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

In the first edition, I argued that the next generation of AI — world models, physical AI — will be won on data that was never put online. That was the thesis.

This edition is the proof, and it's narrower than I expected. Because when you stop theorizing and actually listen to the market — thousands of dated, independent signals from the real economy — it doesn't point at a vague "industrial data" opportunity. It points, over and over, at one dataset.

Here's what two months of listening taught us.

1. What the machine heard: 5,000 signals, one recurring shape

Our platform runs on a simple, deliberately un-hyped principle we call signal-first: we don't start from an opinion about where the data is. We start from facts — a funding round, an M&A deal, a data partnership, a tender, an IoT rollout, a regulatory filing — each one dated, each one sourced from the trade press, each one pointing at a theme that produces data.

Over the last eight weeks, the engine harvested ~5,000 such signals. It then did something important: it refused to trust any single one. A theme only becomes a real "niche" when at least three independent sources converge on it — a guardrail against the hype of one loud headline. On that bar, 195 niches cleared confirmation.

And here's the striking part. Those ~5,000 signals generated thousands of candidate themes — but they collapse into a handful of families, and every one of them is the physical economy in operation:

• Energy, grid, renewables & storage — ~1,000 clusters, ~2,500 signals
• Supply chain, logistics & freight — ~620 clusters, ~2,200 signals
• Robotics & industrial automation — ~455 clusters, ~1,500 signals
• Mobility, EV, automotive & fleets — ~530 clusters, ~1,350 signals
• Mining, critical minerals & materials — ~310 clusters, ~1,000 signals

Not "AI." Not "LLMs." Not the web. The market keeps talking about machines that run in the real world — and, implicitly, about the data those machines produce every second and nobody keeps.

2. Ask "what data, and for whom?" — and it collapses to one niche

Signals tell you where the heat is. The real question is: what dataset does this actually create, and who would pay for it?

We've now mapped 413 real data holders (up from 311 in the last edition — the map is compounding). When we classify what each one is actually sitting on, the distribution isn't flat. One niche towers over the rest:

• 43% of every holder we've mapped resolves to the same use case: Predictive Maintenance — 179 of 413. The runners-up (industrial monitoring, document intelligence, regulatory RAG, diagnostic AI) trail far behind.
• 100% of those 179 are time-series. Every single one. Machine logs, sensor telemetry, fleet telematics — the raw signal of the physical world, timestamped.
• It cuts clean across sectors: industrial (89), mobility/fleets (58), energy & other (23), healthcare (7) — the same families the signals were shouting about.
• And the demand side has one dominant buyer profile: industrial-AI and maintenance-optimization vendors.

That is the most-repeated, most independently-confirmed pattern in everything we've collected. The market isn't asking for "data." It's asking for predictive-maintenance time-series — and it's asking loudly.

3. Why this niche, and not another: the cost of the problem is staggering

Predictive maintenance isn't the most prolific niche by accident. It sits on top of one of the largest unsolved costs in the global economy.

According to Siemens' True Cost of Downtime 2024, the world's 500 biggest companies lose an estimated $1.4 trillion a year to unplanned downtime — equal to 11% of their revenue, roughly the GDP of Spain. At the sharp end, an hour of unplanned downtime in an automotive plant now costs up to $2.3 million — more than $600 every second — and that per-hour cost has roughly doubled since 2019.

That is the exact pain predictive maintenance exists to kill. Which is why the market for it is real, not a promise: independent analysts size the global predictive-maintenance market at ~$13–14 billion in 2025 (Grand View ~$14.2B; Fortune Business Insights ~$13.65B; Mordor ~$14.1B; MarketsandMarkets ~$12.1B). Growth estimates vary widely — from ~11% to ~34% CAGR depending on the firm — so treat the headline growth rate as a range, not a settled number. But the direction is unanimous: up and to the right, for a decade.

A giant, quantified, recurring cost — and the data that would fix it is exactly the data our holders under-exploit.

