipsos survey data, the proportion of respondents who believe that AI has "more benefits than drawbacks" has risen from 55% in 2024 to 59%. The percentage of respondents who say they have a "good understanding" of AI has slightly increased from 67% to 68%. At the same time, 52% of respondents feel "nervous" about using AI products and services. From 2022 to 2025, the distribution of responses from respondents regarding multiple AI-related statements shows a slight increase in optimism, while negative sentiments have also increased. There are significant differences between countries. Southeast Asian countries such as China, Malaysia, Thailand, Indonesia, and Singapore tend to have a positive attitude towards artificial intelligence. The countries with the largest year-on-year positive shifts are Germany (up 12%), France (10%), and the Netherlands (10%). Colombia experienced the largest negative shift (down 6%), contrary to previous trends. The American public's attitude is more cautious than that of other countries. Only 33% of Americans expect AI to improve their jobs, compared to a global average of 40%. The proportion of the American public expecting AI to eliminate jobs rather than create new ones is also the highest. There is an even greater disparity in trust in government AI regulation. In Singapore, 81% of respondents express trust in regulatory agencies, while in the U.S., only 31% do, ranking last among the surveyed countries. Several European countries and Japan also show lower levels of trust. Trust levels are generally higher in Asian and South American countries. The survey on trust in government AI regulation shows Singapore at the top with 81%, while the U.S. is at the bottom with 31%. According to a Pew survey, there is a significant divide between experts and the public regarding the future of artificial intelligence. 73% of experts believe that AI will have a positive impact on the way people work, while only 23% of the American public shares this view. Experts are also more optimistic about the impact of AI on education and healthcare than the public, but both groups agree that AI will have a negative impact on elections and personal relationships. The report also highlights a concerning trend: the most capable modern models today are also the least transparent. Large and powerful models are concentrated in the hands of the largest AI companies, which are increasingly keeping training code, dataset sizes, and parameter counts confidential The Basic Model Transparency Index measures the openness of major artificial intelligence companies in disclosing details about their model training data, computational resources, capabilities, risks, and usage policies. The report shows that the average score of the index has dropped from 58 points in 2024 to 40 points this year. The index particularly notes that the most capable models often disclose the least information. Yolanda Gil, a computer scientist at the University of Southern California and co-author of the report, stated, "There is a lot we don't know about predicting model behavior." She mentioned that this lack of transparency makes it difficult for independent researchers to study how to make AI models safer. Conclusion The AI Index report from Stanford University is essentially a "survival manual" co-written by humans and machines. It tells us in 423 pages that AI has moved past its childhood phase, where it could survive on storytelling, and is now entering adulthood, facing the rites of passage of commercial returns, energy bottlenecks, and global sovereignty competition. For China, a 2.7% gap means that the catch-up has been completed, and the next task is to leverage patent licensing and the scale advantages of industrial robots to be the first to reap the benefits of AI in the real economy. For the United States, how to maintain the innovative vitality of private investment and address high costs and energy issues will determine whether it can continue to lead in the next decade. This $405 billion gamble has just entered the second half, and the ultimate winner may not be the one with the strongest computing power, but certainly the one that can best adapt to physical realities and provide the most commercial value. Tencent Technology Risk Warning and Disclaimer The market has risks, and investment requires caution. This article does not constitute personal investment advice and does not take into account the specific investment goals, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article align with their specific circumstances. Investment based on this is at their own risk