Heavy-frame gas turbines now take about 278 weeks to arrive and transformers about 162, Wood Mackenzie finds.

Anthropic is hiring people who can secure electricity, land, hardware and approvals for its power buildout.

Amanda Goodall tracks Anthropic's recent senior hires and finds many are not researchers but people who get power, land and government approval, including hires from Hitachi Energy and Google's data center supply chain. AI labs are now staffing like utilities and developers. That tells you where the bottleneck sits.

A local Ollama model wrote PLC structured text offline through a CODESYS MCP, but still failed to create the project.

Because machine builders will not paste PLC code into a cloud chatbot, Matt Kurantowicz ran a free model on one GPU with Ollama, no network and no API key. It wrote structured text locally and launched CODESYS through a local MCP client, though full project creation still failed. The offline path is what makes AI usable on controls IP.

ENFORCE 2.0.0 (45 stars): keep your PyTorch model and force its predictions to obey mass balances, co-developed with Shell.

ENFORCE is an MIT-licensed PyTorch framework from TU Delft that adds a projection layer so a neural network's predictions satisfy equality and inequality constraints, such as mass balances. Artur Schweidtmann shows a distillation model that scores well while inventing material, which ENFORCE prevents. Shell supported the research. For process models, hard constraints beat hoping the network learned conservation.

For process engineers and data scientists building surrogate models of plants.

Found via Artur Schweidtmann on LinkedIn