How does Lowdown work?
Lowdown publishes intelligence briefings, built mostly from first-party sources: government press releases, changes to web pages, shipping movements and market data. It runs like a newsroom:
- An editor-in-chief agent plans each briefing and hands the draft to a writer agent.
- A fact-checking and legal desk needs three or four direct sources before a claim about a person goes in.
- An editorial desk handles style, and another desk removes the tells of AI writing.
- Ruby scripts score every draft for banned words, contradictions between stories and flow.
- The editor reads the scores, makes its own call, and usually sends a draft back three or four times.
A fixed pipeline, one phase handing to the next, did worse on my checks and needed more of my time. Story threads stop each briefing retelling the last one, so it reports what changed.
How many people get their news from AI depends on the measure. Ofcom found that 21% of UK adults used an AI service for news in the past month, and 13% used ChatGPT. The Reuters Institute asked about weekly use of AI chatbots for news in 2025 and found 3% in the UK.
What has Lowdown found?
Lowdown’s Trump page asks why he has posted daily on Iran for 99 days without signing a single Iran instrument. Between 12 and 21 September, the Federal Register and the White House listing showed no Iran document at all: Trump signed nothing on Iran this week.
In April, Lowdown reported that Britain’s queue for new electricity demand stood at 125 GW. In September it went back to NESO’s register, the list that figure should come from, and counted it: 2,202 entries, and not one of them is demand alone. NESO’s own reformed pipeline says 99 GW. Lowdown published the correction against itself in Data Centres, edition 14.
A headline said 45% of English universities were in deficit. That is the regulator’s modelled scenario, the one where nobody takes mitigating action. Its actual count for the same year was 36.6%, 102 of 279 providers, and that count went unreported for three and a half months. One number is a model, one a count.
FEMA gave the 11 US host cities of the World Cup $625m for security. It gave host states another $250m for counter-drone work. That makes $875m, a total FEMA never published and no outlet added up. World Cup, edition 44.
Who is Plotbeacon for?
The warehouses at Astley Business Park in the Wigan borough show the problem. Wigan Council approved them in June 2024, and a required screening application went in only a month before, without any public consultation.
Plotbeacon collects applications from council portals. There is no national feed, so it scrapes them. An AI model scores each one out of ten for how much neighbours will care and rewrites it in plain English. Each application goes on a map that crosses council boundaries, coloured from green for a new shop sign to red for a warehouse. On 24/09/2026 it covered 400 planning authorities, Barnet among them, and 22,385 applications.
The same data can help applicants: the flood zone, the heritage status and nearby precedents, before they submit, so the council receives an application it can approve. The public site does not offer that side yet.
Why host everything in Europe?
Everything I run is hosted by European providers, from the CDN to the servers. The trade-offs are some missing features and more to manage yourself. I wrote up the move in On the European stack.
The setup spans three providers: Hetzner in Germany, a replica at OVH in France, and a backup copy at Civo in the UK. The database streams its write-ahead log to object storage continuously, so I can restore it to any moment. When an agent deleted Plotbeacon’s production database, I restored it to a point ten minutes before the deletion. Keep production credentials away from agents, and test the restore before you need it.
Why use AI for family history?
Most family-history work is reading records about the wrong person. The AI holds every person and record in view at once, and it follows the same rule as Lowdown: a claim is only as good as the source under it. I tie each claim to its source, mark what is inference and keep the contradictions.
I am building this as Tin Chest, which records what every claim rests on.
In my own tree, one entry on a family-tree website held up a branch of 24 people and 217 documented facts, a New Zealand prime minister among them. That branch stays frozen until an 1833 baptism register settles the link.
Tin Chest’s family-history guides write up the method.
Doesn’t AI use too much energy?
It uses a lot, and measuring it is harder than the headlines suggest.
Most of the alarming numbers are connection queues. A queue counts capacity someone has asked the grid for, which is a long way from electricity used. Britain’s data-centre queue was about 50 GW across roughly 140 sites in February, against a national peak of 45 GW on 11/02/2026. After Lowdown’s September count, I treat any queue total as a pile of requests.
Finland’s grid operator, Fingrid, warned that 5 GW of connections already signed may never be built. Denmark has about 60 GW of requests against a peak near 7 GW.
Ireland has a real measurement. Data centres used 23% of its metered electricity in 2025, up from 5% in 2015.
In September, three governments acted on those queues. Texas halted every data-centre environmental permit on 21 September; its grid queue is 438 GW, about 90% of it data centres (Lowdown on Texas). In Scotland, a ministerial Direction holds data-centre planning decisions until national guidance arrives, and the Scottish Parliament backed a pause by 80 votes to 26 (Lowdown on Holyrood). Ireland’s planning appeals board refused a €1.5bn data centre in County Mayo that had no fixed grid connection and no renewable power contract (Lowdown on Mayo).
Coverage of OpenAI pausing Stargate UK put it down to regulation and to UK industrial electricity, at more than four times the price in the US and some Nordic countries.
Giant Robots gave this a whole episode: episode 616, with Dr Victoria Plutshack. Lowdown follows it in Data Centres: Boom and Backlash, including Fingrid’s warning and Ofgem’s proposed fee for a place in the queue.
