Analysis

Our thesis, in depth

Long-form, sourced analysis on where the data for the next generation of AI actually lives — each edition with its film.

x algorithmopen sourcealgorithmic transparencyAugust 2026

X published its algorithm. We read it, we got one thing wrong, and the source caught us.

A repost is worth two likes, not twenty. A profile click is worth nothing. The heaviest positive signal in X's published code is someone copying your link. And one filter names 665 accounts outright, by handle — which is the claim we got wrong, in public, before our own source corrected us. What the code says, what it cannot say, and why the correction matters more than the findings.

Read & watch

freightmaritime datasupply chainAugust 2026

Everybody says the supply chain needs data. Nobody is buying it. They're buying the companies.

The open, neutral supply-chain data platform died in public: Maersk and IBM switched off TradeLens saying the industry never showed up. The same data, owned by an integrated company, sold for $241m — and the seller kept the US government slice. Here is who actually pays, why the only regulator to look at the consolidation closed the file without examining it, and why this is one cluster where our own model does not fit.

Read & watch

energygrid datadata centersJuly 2026

The grid publishes everything about itself. It just found the one thing it can't see.

Energy is 29.7% of our signals and our largest cluster — yet grid data is free by law (ENTSO-E, the EIA, 'presumed open'). The scarce data is the behaviour of the huge new loads the grid can't see; NERC just issued its most serious alert asking for exactly that. The grid became the buyer — and no market for that data exists yet.

Read & watch

mobilitytelematicsphysical aiJuly 2026

A turbine only ever talks about itself. A truck brings back the world.

Mobility is 30.8% of our data holders but 34.6% of our signals, and the only genuinely multimodal niche — a world, not one machine. It can't be scraped, the EU just made it a priced licensing channel, and the real buyers are insurers, maintenance-AI and RegTech — not the AV labs.

Read & watch

world modelsphysical aidata licensingJuly 2026

Web scraping won't train the next AI

Public text is finite, self-polluting and locking shut. World models and physical AI need real, operational, multimodal data that was never published — and it belongs to operators, not labs.

Read & watch

predictive maintenancetime seriesindustrial aiJuly 2026

The most valuable dataset in AI is the one nobody is keeping

Two months of signals point at one dataset: predictive-maintenance time-series from the physical economy. 43% of every holder we mapped, 100% time-series, on a $1.4-trillion-a-year problem that can't be scraped.

Read & watch