AI Adoption (Consumer + Enterprise)
800 million#1 weekly active users of ChatGPT reported by OpenAI in early 2025, up from 300 million a year earlier.
Source: OpenAI DevDay 2024 announcement and subsequent OpenAI updates (2025)
39%#2 of US adults aged 18-64 reported using generative AI in their work or personal life in 2024.
28%#3 of US workers reported using generative AI on the job in the same NBER survey.
Source: NBER w32966, Bick et al. 2024
78%#4 of organizations reported using AI in at least one business function in 2024, up from 55% the prior year.
71%#5 of organizations reported using generative AI specifically in at least one function in 2024.
Source: McKinsey State of AI 2024
65%#6 of organizations were using generative AI regularly in 2024, nearly double the 33% reported in 2023.
Source: McKinsey State of AI 2024
Marketing & sales#7 remained the most common business function for generative AI deployment, with 34% of organizations using it there.
Source: McKinsey State of AI 2024
3.5 billion#8 monthly visits to ChatGPT's web interface reported in late 2024 according to publicly cited Similarweb data.
~270 million#9 weekly active users for Google's Gemini app reported by Sundar Pichai during Alphabet's Q4 2024 earnings call.
85%#10 of US developers said they had used or were currently using AI in their development workflow in 2024.
76%#11 of professional developers were using or planning to use AI tools in their development process in 2024, up from 70% in 2023.
62%#12 of developers said they were currently using AI in their development process, with another 14% planning to soon.
81%#13 of developers cited "increased productivity" as the top benefit of AI tools in their workflow.
~30%#14 of code in new files at Google was generated by AI assistance and accepted by reviewers, per Sundar Pichai.
8 in 10#15 US K-12 teachers reported having used AI tools at school by mid-2024.
API & Token Economics
$2.50#16 per million input tokens for OpenAI's GPT-4o list price as of late 2024, with $10.00 per million output tokens.
Source: OpenAI API pricing page
$0.15#17 per million input tokens for OpenAI's GPT-4o mini, roughly 1/17 the cost of GPT-4o input tokens.
Source: OpenAI API pricing page
$3.00#18 per million input tokens for Anthropic's Claude 3.5 Sonnet list price, with $15.00 per million output tokens.
Source: Anthropic API pricing page
200,000#19 token context window for Claude 3.5 Sonnet on the standard API tier.
Source: Anthropic model documentation
128,000#20 token context window standard for OpenAI's GPT-4o family.
Source: OpenAI models documentation
2 million#21 token context window available on Google's Gemini 1.5 Pro, the largest among major frontier APIs.
~280x#22 drop in cost per million tokens for GPT-3.5-class capability between November 2021 (Davinci) and mid-2024 (GPT-4o mini).
10x per year#23 approximate decline in price per million tokens at constant capability tracked by Epoch AI's LLM pricing analysis.
Source: Epoch AI LLM cost analysis
90%#24 discount available on OpenAI's batch API tier versus standard pricing for non-realtime workloads.
Source: OpenAI batch API pricing
50%#25 discount on Anthropic's prompt-cached input tokens versus uncached, with 90% off on cache reads.
$3.7 billion#26 in 2024 revenue reported for OpenAI, the bulk from API and ChatGPT subscriptions.
$1 billion#27 annualized revenue run-rate disclosed by Anthropic in late 2024.
~73%#28 share of Anthropic 2024 revenue from API customers rather than direct chat subscriptions, per The Information.
$20#29 per month standard ChatGPT Plus consumer subscription price unchanged since 2023.
Source: OpenAI ChatGPT pricing page
$200#30 per month price of ChatGPT Pro tier launched in December 2024 with unlimited o1 access.
Frontier Model Release Cadence
51#31 notable machine-learning models released by industry in 2023 alone, per Epoch AI's tracking cited in the AI Index.
Source: Stanford HAI AI Index 2024
~3 months#32 typical interval between major frontier-model releases by the top three labs (OpenAI, Anthropic, Google DeepMind) during 2024.