4. Why it can't be scraped: this data lives behind the firewall

Here's what makes this niche structurally different from the web-text market the labs already exhausted.

Predictive-maintenance time-series is not on the internet, and it never will be. It lives inside SCADA systems, plant historians, CMMS databases and PLCs — behind the operational firewall, in formats built for control, not for publishing. There is no crawler that reaches it.

And most of it is never even used. The most-cited number here — Seagate and IDC's Rethink Data — found that only 32% of the data available to enterprises is ever put to work; the other 68% goes unleveraged (2020). Forrester put the share of enterprise data unused for analytics at 60–73% (2016). Those figures are older, and they're about enterprise data broadly — but they describe exactly the reservoir sitting idle inside every factory, grid and fleet.

The scale is visible wherever a company has managed to aggregate even a slice of it. In its SEC filings, the connected-operations company Samsara reports over 25 trillion data points flowing through its platform each year — and states plainly that "the cost and availability of sensors, compute… have prevented widespread analysis of physical operations data." Twenty-five trillion points, and that's one private platform. The rest is dark.

That's the definition of a scarce asset: enormous, valuable, and locked away from everyone who'd want to train on it.

5. The buyers are already arriving — and they're a brand-new class of model

For years, "who would buy operational time-series?" was a fair question. In 2024, it stopped being one — because a whole new category of AI showed up needing precisely this.

The last two years produced the first wave of time-series foundation models: Google's TimesFM, Amazon's Chronos, Salesforce's Moirai, IBM's Granite / Tiny Time Mixers, CMU's MOMENT, Nixtla's TimeGPT. The same "pretrain one big model, generalize everywhere" recipe that transformed text and images — now aimed at signals over time.

But they hit the same wall we've been describing. Google's TimesFM was pretrained on ~100 billion time-points drawn largely from Google Trends and Wikipedia pageviews — public web data, not a single hour of SCADA or sensor telemetry. And the researchers building these models say the quiet part out loud:

"Foundation models have transformed vision and language by pretraining on large, structurally coherent corpora — yet no analogous substrate exists for industrial time-series." — FactoryNet, 2026

"Unlike language, which has abundant public pre-training data in terabytes, time-series data is relatively scarce, very diverse and publicly limited." — IBM Research, Tiny Time Mixers (NeurIPS 2024)

The capital is following the same logic. Machine-health specialist Augury raised $75M at a $1B+ valuation (Feb 2025); industrial-maintenance platform Tractian raised $120M (Dec 2024). The buyers of predictive-maintenance data aren't hypothetical anymore. They're funded, they're building, and they're starving for exactly the reserve our 413 holders are sitting on.

The bottom line

Edition #1 said the next AI needs data that was never online. Edition #2 tells you which data, because the market wouldn't stop pointing at it: predictive-maintenance time-series from the physical economy — machines, grids, fleets, mines.

It's the most-repeated signal we collect. It's the single largest niche we've mapped (43% of every holder). It's 100% time-series. It sits on a $1.4-trillion-a-year problem. It can't be scraped. And a brand-new class of foundation model has just arrived unable to train without it.

On one side: operators sitting on a scarce, un-exploited asset — often without realizing it's the fuel of the next decade of AI. On the other: a new generation of models, and the vendors around them, ready to pay for it. What's missing between them is the infrastructure to match, qualify and trust the exchange.

That's what we're building.

If you operate machines, a grid, a fleet or a plant — you are almost certainly sitting on predictive-maintenance time-series that someone is now willing to pay for. If you're building AI that needs real operational signal — this is where it lives. Either way, now's the time to talk.

— Salim Labriki, d-nvest

Methodology note: market-sizing figures are proprietary vendor estimates, attributed by firm; where analysts disagree (notably on CAGR), we show the range rather than pick a number. A "signal" is a dated external fact (press/registry); a niche is counted as qualified only when ≥3 independent sources converge on it.