Situational Unawareness

Deal flow leads, public valuation lags. When capital and capacity pile onto a layer faster than the market reprices it, that layer is the next bottleneck: the next GPUs, the next HBM. The edge was never predicting that AI gets bigger. It is knowing which layer binds next, while the multiple still says otherwise.

We scraped public announcements of deals between companies and drew an edge for every one. The result is the AI economy as a graph, stacked across six layers from Capital at the floor to Application at the top, with the timeline replaying how the web thickened from 2020 to now. It’s meant to be wandered through.

Open the map →See the layer dashboard →

To find the next bottleneck we score each supply-side company on three things and multiply them: how fast deals are converging on it, how little the market has already repriced it, and whether the demand behind it is real rather than just a cheap stock. High only when all three line up.

How the score works ⓘ

The only honest test of a predictor is whether it would have called something beforeit happened. So we froze the model in January 2025, fed it nothing but deals and prices from on or before that date, and asked for the next bottleneck. It put SK Hynix third, at a moment the stock had gone nowhere for a year.

Then memory became the story, and it ran about twelve-fold. As it ran, its score decayed toward zero. The model goes quiet exactly when the trade becomes consensus, which is the point.

Run it on today’s data and the signal has moved off the chips. The highest scores cluster around the buildout’s physical limits, and the cleanest unrepriced weight is in power: the megawatts to run the data centers, not the GPUs inside them.

The shift we’re arguing for is easy to state and hard to practice: stop reading AI as one curve and start reading it as a stack where the binding constraint keeps moving. Last cycle it was GPUs, then memory, and each time the layer that mattered was invisible in the consensus until it wasn’t. This cycle the smart money still says compute. The deals say power.

That is the proof of concept. A map built only from public announcements, scored with three transparent inputs and no private data, would have put SK Hynix at the top of the memory layer in early 2025, while the stock was flat and the story was still GPUs. It then ran roughly twelvefold, and the score went quiet as the trade became consensus. We are not claiming a crystal ball. We’re claiming something smaller and more useful: the information was already public, and the only thing missing was a lens wide enough to hold the whole stack at once.

What building it taught us is that the bottleneck is physical and financial long before it is computational. The frontier needs memory it can’t buy, megawatts it can’t build fast enough, and capital that has to be structured before a chip ships. Map those dependencies and you watch the next constraint form. Watch only the frontier and you meet it at the price.

Situational awareness got the destination right. It’s the road that’s mispriced. Read the whole stack, or keep getting surprised by the layer you weren’t looking at.