GPT-4 to GPT-4o#33 took roughly 14 months (March 2023 to May 2024), the slowest gap between OpenAI flagship releases since 2020.
Source: OpenAI GPT-4o launch post
o1-preview#34 launched September 12, 2024, was OpenAI's first publicly released reasoning-trained model.
Source: OpenAI o1 launch announcement
3#35 distinct Claude 3 family tiers released by Anthropic on March 4, 2024 (Haiku, Sonnet, Opus).
June 20, 2024#36 release date of Claude 3.5 Sonnet, which Anthropic positioned as outperforming Claude 3 Opus at a fraction of the price.
Gemini 1.5#37 Pro launched February 15, 2024, introducing the 1M-token context window publicly previewed by Google DeepMind.
Llama 3#38 released April 18, 2024 in 8B and 70B sizes, with the 405B Llama 3.1 released July 23, 2024.
Source: Meta AI Llama 3 announcement and Llama 3.1 announcement
15.4 trillion#39 tokens used to pretrain Llama 3, more than seven times the volume used for Llama 2.
Source: Meta AI Llama 3 model card
DeepSeek V3#40 released open-weight on December 26, 2024, claimed pretraining cost of approximately $5.6M for a 671B-parameter MoE model.
17 months#41 elapsed between OpenAI's GPT-4 release (March 2023) and the o1-preview release (September 2024) marking the start of the reasoning paradigm.
Source: OpenAI GPT-4 research page and OpenAI o1 announcement
Open-Source vs Closed-Source Share
~100#42 notable open-weight models released in 2024, roughly double the count released in 2023 per Epoch AI tracking.
Source: Epoch AI Notable Models database
1.7 million+#43 public model repositories on Hugging Face by late 2024.
Source: Hugging Face 2024 year-in-review
650 million+#44 cumulative downloads of Llama models reported by Meta in late 2024, a 10x increase year-over-year.
Source: Meta Llama download disclosure
~1.5 years#45 typical lag between a closed-source frontier capability and the first open-weight model matching it on standard benchmarks per State of AI 2024.
Source: State of AI Report 2024
Llama 3.1 405B#46 was the first open-weight model to substantially match closed-source frontier scores on MMLU, GSM8K and HumanEval.
$0.27#47 per million input tokens for DeepSeek V3 hosted inference, the lowest publicly listed price for a 600B+ parameter open model in early 2025.
~70%#48 share of new arXiv cs.CL papers using or fine-tuning open-weight models in 2024 versus closed APIs, per ACL Anthology indexing.
Apache 2.0 / Llama license#49 dominate the open-weight ecosystem, with Apache 2.0 used on Mistral, Qwen, and DeepSeek releases and Meta's custom Llama Community License governing Llama 3 family weights.
Source: Meta Llama 3 Community License
Qwen2.5#50 released by Alibaba in September 2024, became the most-downloaded open-weight model family on Hugging Face in late 2024.
Source: Hugging Face 2024 year-in-review
Enterprise AI Deployment Patterns
RAG#51 remained the most common enterprise generative-AI architecture in 2024, with 51% of deployments incorporating retrieval-augmented generation per McKinsey.
Source: McKinsey State of AI 2024
63%#52 of organizations reported using foundation models from external providers without modification, per Stanford AI Index 2025.
Source: Stanford HAI AI Index 2025
38%#53 of organizations had production-grade generative AI deployed in 2024, up from 19% in 2023.
Source: McKinsey State of AI 2024
44%#54 of organizations reported a measurable cost reduction from generative AI in at least one business function.
Source: McKinsey State of AI 2024
63%#55 of organizations reported a revenue increase from generative AI deployment in at least one function.
Source: McKinsey State of AI 2024
Inaccuracy#56 remained the top reported risk from generative AI deployments, cited by 63% of respondents in McKinsey's 2024 survey.
Source: McKinsey State of AI 2024
Cybersecurity#57 and intellectual-property infringement were the next most-cited generative AI risks, at 50% and 40% of respondents respectively.
