The Stanford Institute for Human-Centered Artificial Intelligence (HAI) released its 2026 AI Index Report, and it reads like two documents stitched together. One tells the story of a technology achieving things that seemed impossible 18 months ago. The other tells the story of every system around that technology, from workforce pipelines to transparency standards to environmental accounting, failing to keep pace.
The Stanford Institute for Human-Centered AI (HAI) is an interdisciplinary institute established in 2019 to advance AI research, education, policy, and practice. Stanford HAI brings together thought leaders from academia, industry, government, and civil society to shape the development and responsible deployment of AI.
The mission of the AI Index is to provide unbiased, rigorously vetted, and globally sourced data for policymakers, researchers, journalists, executives, and the general public to develop a deeper understanding of the complex field of AI. To achieve this, we track, collate, distill, and visualize data relating to artificial intelligence.
Every year I look forward to this report. Kudos to Raymond Perrault, Yolanda Gil , Erik Brynjolfsson , Vanessa Parli , Russell Wald, Jack Clark, Russ Altman, Carla E. Brodley, Virginia Dignum, James Landay, James Manyika, Juan Carlos Niebles, Yoav Shoham , Sha Sajadieh, Loredana Fattorini, Nestor Maslej , Juan N. Pava, Lapo Santarlasci , Sukrut Oak & Henry T. Zhang, and many, many more who contributed to this incredibly important initiative. The full report runs 423 pages (read it here). and what follows are the 10 numbers that struck me as important for directors and executives to have top of mind.
1. $581.7 Billion: The Investment Surge
Global corporate AI investment hit $581.7 billion in 2025, up 130% from the prior year. Private investment alone reached $344.7 billion, a 127.5% increase. The United States accounted for $285.9 billion of that, 23.1 times greater than China's $12.4 billion in tracked private investment.
To put that in context: total AI investment has grown roughly fortyfold since 2013. And unlike the 2021 record of $360 billion, which was driven by mergers and acquisitions, 2025's surge was led by private investment into AI companies. The capital markets are not hedging. They are going all in. Is this a bubble? No one knows, but this is capital being deployed like never before.
The United States also led in new AI companies, with 1,953 newly funded startups in 2025, more than 10 times the next closest country (the U.K.). The largest AI companies, including OpenAI (private) and Anthropic (private), are heading toward IPOs later this year.
2. 2.7%: The Vanishing U.S.-China Gap
For years, U.S. AI models held a comfortable lead over Chinese competitors. That lead has effectively evaporated. As of March 2026, Anthropic's top model leads China's best by just 2.7 percentage points, and the two countries have traded the number-one position multiple times since early 2025.
The U.S. still produces more top-tier models (50 "notable" models in 2025 vs. China's 30, per Epoch AI) and generates higher-impact patents. But China leads in total patent output, publication volume, citations, and industrial robot installations (295,000 robots installed in 2024, compared to 34,200 in the U.S.).
For business leaders, this has immediate implications. Your AI vendor's competitive moat may be thinner than you think. The supply chain for AI capability is globalizing faster than the regulatory frameworks designed to govern it.
3. 66%: The Agentic Leap
This may be the single most consequential number in the report for enterprise strategy.
On OSWorld, a benchmark that tests AI agents on real computer tasks across operating systems (Ubuntu, Windows, macOS), the best model jumped from roughly 12% success in early 2024 to 66.3% today. That is within 6 percentage points of the human baseline of 72.35%. On WebArena, which tests autonomous web agents, success rates climbed from 15% in 2023 to 74.3% in early 2026, now within 4 percentage points of human performance.
On SWE-bench Verified, which measures autonomous software engineering, performance rose from 60% to near 100% of meeting the human baseline in a single year. AI agents handling cybersecurity issues solved problems 93% of the time, up from 15% in 2024.
These are not chatbots answering questions. These are autonomous systems completing multi-step tasks across real software environments. The agentic enterprise is no longer a slide deck concept. It is arriving in production, function by function.
