The Daily Receipts

AI training's carbon footprint has scaled by orders of magnitude. Stanford's 2026 AI Index estimates that training AlexNet in 2012 produced about 0.01 tons of CO2 equivalent, while training xAI's Grok 4 in 2025 produced roughly 72,816 tons of CO2 equivalent — over 1,150 times the lifetime emissions of an average car

McKinsey's State of AI survey, cited in Stanford's 2026 AI Index, asked organizations already using AI where they had seen the clearest gains. The single most commonly cited outcome in 2025 was improved innovation, named by 64% of respondents, more than any other organizational outcome measured

Epoch AI's FrontierMath Erdős benchmark tasks AI models with proving or disproving 68 significant unsolved Erdős problems in Lean, each within a fixed $300, 72-hour budget. Running five frontier models in August 2026, only one produced any verified solution: GPT-6 Astra solved 2 of 68 problems (3%) — every other model tested scored 0%

Innodata's data engineers annotate, label and curate the training data behind frontier AI models, and demand for that work keeps climbing. Its second-quarter 2026 revenue hit a record $92.1 million, up 58% year-over-year

TrendForce sees a slow shift away from merchant GPUs inside AI servers as Google, Meta and other North American cloud providers scale up in-house chips. It expects GPUs to still lead 2026 AI server shipments at 69.7%, while pegging the ASIC-based share at 27.8% in 2026, the highest since 2023

Epoch AI's new AI Chip Users explorer estimates that OpenAI's training-and-inference computing capacity, in H100-equivalents, grew over 2024 and 2025 combined to 17 times OpenAI's 2023 computing capacity, per Epoch AI's H100-equivalent estimate