Sand to Sentence — the machine-intelligence stack

A self-guided descent through 27 strata

Every AI answer
begins as sand.

Between a quartz seam in North Carolina and a sentence on your screen sit roughly 27 layers of civilisation — crucibles and crystal pullers, tin-plasma light sources, copper damascene, stacked memory, cold plates, gas turbines, all-reduce collectives, preference data, attention heads, and finally a person deciding whether the answer was worth anything. This is the whole chain, one layer at a time, with the companies that own each link.

01 · lithosphere 27 · return & residue
How to read this

Each stratum is a layer of the stack; each station inside it is a distinct technology, process or market with its own physics and its own incumbents. The coloured pips on a card are a criticality read — how concentrated and how substitutable that link is. Three magenta pips means a genuine single point of failure for the whole industry.

Figures reflect public reporting through mid-2026 and move fast — treat capacities, valuations and roadmap dates as of that vintage, not as live data. Company lists are illustrative of who matters at each link, not exhaustive or ranked, and nothing here is investment advice. Built as a map for understanding, not a database of record.

The dependency web

Every station, stacked by depth. Hover a node to light its supply cone; click to open it. Drag to pan, scroll to zoom.

what it needs (upstream) what it feeds (downstream) zoom in to read names

The same column, re-weighted

Where the money actually sits.

You have already walked down this column once, sized by how many stations each layer holds. Here it is again, sized by capital. Recognition and rearrangement teach more than either does alone — and every figure is derived from something committed, or it is a dash.

How to read this

Bar length is the layer's total. Bar saturation is concentration — the Herfindahl index computed on attributed value, so a layer where one company holds most of it reads darker than one where twenty share it evenly. The hairline beneath each bar is coverage: how much of the layer's cast is in the ticker spine at all. A long bar over a short hairline is a lower bound wearing a confident face, and the panel says so out loud below 60%.

A company at nineteen stations must not be counted nineteen times. Every one is split across its stations by weights that sum to one. The default is an even split — no judgement, fully reproducible. Thirteen entries carry hand-set weights instead, because an even split would put five per cent of NVIDIA in the scheduling layer; each says why, and the toggle above shows you exactly how much that judgement moved the answer. None of them are segment disclosures.

Private companies have no market value here at all. They are counted in the cast and excluded from every total, because mixing a two-year-old private mark with a live public price produces a number that means nothing. Divisions carry an estimated share of their listed parent, flagged as an estimate. Attributing a parent's whole market cap to a division is wrong; attributing zero is also wrong.

Nothing here is investment advice. No recommendations, no price targets, no view on any security. Sources, formulas and known weaknesses are on Method.

Seventeen orders of magnitude, drawn to scale

Feel a nanometre.

Everything here is drawn at its true relative size. Scroll, drag or use ← → to travel from the silicon lattice to the Earth. Click anything to centre it.

across the view
How to read this

The horizontal axis is logarithmic: every 300 pixels is one power of ten. Objects are drawn at true relative size — if something looks a hundred times bigger than its neighbour, it is a hundred times bigger. That is why most of the ruler looks empty. Seventeen orders of magnitude do not fit on a screen, and pretending otherwise is the usual lie in diagrams like this.

Shapes are schematic; proportions are not. Each object carries a precision flag — exact for defined or measured quantities, typical for a representative production value that varies by vendor and node, approx for the right order of magnitude and not much more. Sources and reasoning are on the Method page.

Forty jurisdictions, drawn to scale

Every chokepoint is a place.

Fifty-six sites, each resolving to stations in the corpus. Drag to pan, scroll to zoom, click anything to open it. Circles are true to the ground — the ring around Spruce Pine really is thirty kilometres across.

across the view
How to read this

Rings are geodesic: a hundred points, each exactly the stated distance from the centre, projected like any other geometry. That is why they widen towards the poles — an equirectangular map stretches east–west by 1/cos(latitude), and a ring that stayed circular on screen would be lying about the ground. Each site carries a precision flag saying what kind of coordinate it is: sited where the operator publishes the location, approx where it is reconstructed from public reporting, area where the point is the centre of a district.

Coordinates are for published corporate and public locations only. Nothing here is more precise than the operator has itself made public, and nothing is derived from imagery, filings or any other private source. The export-control regime layer is a reading of who writes and who is bound by the leading-edge tool controls — it is judgement, labelled as such on Method, and not a legal finding. Boundaries are Natural Earth's, drawn for orientation.