Yet agent deployment still sits in single digits across nearly all business functions. The gap between what agents can do on benchmarks and what companies are actually deploying represents both the largest near-term opportunity and the largest near-term risk for enterprise AI strategy.
4. 53%: Faster Than the Internet
Generative AI reached 53% population adoption within three years of its first mass-market product (the launch of ChatGPT on November 30, 2022). That is faster than the personal computer and faster than the internet over comparable time frames.
But adoption varies dramatically by country, and correlates strongly with GDP per capita. Singapore leads at 61%. The United Arab Emirates is at 54%. The U.S. ranks 24th globally at 28.3%.
Organizational adoption tells a different story: 88% of organizations in the tech sector now use AI. The estimated value of generative AI tools to U.S. consumers reached $172 billion annually by early 2026, with the median value per user tripling between 2025 and 2026.
The adoption curve is steep, but it is also uneven. Companies that assume uniform AI literacy across their workforce, their customers, or their markets are planning against data that does not exist.
5. 20%: The Entry-Level Squeeze
Here is the number that should appear on every CHRO's desk this week.
Employment among U.S. software developers aged 22 to 25 fell nearly 20% from 2024, even as headcount for developers aged 26 and older continued to grow. The same pattern is emerging in customer service roles. Studies now show 14% to 26% productivity gains in customer support and software development, with weaker or negative effects in tasks requiring more judgment.
Stanford's framing is direct: "The disruption is targeted and just beginning." Firm surveys indicate executives expect planned headcount reductions to outpace recent cuts. Among early-career workers in the most AI-exposed occupations, employment has fallen roughly 16% relative to the least-exposed, after controlling for firm-level and industry effects.
This is not a forecast. It is a measurement. And it raises questions every executive team should be answering now: What is our plan for the entry-level pipeline that feeds our senior talent in five years? Are the productivity gains we are capturing today creating a skills gap we will pay for tomorrow?
6. 58 → 40: The Transparency Collapse
The Foundation Model Transparency Index, which measures how openly AI companies disclose training data, compute, capabilities, risks, and usage policies, saw its average score drop from 58 to 40 in a single year. IBM leads at 95. Writer follows at 72. xAI (Grok) and Midjourney score just 14.
The pattern is striking: the most capable models now disclose the least. Data acquisition transparency averages just 31% across major providers. Data properties disclosure sits at 15%. Compute disclosure: 26%.
For any business building on top of foundation models, this trend creates a compounding risk. You are increasingly dependent on systems whose training data, failure modes, and risk profiles their own makers will not disclose. The transparency you need to conduct due diligence on your AI vendors is declining precisely as your dependence on those vendors grows.
7. 72,816 Tons of CO₂: The Environmental Bill
Training xAI's Grok 4 generated an estimated 72,816 tons of CO₂ equivalent, roughly equal to driving 17,000 cars for a year. That is up from 5,184 tons for GPT-4 and 8,930 tons for Llama 3.1 405B. Epoch AI independently estimates the Grok 4 figure may be as high as 140,000 tons.
AI data center power capacity rose to 29.6 gigawatts, comparable to powering the entire state of New York at peak demand. Annual GPT-4o inference water use alone may exceed the drinking water needs of 12 million people. The cumulative power demand of AI systems is now comparable to the national electricity consumption of Switzerland or Austria.
Inference efficiency varies wildly: the least efficient models consume more than 10 times the carbon of the most efficient for the same task. DeepSeek's V3 models use roughly 23 watts per medium-length response; Claude 4 Opus uses approximately 5 watts.
For companies with climate commitments, sustainability reporting obligations, or ESG-conscious investors, these numbers create a new category of indirect emissions that most reporting frameworks have not yet addressed.
8. 362: AI Incidents Rising
Documented AI incidents rose to 362 in 2025, up from 233 in 2024. The six-month moving average hit 326, with a single-month peak of 435 in January 2026.
