In partnership with
An energy owner answers to its regulator for every decision and every data flow. An AI answer it cannot trace, or that leaves its boundary, is one it cannot act on. On 2 November the Abu Dhabi National Oil Company (ADNOC) showed what that means in a contract. With G42, it signed a deal with OpenAI to deploy AI across its operations, Reuters reported.
The models will run on G42's sovereign AI infrastructure in the United Arab Emirates (UAE). ADNOC set its own condition. "All solutions will operate within ADNOC's oversight, ensuring that proprietary data and critical infrastructure remain protected," the company said. Human experts will keep oversight of consequential decisions.
Regulated owners want AI to run inside a boundary they control. We set out why, and what an owner should ask of the AI itself. Then we show one example running on Red Hat OpenShift.
Rules we read in the EU, Britain and South Korea hold the owner to account and set no location
In the European Union (EU), the AI Act treats some energy AI as high-risk. That covers safety components in the supply of water, gas, heating or electricity. An amending regulation of 8 July 2026 moved those duties to 2 December 2027.
The EU's directive on network and information security, NIS2, already applies. It names energy first among its sectors of high criticality. It requires supply chain security, including each entity's relationships with its direct suppliers and service providers. An AI service falls within that duty.
Britain's energy regulator is more direct. Ofgem's May 2026 guidance states that an organisation cannot outsource responsibility or accountability for safety or security. In South Korea, the AI Framework Act counts AI used in energy supply as high-impact AI. Operators of such AI must ensure human management and oversight.
France's EDF keeps ownership of its AI data, and Bull and Schwarz Digits plan European AI computing
Owners read the accountability rules the same way. In May, EDF set out its AI partnership with Mistral AI. The data will remain the property of EDF, it said.
The data will sit on a sovereign cloud or in EDF's own data centres. The tools will not touch nuclear plant control systems.
Suppliers are building for it. On 3 November France's Bull and Germany's Schwarz Digits signed a preliminary agreement. They plan AI computing services operating entirely under European jurisdiction, Reuters reported.
In September we argued that sovereign AI is an energy policy. This piece covers its deployment side: where the model runs, and who can see what it reads.
Three nuclear regulators describe AI that supports the operator's own decision
Location is half the requirement. The other half is how the AI behaves once it runs inside the boundary.
Canada's, Britain's and the United States' nuclear regulators set out shared considerations in 2024. Many AI capabilities, their joint paper says, may augment human decision-making rather than replace it. They write that integration should maintain appropriate human oversight.
The International Atomic Energy Agency (IAEA) met on the same question in Daejeon, South Korea, in November 2025. Its information sheet for the meeting warned that AI requires rigorous and comprehensive safety evaluations in nuclear plants. The United States' National Institute of Standards and Technology (NIST) states a general principle. Trustworthy AI depends upon accountability, and accountability presupposes transparency.
EPRI's SAFERai framework leaves consequential decisions with domain experts
For an owner, those considerations describe an agent it can put in front of an auditor. The agent reads the owner's own record of the asset, and cites the standard it checked that record against. When the record is silent, it declines rather than guesses.
The owner's engineer makes the decision. The Electric Power Research Institute (EPRI) writes the same into its SAFERai framework: domain experts retain authority for consequential decisions.
Where would your organisation let an AI agent read its asset records?
Our governed agent is certified on the Red Hat Ecosystem Catalog and demonstrated on a lab cluster
We built our own agent to behave this way. It runs on Red Hat OpenShift, the platform IBM's Maximo asset management suite also runs on. It is listed on the Red Hat Ecosystem Catalog, certified since 13 October 2026.
Its Helm chart, the package an OpenShift team installs from, was certified on 15 October. The current release was certified on 9 November. Red Hat's platforms, the company said in May, provide the software backbone for air-gapped, sovereign and private clouds.
We demonstrated the agent on Red Hat's OpenShift Partner Lab. It ran against GIGAWATT-1, the reference AI-factory campus we built to test it. Its 17 assets were held in a live Maximo record.
Asked about a work order, the agent read the record and added an audit line. Its readiness check passed a transformer's record. It failed a switchgear record and listed what was missing against the named standard.
A question about a figure the record does not hold drew a refusal, with the reason. The agent and the asset record sit on the same cluster, and the agent reads the record there.
The agent's model is set in its configuration. The same certified image runs unchanged on the lab cluster and on a second cloud platform.
The EU's 2027 date gives owners a year to choose where their AI runs
In Europe the high-risk duties apply from 2 December 2027. NIS2's supply chain duty applies now. An owner weighing an AI proposal can start with where the model runs and what record it reads. It can then ask which standard the answer cites, and who signs the decision.
The bottom line
Regulated owners answer for every decision their AI touches. ADNOC and EDF have each set the boundary their AI will run inside. Three nuclear regulators describe AI that keeps human oversight. An owner that keeps its AI beside its own records has an easier answer to give its regulator.
Next week: Small modular reactor (SMR) size classes against AI factory classes, and which reactor fits which campus.
The voice-to-text app that built its own speech model
Most dictation apps are a wrapper around someone else's speech model. Wispr Flow trained its own.
It's called Canto, and it was built for the way people actually talk to AI: thinking out loud, changing direction mid-sentence, half-whispering in an open office.
Wispr tested it on 10 hours of real dictations from more than 2,300 speakers. Canto had the lowest word error rate of every model tested, including models from Google, OpenAI, AssemblyAI and Deepgram. That means your prompts to ChatGPT, Claude and Cursor come through the way you said them.
Talk naturally in any app and Wispr Flow types clean, formatted text, 4x faster than typing.
