In partnership with
In a 1.2 GW data centre commissioning hall in Abu Dhabi, a senior engineer signs the permit to operate. The time is 09:47 on a Tuesday. The paper is on the desk. The seal is in her pocket. The asset register is open on her laptop. She has spent eleven weeks reading the field against the design. Three subcontractors are watching her sign.
When she puts the pen down, the facility is permitted. Not before. Not by anyone else.
The AI did everything else.
A fortnight ago, Edition 8 named the symptom. The handover specialist does not scale. The cause is older. It gets harder as the model gets better.
No AI signs the permit.
Everything except the signature
In 2026, AI compresses the work that used to take a commissioning engineer eighteen months. Extraction engines ingest 30,000 documents in hours. Standards-harmonisation tools reconcile 25 conflicting specifications without human intervention. Validation models flag missing welds, missing as-builts, and missing baseline performance tests faster than any junior auditor.
The signature is a different category of object. It is not a piece of work that can be compressed. It is the moment a named, competent person accepts personal accountability that the facility is fit to operate.
That accountability is not delegable. It is not transferable. It is not, in any jurisdiction worth permitting in, performable by software.
Every regulator is sovereign
The competency frameworks are sovereign. The signatures do not transfer.
The world has around 32 countries with operating civil nuclear power. Each has its own regulator with its own competency framework.
The United Kingdom has the Office for Nuclear Regulation. The United States, the Nuclear Regulatory Commission. France has the Autorité de Sûreté Nucléaire. The United Arab Emirates has the Federal Authority for Nuclear Regulation. Saudi Arabia has the Nuclear and Radiological Regulatory Commission. South Korea has the Nuclear Safety and Security Commission.
A senior reactor operator licensed in the United States cannot sign a permit to operate in France. A French commissioning engineer cannot sign in Abu Dhabi.
Data centres run a parallel pattern with different acronyms. UL in North America. TÜV in Germany. DNV across Northern Europe. The Uptime Institute certifies facilities at the tier level. The underlying competent-person sign-off lives with named individuals under local frameworks.
Every sovereign AI infrastructure programme sits on top of a domestic regulator's certificate. Stargate UAE. Humain. The French AI gigaprogramme. The UK AI Growth Zones. The Indian gigafactory build. The certificate, in every case, sits in a human hand.
The bar rises as the model improves
AI accelerates the work in handover. AI does not reduce the accountability that someone has to take for the result. As the model gets better, the things a competent person can responsibly sign for go up. The competency required to assess what the model got right also goes up.
A senior commissioning engineer in 2020 read maybe 2,000 documents end to end before signing. The same engineer in 2026 reviews the same 2,000. The difference is that AI has now pre-classified, summarised, flagged, and structured the exceptions.
The exception report is faster to read. The accountability for its accuracy is hers.
If she signs and the model missed a fire-suppression interlock, the regulator does not visit the model vendor's offices. It visits hers.
This is what expert in the loop actually means in regulated infrastructure. AI augments expert judgement. Operators, regulators, and engineers decide. As the augmentation strengthens, the decision gets heavier.
The signature does not get easier. It gets heavier.
The pipeline you cannot accelerate
The Office for Nuclear Regulation publishes its inspector pathway publicly. The route from baseline qualification to lead inspector takes five to ten years depending on discipline. The US NRC's Senior Reactor Operator path runs similarly long, with cycles documented under 10 CFR Part 55.
Data centre commissioning has no global framework of equivalent rigour. Competence is built one project at a time. It is signed off one project at a time. It is aged out one retirement at a time.
The supply curve has a fixed lead time. Demand does not.
In 2026, the demand side is announcing a 1 GW campus a month. Stargate Abilene scales past 2 GW. NVIDIA and OpenAI commit to 10 GW. Lenovo announces a Gigawatt AI Factories Programme at GTC 2026. The supply side trains specialists at the rate of a single national framework. The cohorts run in tens per discipline per year, across years of supervised work.
You cannot hire your way to a senior reactor operator in eighteen months. You cannot hire your way to a competent data centre commissioning lead in six. The pipeline is the pipeline.
Who signs the permit?
The commissioning or handover specialist does not scale because the signature does not scale.
Edition 1 set the cost-of-delay floor at £79m ($100m, €95m) per day for one 1 GW AI factory sitting dark. That floor is now stretched across every campus committed since. The number of facilities chasing finite signature supply has grown by every announcement. The number of competent signatures has not.
Edition 7 named the grid as the structural constraint. Edition 8 named coordination as the bounded function. Edition 9 names the irreducible one.
The story of Permit to Operate is not the model that reads. It is the person who signs.
What comes next
Edition 10: The standards committee meets next year. The slowest clock in the system.
Edition 11: What the gap looks like five years out. Who closes it. Who lives in it.
Edition 12: What an AI co-signer would actually need to be. The end of the inversion, traced.
If you build, operate, invest in, or regulate infrastructure anywhere in the world, this is written for you. Subscribe to Still Dark.
This newsletter lives in the gap between digital delivery complete and permit to operate. That gap is where value dies, and where it can be recovered.
I also co-author The Vistergy Brief at vistergy.com/archive. Satellite and geospatial monitoring, facility lifecycle intelligence, and standards architecture across LNG, nuclear, data centres, utilities and construction. Subscribe to both for the full picture.
The podcast is live. Permit to Operate is available at vistergy.com/pod, Apple Podcasts, and Spotify.
Permit to Operate.
Crash Expert: “This Looks Like 1929” → 71,105 Diversifying Here
Mark Spitznagel, who made $1B in a single day during the 2015 flash crash, warned markets are mimicking 1929. Seems extreme but we did just see the worst quarter for the S&P since 2022.
So it’s not so surprising that Vanguard and Goldman Sachs forecasted 5% and 3% annual S&P returns respectively for 2024-2034.
Late last year, Apollo’s chief economist Torsten Slok put it this way: "expect zero in return in the S&P 500 over the coming decade."
Almost no one knows this, but postwar and contemporary art appreciated 10.2% annually with near-zero correlation to equities from 1995–2025 overall.*
And sure… billionaires like Bezos can make headlines at auction, but what about the rest of us?
Masterworks makes it possible to invest in legendary artworks by Banksy, Basquiat, Picasso, and more – without spending millions.
29 exits. Net annualized returns like 16.5%, 17.6%, and 17.8% on works held over 1 year+. $1.3 billion invested. 500+ offerings.*
Shares in new offerings can sell quickly but…
*According to Masterworks data. Past performance is not indicative of future returns. Investing involves risk. Important Reg A disclosures: masterworks.com/cd.