Every capability in the stack, with two dates

Everything here is older than it looks.

The left end of each bar is the year a capability first worked. The right end is the year it arrived in volume. Drag the year, or sweep it, and watch the stack fill in. Click any bar to read what it was waiting for.

How to read this

Invented is the year the capability was first demonstrated in a form a specialist would recognise as working — a paper, a lab device, a proof of principle. Shipped is the year it reached volume, or general availability at a scale that changed what the layer above could assume. Both are judgement calls at the margin, so each bar carries a confidence: dated where a specialist would recognise both years, contested where one of them is a matter of where you draw the line.

While the scrubber is short of today, each bar is drawn twice: solid as far as the year you are standing in, and faint for the rest of the wait. The faint part is the answer — it is what the wait turned out to be, not what anyone knew at the time.

Four entries have no right-hand end. They work in a lab or in early production and have no volume date, and this view will not invent one — an arrow is not a forecast. A handful of bars run off the left edge: they were finished before 1947 and waited for an industry to exist. Their true year is in the panel.

The events are chosen, not enumerated. The medians in the headline are properties of this selection, not measurements of the industry — honest about a pattern, useless as a dataset. There is a survivorship problem built in, too: capabilities that never shipped at all mostly do not appear, because nobody writes their history. Said again, at more length, on Method.

Exposure mapping, not prediction

Cut one link. Follow it up.

Remove a station and the dependency graph says what sits downstream of it. That number is arithmetic and it is not damage — what actually reroutes, and what dead-ends, is a separate reading in a separate voice, and this page keeps them apart.

How to read this

Each row is a stratum and each cell a station. A white cell is the link the scenario removes. Everything faintly coloured depends on it somewhere up its chain — that is the graph talking, and it makes no claim about whether anything breaks. Amber and magenta cells are a hand-written reading laid on top: amber where a substitute exists and the entry says what it is and roughly how long qualifying it takes, magenta where the reading is that nothing substitutes on a timescale anyone would call a response. The count beside each stratum is how many of its stations the shock reaches.

The unclassified remainder is deliberately left visible. On most scenarios it is the large majority of the blast radius, and pretending otherwise would be the single easiest way to make this page dishonest. Lead times are years to a substitute carrying real volume — not first sample, not an announcement — and where a comparable substitution has already happened, the entry links to it on the Lag chart rather than asking you to take the number on trust.

Three things this cannot do. It has no per-edge coupling: every dependency is present or absent, when some are fatal and some are an inconvenience. It has no inventory: a shock absorbed by six months of stock looks identical to one absorbed by nothing. And its unit is the station, which is sometimes coarser than the real event — where that matters, the scenario says so in its own words. Nothing here is a forecast, and nothing here is investment advice.

Sand to sentence, actually quantified

What does one answer
really cost?

Follow 1,000 output tokens — roughly one substantial answer — backwards through the whole stack, converting units at every step, until it arrives at rock. Every factor below is sourced or derived in the open; press how on any step to see the working. Change the assumptions and the whole chain recomputes.

How to argue with this

Each step is a single multiplication, and each multiplier is either cited or derived from first principles in the panel behind its how button. The ranges are not error bars in the statistical sense — they are the span between defensible low and high parameter choices, propagated through the chain. Where they are wide, the honest answer is wide.

The largest single uncertainty by far is energy per output token, which varies by two orders of magnitude between a distilled model answering directly and a frontier model reasoning at length. Nothing else in the chain comes close. Figures last reviewed .

Sources, assumptions and limits

How this site
knows what it claims.

What the words mean

Where each kind of claim comes from

Grouped by how much you should trust it, weakest first. Judgement calls are listed before cited figures on purpose — they are the ones that shape the site most and defend themselves least.

Against the grain

Findings that contradict what most people assume, collected in one place. Each is load-bearing: it changes how the rest of the stack reads.

The assumption ledger

Every parameter behind the Cascade, generated directly from data/static/cascade.json — the same file the calculation reads. If a number changes there, it changes here, and the build fails if it loses its source or falls outside its own range.

Switchable assumptionOptionsWhat it means
ParameterCentralRangeDerivationSource

Where this is most likely wrong

Written by the person who built it, which makes it incomplete by construction. Additions welcome.

Corrections

Figures reflect public reporting through and move fast. Company lists are illustrative of who matters at each link, not exhaustive or ranked. Nothing here is investment advice. Built as a map for understanding, not a database of record. This page last reviewed .

CompanyWhat it does here StationBase