AI Meets the Physical World

Follow the conversation from AI models and useful workflows to the power, equipment, and cooling behind them. Explore the demos, research, and source material below.

Funding figures are dated announcements, not live valuations. Forecasts and vendor specifications retain their original scope; worked cooling examples are illustrative calculations.

Try the tools & demos

• Google Nano Banana 2 — Google’s introduction to its image-generation model, February 26, 2026.
• ChatJimmy — Try the fast-inference chatbot discussed in the episode.
• Transformer Explainer — An interactive walkthrough of how a transformer processes text.
• Taalas, Products — Vendor reports 17k tokens/s/user on Llama 3.1 8B. Does not establish equal model quality or energy/task.

Investment, electricity & the data-center buildout

• McKinsey, The $7 trillion data center build-out — March 27, 2026: roughly $7 trillion in global data-center spending through 2030. A cumulative forecast across categories, not annual cooling revenue or committed orders.
• IEA, Key Questions on Energy and AI, executive summary — IEA’s 2026 outlook: 485 TWh of global data-center electricity use in 2025, with a central projection of 950 TWh in 2030. Covers all data centers, not AI alone.
• Berkeley Lab, 2024 United States Data Center Energy Usage Report — 2023 US estimate and 2028 scenario range. Different geography and dates from IEA.

Company valuations & funding

• OpenAI: February 2026 funding announcement — $110 billion in new investment at a $730 billion pre-money valuation. Pre-money excludes the new investment.
• OpenAI: March 2026 funding close — $122 billion in committed capital at an $852 billion post-money valuation; a later announcement than the February figures.
• Anthropic: February 2026 Series G — $30 billion raised at a $380 billion post-money valuation.
• Anthropic: May 2026 Series H — $65 billion raised at a $965 billion post-money valuation; a later announcement than Series G. These are historical funding snapshots, not current market prices.

Models, benchmarks & AI workflows

• Vaswani et al., Attention Is All You Need — Transformer foundations. Dense self-attention has quadratic sequence-length complexity.
• Child et al., Generating Long Sequences with Sparse Transformers — Efficient attention changes scaling. Avoid universal context-cost laws.
• Anthropic, Building effective agents — Distinguishes predefined workflows from model-directed tool use.
• Model Context Protocol, Tool Annotations as Risk Vocabulary — Tool descriptions are hints, not enforced security controls.
• ARC Prize leaderboard — Results depend on the benchmark version and test configuration; consult the live source for scores.
• ARC Prize testing policy — How verified results are evaluated.
• Humanity’s Last Exam — The benchmark behind the HLE discussion; archived slide charts are not live rankings.
• Pew Research Center: AI concern and excitement by age — Source for the public-edition sentiment chart; preserve the survey date and age-group scope.

Chips, cooling & thermal engineering

• NVIDIA, Vera Rubin NVL72 product specifications — Vendor specification: 45°C inlet. Do not generalize to every GPU or deployment.
• NVIDIA, Hotter Than a Hot Tub: The 45°C Breakthrough — Vendor description of liquid cooling and climate-dependent dry-cooler operation.
• ASHRAE TC 9.9, Emergence and Expansion of Liquid Cooling in Mainstream Data Centers — Background on liquid cooling, thermal limits and facility interfaces. No figures reproduced.
• ASHRAE, AI Data Center Energy Performance Framework: Energy and Thermal Efficiency — Mixed cooling architectures and facility resource metrics. Guidance is not a project selection.
• NASA, Small Spacecraft Technology State of the Art: Thermal Control — Vacuum has no convective heat rejection. Radiators require area, temperature and suitable view factors.
• ASML: EUV lithography systems — Background on the lithography technology used in advanced chip manufacturing.