Stanford's report adds a critical finding: improving one responsible AI dimension (such as safety) can degrade another (such as accuracy). Almost all leading frontier AI model developers report results on capability benchmarks, but reporting on responsible AI benchmarks remains inconsistent.
Hallucination rates across 26 models range from 22% to 94%. The jagged frontier of AI capability, where models earn gold medals at the International Mathematical Olympiad but read an analog clock correctly just 50.1% of the time, is not a technical curiosity. It is an operational risk that current testing frameworks do not adequately capture.
For enterprises deploying AI agents that take actions on behalf of the company (approving transactions, responding to customers, writing code that ships to production), the question is no longer "Does this model perform well on benchmarks?" It is: "Do we understand where this model fails, and who is accountable when it does?"
9. 89%: The Talent Drain
The number of AI researchers and developers moving to the United States has dropped 89% since 2017. That decline is accelerating: down 80% in the last year alone.
The U.S. still hosts the most AI researchers of any country by far, and it hosts 5,427 data centers, more than 10 times any other nation. But the inflow that built that lead is drying up. Meanwhile, AI-related computer science publications have more than doubled over the past decade, from 102,000 to 258,000, with over 68% originating in academia.
On GitHub, AI-related projects reached 5.58 million through 2025, a roughly fivefold increase since 2020. The developer community is growing globally, even as the U.S. share of new talent shrinks.
For companies competing for AI engineering talent, the implications are clear: your hiring pool is becoming more global and more competitive. The assumption that the best AI talent will naturally flow to U.S.-based companies no longer holds.
10. 31%: America's Trust Deficit
Only 31% of Americans trust their government to regulate AI, the lowest of any country surveyed. Singapore leads at 81%. Globally, the EU is trusted more than either the United States or China to regulate AI effectively.
This is not just a policy statistic. It is a business environment indicator. In a low-trust regulatory environment, companies face higher compliance uncertainty, more fragmented state-level regulation, and greater reputational risk from AI deployments that go wrong. The absence of a trusted federal framework means every company is effectively building its own governance standard, whether it knows it or not.
Meanwhile, 59% of global respondents said AI's benefits outweigh its drawbacks (up from 52%), and 52% said AI products make them "nervous." The public is simultaneously adopting AI faster than any previous technology and expressing deep skepticism about the institutions managing it.
What This Means for the Agentic Enterprise
Read these 10 numbers together and one pattern emerges: the capability curve is pulling away from every other curve.
AI agents are approaching human-level performance on real computer tasks (66.3% on OSWorld, 74.3% on WebArena). Investment has doubled in a year ($581.7B). Adoption is outpacing the internet (53% in three years). But transparency is collapsing (58 → 40). AI incidents are rising (362 in 2025). The talent pipeline is narrowing (89% drop in U.S.-bound AI researchers). And the entry-level workforce is already being displaced (20% decline among young developers).
For any business leader thinking about the agentic enterprise, one thing is clear from Stanford's data: the agents are ready before the institutions are. The models can complete real tasks. The question is whether your organization has the governance infrastructure, the workforce strategy, the vendor due diligence, and the accountability frameworks to deploy them responsibly.
Three things I would do this week with this report:
Pull the vendor transparency data. Figure 3.8.2 in the report ranks every major foundation model on transparency. If your company is building on a model that scores below 40 (the current average), your risk committee and procurement team should know that, and you should properly understand your exposure.
Benchmark your workforce data against Figure 4.4.29. If your early-career hiring trends in AI-exposed roles mirror the 20% decline Stanford documents, your talent pipeline has a structural problem that no amount of productivity gains will solve in five years.
Assign the agentic AI section to your operations leadership. Chapter 2.6 of the report details where agents are approaching human performance and where they still fail. Any executive evaluating agentic deployments needs to understand the jagged frontier, not as an abstract concept, but as the specific capability profile of the systems they are buying.
Stanford gave us the data. What we as leaders do with it is the test.
The full 2026 AI Index Report is available at hai.stanford.edu/ai-index/2026-ai-index-report
This article was originally posted on Alpha BoardBrief.
