DEADRISK

Internal memo — do not forward

FREIGHT DESK / ALT-2029 / RE: THE EIGHT THINGS

Somebody upstairs asked why we only move eight lines when the catalogue runs to four hundred. Here is the answer, and I would like it to stop being asked.

A datacentre is not a building. It is a queue. Four queues, actually, and you join all four or you join none: the trays that do the arithmetic, the stacks that feed them, the iron that makes the electricity, and the plant that gets the electricity into the room. Miss one and the other three are inventory.

Everything else on the catalogue can be bought on a Tuesday. These eight cannot, and that is the whole business. Three of them come out of one country. Two of them need a licence to cross a line on a map. One of them has an order book that ran out before the decade did.

The part nobody upstairs believes: the compute order and the power order are the same order, placed six weeks apart by two departments who do not have each other's numbers. When the trays go short, the grid desk finds out last and pays most. We have made a living out of that gap for two years and I would rather we kept it.

Buy the thing that is cheap where it is made. Sell it where the queue is. Do not fall in love with a container.

— V., forwarding desk

Every run this machine remembers

Kept on this machine and nowhere else — there is no server, no account and no leaderboard. Clear the browser and the log goes with it.

Terms of use

1. This is a game. Every run is stamped with a year that has not happened and a branch that does not exist. The console says so in hazard tape for the whole month, and it is not decoration.

2. Nothing here is a security. The eight tradeable lines are equipment and commodities. No ticker, no equity, no option, nothing with a real quote anywhere in the world.

3. No company appears in this game. Events name a role — a Dutch toolmaker, a hyperscaler, a west-coast foundry — because the places are real and the actors never are. Nobody is quoted, credited or blamed.

4. Every price is invented. They are drawn from a seeded random number generator against ranges chosen by hand. Nothing on this page has ever read a market feed and nothing on this page ever will — a live feed would break the determinism the whole thing rests on, and it would make this a different kind of page with a different set of duties.

5. It is not advice, and it could not be even if it wanted to be. The real record — the part with receipts and sources — is the rest of the site.

These five are checked by a script before every deploy. They are load-bearing, not boilerplate.

Buildout economics — working file

FREIGHT DESK / ALT-2029 / NOT FOR CIRCULATION

Somebody keeps asking whether the pad is worth taking. Here is the arithmetic, at base prices, which nobody ever actually pays.

Four walls, and where you meet them

Everything in this game is invented — the prices, the branches, the year. The shapes are not. Each of the four below is a real constraint on the real buildout, and each one is a mechanic you have already been playing. The link at the end of each goes to the part of this site that keeps the receipts.

1 · The memory wall

The bottleneck nobody priced. Everyone modelled the arithmetic and almost nobody modelled the memory feeding it — and a tray of accelerators without stacks is a paperweight, so the scarce thing turned out to sit one layer below the thing everyone was watching.

In the game: HBM has the highest weight of any edge into accelerators, and it moves the day after. When you see stacks go, you have one day.

The record: the memory wall — a third of this year's big-tech capex now goes to memory alone.

2 · The power wall

The limit is not money and it is not silicon. It is permission to plug in. Interconnect queues run years. Turbine order books ran out past the end of the decade. A campus with every chip it needs and no transformer is a field with a fence round it.

In the game: this is the heaviest pair of edges on the board — accelerators to turbines and accelerators to grid plant, two days later, both ways. A rack that gets built is a power order that gets placed, and the grid desk finds out last.

The record: the power wall — grid queues, moratoriums, and who is on the scarce side.

3 · The circle

The most important shape here and the hardest to see, because it does not look like a risk — it looks like demand. A supplier takes a stake in his customer. The customer borrows against a contract with a third party to buy the supplier's product. Every leg books a sale. Very little cash ever leaves the group.

In the game: the VENDOR FINANCE offer. He books the sale, you book the debt, and nothing moves. It is usually a bad trade and occasionally the right one — but the loan does not care which, and it compounds at ten percent a day either way.

The record: the capex supercycle — not a spending number, a financing shape.

4 · The export line

A line on a map that some goods may not cross. It does not stop a buildout; it forks one. Two stacks, two price regimes, and a very profitable gap between them for anybody willing to carry the paperwork risk.

In the game: the two goods marked LIC, the customs percentage on every leg, and gallium at 0.28 in Baotou against 1.32 in Santa Clara — the widest single gap on the board, and it exists because of a licence rather than a shipping cost.

The record: the forked stack — controls did not stop it, they split it.

The method, in one line

A number means nothing without a baseline. That is the whole of it, and it is why this game's board stopped telling you a price is “+42%” against some global average and started telling you whether it is CHEAP, FAIR or DEAR against what that good costs in the market you are standing in. Same discipline, smaller stakes.

How the rest of this site does it: methodology.

What you are trading

Buy low in Shenzhen, run it past customs, sell high in Santa Clara. The debt compounds at ten percent a day whether you trade or not.

The original was a school project in 1984, copied onto more floppy disks than anyone counted. Thirty days, six goods, six boroughs and a loan shark. The whole game was one question asked sixty times: is this price good, or is the next one better? You never saw another market's board. You held the ranges in your head or you went broke.

Same game, different contraband. Every good below is a real chokepoint — a thing that, when it goes short, stops a datacentre being built no matter how much money is pointed at it. Nobody wins the buildout by being clever. They win it by having the crate.

The prices are invented. The reasons they move are not.

For scale: the 1984 original's dearest line topped out around $29,000 a unit. One gas turbine on this board is a hundred of those, and it eats forty of your hundred slots to carry.

