CrudeCode is an Apache-2.0 MCP server that lets Claude or ChatGPT query state well filings with guarded SQL and run decline forecasts, valuation and maps. The repo is small, with one maintainer. It is a working upstream tool, not a pitch.
For upstream engineers: query well data and run decline and NPV economics in plain language.
OpenAI's Grid Protectors program asks its future campuses to reduce load when the US grid is under strain. It is non-binding, and the model is OpenAI's existing 25-year Georgia Power deal for the 3.2 GW Effingham County campus, where the utility can curtail up to 1,000 MW at peak and OpenAI pays full grid costs. The open question is how often utilities actually get to call it.
After FERC rejected most of its first plan, PJM refiled its reliability backstop with a March 1 start. Jefferies says utilities cut near-term large-load forecasts by 6.5 GW, so the purchase could fall to about 630 MW, while a CRA study still sees an 11.5 GW shortfall by 2030. The same grid looks short or nearly fine depending on whose forecast you use.
At Aurora Energy Research's forum, 2030 data center load forecasts ranged from about 22 GW to 131 GW. Grid operators expect about 63 GW more demand than Aurora's base case, and getting it wrong could mean $79B to $624B in extra capex. Ratepayers pay for the miss either way.