Scroll sideways for the source links.
| Figure | What it means | Source link |
|---|---|---|
| About 50 GW across roughly 140 sites; 45 GW peak | Britain’s data-centre queue compared with peak demand on 11/02/2026. That day’s peak is a reference point, not an all-time record. | The Register, citing NESO and Ofgem |
| 41 GW to 125 GW; at least 80 GW from data centres | Britain’s wider demand-connection queue by Ofgem’s figures, November 2024 to June 2025. Lowdown could not reproduce 125 GW from NESO’s register in September. | Ofgem, dated comparison, Lowdown’s check |
| £237,500 to £712,500 per MW | Ofgem’s proposed commitment fee on accepting a connection offer. The proposal refunds it at energisation and forfeits it on early exit. | Ofgem’s consultation announcement |
| More than 4 times | Reported UK industrial electricity prices against the US and named Nordic countries, in coverage of OpenAI’s Stargate UK pause. Regulation also mattered. | The Next Web |
| Up to 35 data centres; £10 billion planned | Slough’s cluster faces grid constraints, while Blackstone’s proposed QTS campus near Blyth received outline planning permission. These are separate developments. | Slough reporting, Blyth planning report |
| Ireland: 5% to 23% | Data centres’ share of metered electricity rose between 2015 and 2025. This measures electricity consumed, unlike a connection queue. | Central Statistics Office |
| Denmark: about 60 GW against roughly 7 GW | Requested new demand compared with national peak demand. Energinet describes the pressure on grid expansion and warns that many projects may never materialise. | Energinet’s own explanation |
What about water?
Water is a local problem. A fleet total says little; one evaporatively cooled site can become the biggest customer of a small, stressed water system. Withdrawal is water taken from a source, and consumption is the part not returned, mostly because it evaporates.
The local numbers tend to stay private until someone forces them out. The Dalles, in Oregon, only published Google’s share of its water, 29% in 2021, after the city settled its lawsuit against journalists. Laws are starting to do the forcing. Virginia now makes utilities report data-centre water sales separately (Lowdown on Virginia), and Brazil has written a water cap into law: its new data-centre tax break limits claimants to 0.05 litres per kilowatt-hour (Lowdown on Brazil).
Cooling design changes the demand. Google says its Mesa site in Arizona uses air cooling, and a 300 MW operator in Johor says it runs on treated wastewater (Lowdown on Johor).
Scroll sideways for the source links.
| Figure or record | What it means locally | Source link |
|---|---|---|
| London and Slough: drought investigation in 2022 | Thames Water examined data-centre demand and alternatives to drinking water. The report establishes the investigation, without quantifying a resulting shortage. | Thames Water reporting |
| The Dalles: 29% in 2021 | Google’s data centres accounted for this share of the city’s water consumption. Records became public after the city settled its lawsuit against journalists. | Released water records, reported by DCD |
| Google: about 23 billion litres in 2023 | Reported consumption across Google’s data centres. This fleet total cannot identify the pressure on an individual water supply. | Google’s environmental report |
| Newton County: 1 million US gallons daily in 2030 | Appendix A projects this combined demand for Meta’s 2 developments. The document also models lower demand; these are forecasts, not measured use. | County water analysis |
| Mesa: air cooling | Google says its Arizona facility uses air cooling. Local design changes the demand, even where the surrounding region faces water stress. | Google’s Mesa account |
Neighbours of Meta’s planned sites in Newton County, Georgia, have reported well problems. The county’s study has no baseline for them, so the cause is still disputed in the reporting on the complaints.
Is AI taking the jobs?
Giant Robots called it in episode 606, “AI layoffs are BS”, and Lowdown’s evidence mostly agrees. Companies cut staff, then argue about the label. Microsoft cut 4,800 and denied AI did it. Goldman cut 2% on record revenue and refused the AI excuse. Wells Fargo cut too, and named AI as the reason.
The data outside company statements points the same way. The New York Fed found adoption rising while sackings stayed rare. In a survey of 1,000 US hiring managers, 59% said they emphasised AI when explaining freezes or layoffs, but only 9% said AI had fully replaced roles. New York’s layoff notices covered 162 employers and nearly 28,300 workers with no AI attributions, WIRED reported on 09/02/2026. A Federal Reserve comparison of late-2025 surveys put adoption at about 18% of firms and 41% of workers.
More in AI: Jobs, Power & Money.
What Bannermedia makes.
- Lowdown: live. Intelligence briefings from first-party sources, written by an AI newsroom I run.
- Plotbeacon: beta 2026. Scores and explains council planning applications on a map, for neighbours and for the trade.
- Tin Chest: in development. An evidential family-history archive that records what every claim rests on and what contradicts it.
- Resurface: in development. Finds the moment you remember in years of your own footage, searching a video archive on your own machine.
- Accountably: in development. A focus tool for working alongside friends without a chat feed demanding attention.
- Lettergraph: in development. A newsletter publishing platform for writers who want to keep control of their audience.
I also work as a fractional CTO, helping teams make technical decisions and review the software those decisions produce.
Get in touch.
Email hello@bannermedia.ltd about a product, a technical problem or a source that changes a claim on this page.
Find me on LinkedIn for professional contact.