Source: McKinsey State of AI 2024
87%#58 of enterprise generative-AI workloads ran on a third-party hyperscaler in 2024 (AWS, Azure, GCP) rather than on-premises.
Source: Stanford HAI AI Index 2025
~2/3#59 of enterprises with generative-AI deployments report using two or more model providers to mitigate vendor lock-in.
Source: State of AI Report 2024
Customer support#60 was the leading high-ROI generative-AI use case identified in the McKinsey State of AI 2024 survey.
Source: McKinsey State of AI 2024
AI Tool Spend by Category
$235 billion#61 total worldwide AI software, hardware, and services spending in 2024 per IDC's Worldwide AI and Generative AI Spending Guide.
$632 billion#62 projected worldwide AI spending by 2028 per the same IDC guide, implying a 29% CAGR from 2024.
Source: IDC Worldwide AI Spending Guide
$13.8 billion#63 enterprise generative-AI software spending in 2024, six times the 2023 figure of $2.3 billion per Menlo Ventures.
29%#64 of 2024 enterprise generative-AI spending went to code-generation and developer-productivity tools, the largest category.
~12%#65 of enterprise generative-AI software budget went to customer-support and contact-center applications.
$25.2 billion#66 private investment into generative AI startups globally in 2023, an 8x increase from 2022.
Source: Stanford HAI AI Index 2024
$33.9 billion#67 private investment in generative AI in 2024 per Stanford AI Index 2025.
Source: Stanford HAI AI Index 2025
$67.2 billion#68 US private AI investment in 2023, far ahead of China at $7.8 billion and the UK at $3.8 billion.
Source: Stanford HAI AI Index 2024
GitHub Copilot#69 reported 1.8M paying individual subscribers and 77,000+ organizations as customers in fiscal year 2024 disclosures.
Source: GitHub Octoverse 2024 report
$10/month#70 standard GitHub Copilot Individual subscription price, unchanged since 2022.
Source: GitHub Copilot plans page
AI Labor Displacement & Productivity
26%#71 productivity uplift among customer-support agents using a generative-AI assistant, with the largest gains for newer workers, in a 5,000-agent randomized study.
14%#72 increase in resolutions per hour for novice support agents, versus 0% gain for highly experienced agents, in the same study.
55.8%#73 faster task completion for developers given GitHub Copilot in a controlled study by GitHub Research.
37%#74 reduction in time spent on a writing task and improved output quality among professionals given access to ChatGPT, in an MIT-led randomized field study.
12.2%#75 performance improvement for management consultants on a creative-product task when given GPT-4, in a Harvard-BCG controlled study, with quality also rising 40%.
23%#76 of US workers in 2024 said they expected AI to make their job obsolete within five years, up from 19% the prior year.
~25%#77 of work tasks across the US economy are exposed to large-language-model automation per OpenAI and University of Pennsylvania researcher analysis.
19%#78 of US workers held jobs in which at least 50% of work tasks are highly exposed to LLM automation per the same study.
Compute & Training Costs
~$78 million#79 estimated compute cost to train GPT-4 per Stanford AI Index 2024 cost estimates.
Source: Stanford HAI AI Index 2024
~$191 million#80 estimated compute cost to train Google Gemini Ultra, the most expensive model training disclosed by the AI Index 2024.
Source: Stanford HAI AI Index 2024
2.4x#81 annual growth in training-compute requirements for frontier models since 2010, per Epoch AI's compute scaling analysis.
15.4 trillion#82 training tokens used for Llama 3, processed on two custom-built 24,000-GPU H100 clusters operated by Meta.
Source: Meta AI Llama 3 announcement
~$5.6 million#83 compute cost claimed by DeepSeek for the V3 pretraining run, two orders of magnitude below Western frontier-lab estimates.
$30,000+#84 typical end-customer price of a single Nvidia H100 SXM GPU during the 2023-2024 supply crunch reported by The Information.
~3.4 million#85 Nvidia H100-class GPUs shipped to AI labs and hyperscalers during 2024, per Omdia channel-tracker estimates cited in State of AI 2024.
Source: State of AI Report 2024