Where you trade it

The original had no geography. Every borough drew from the same distribution, so location told you nothing and the game was pure nerve. For a supply chain that would be a lie: gallium is cheap in Baotou because that is where it is refined, and a transformer is dear in west Texas because that is where the queue is.

So each market tilts the roll. The tilt is mild on purpose — the numbers below run from 0.28 to 1.78 against price ranges five to eight times wide, so a bad roll in a cheap market still beats a good roll in a dear one about as often as not. Geography loads the dice. It does not throw them.

A blank means the market does not trade that good at all. Heat is the chance customs opens an audit when you arrive carrying something that moves under licence, scaled by how much of your load it is.

The chain

A shock does not stay where it lands. That is the one fact about this industry worth building a game around, and until recently this game did not have it: every price event was an island, so reading the wire could not pay and the whole thing was a slot machine with good copy on it.

Now the shocks travel, and they travel the way the real ones do — down the dependency chain, with a lag, losing force at every hop. Squeeze the tools and memory feels it two days later. Squeeze memory and the accelerator that is nothing without it feels it the day after. Squeeze accelerators and, two days on, so does power — because a rack that gets built is a power order that gets placed, and the grid is where this buildout actually queues.

Three things make it playable rather than decorative. The knock-on is global: a licence freeze in Baotou is a fact about the world, so it lands on every board that quotes the good. You cannot reach today's spike in a market you are not standing in — but you can be standing in the right one when its consequence arrives. The wire reports all eight desks, so you see the cause even when you were somewhere else. And about a fifth of the wire is noise: a line that reads exactly like the others and moves nothing, in the market it names or anywhere downstream. Which is the difference between knowing the chain and trusting the news.

Read off real arcs — the gallium export licensing, the memory sell-outs, the transformer and turbine order books running out past 2030 while compute demand kept being revised up. What is borrowed is the shape: who waits on whom, in what order, and roughly how hard. No real price is used, nothing here reads a feed, and every number on this page is invented.

The rules

• 30 days. Every move burns one. You cannot stay put — the only way to pass a day is to go somewhere.
• You start with $50,000 and owe $137,500 at 10% a day, compounding. Do the arithmetic before you decide it can wait: untouched, that debt is about $2,181,175 on the last day.
• 100 slots. A tray of memory takes one. A gas turbine takes forty. Freight charters turn up on the wire and are the only way to get more.
• One credit desk, in Santa Clara, and nowhere else. Repay there, draw on your line, or park cash in escrow at 5% a day where an audit cannot reach it.
• The line is a $250,000 floor plus 60% of what your book is worth over the loan. The book counts cash, escrow, the pad and the cargo — cargo at the lower of what you paid and what it fetches where you stand. So credit is widest when the thing you are holding is dear, and a shock down the chain closes your line on the same morning it hits your hold. Drawing does not widen it. Selling well does.
• Once 3 racks are running you can sell a share of the buildout at the same desk. What a buyer pays is a multiple of the run-rate you have left — racks times the day rate times the days remaining — so the same site is worth a fortune on day fifteen and very little on day twenty-eight, and nothing about it changed but the calendar. The bid averages nine tenths of that arithmetic and clears it about a third of the time. You can see which kind of day it is before you sell; waiting for a better one costs you the days you waited.
• You see one board at a time. The market you are standing in, and no other. That is the game.
• VS USUAL is the column that decides. It scores a quote against this market's own range — base, tilted by the bias in the table above, across the good's band — not against some global average. The two are not the same and the difference is the whole read: substrate at its plain base price is a bargain in Hsinchu, where the bias runs 0.60, and robbery in Santa Clara, where it runs 1.34. Buy the bottom of the range, sell the top. SPIKE means a quote has left the range altogether, which only an event does.
• VS HERE is the column that decides. It scores a quote against this market's own range — base, tilted by the bias in MARKETS.XLS, across the good's band — not against some global average. The two are not the same and the difference is the whole read: substrate at its plain base price is a bargain in Hsinchu, where the bias runs 0.60, and robbery in Santa Clara, where it runs 1.34. Buy the bottom of the range, sell the top. Every row says which it is in a word — CHEAP, FAIR, DEAR — because the percentage alone was being read backwards: “+42%” means the pallet in front of you costs 42% more than a pallet normally costs here, not that here is 42% cheaper than somewhere else. SPIKE and DUMPED mean the quote has left the range altogether, which only an event does.
• Offtake is a sale agreed before you own the goods. A buyer posts what they want, where, by when, and what they will pay — 7 HV transformers into Abilene by day 14, $2.39M — and that price is fixed the moment it appears. Fill it and you get the money whatever the board says on the day. Miss the date and nothing happens except that you are holding seven transformers nobody promised to buy. It is the only thing here that pays you for assembling a kit rather than flipping one line, and the only price in the game that does not move.
• A shock travels sideways. See CHAIN.TXT: what you read on the wire today is a bet you can place tomorrow, in a market the shock never touched.
• Two goods move under licence. Carry them into a hot market and customs may open an audit. Declare and lose the cargo; contest it if you have counsel on retainer; or expedite, pay, and find out.
• Standing starts at 100. A bad audit costs some. At zero you are debarred and the run ends there, whatever the money says.
• Score is cash plus escrow minus debt on day 30. Cargo still in the hold on the last day is worth nothing, so sell before the month runs out.
• A site, if you can afford one. Once, somewhere between day 8 and day 12, a shell pad is offered — but only if you are holding enough to actually start, which is the pad plus a whole turbine and a whole transformer. Racks go up there and nowhere else, so from that moment you are trading around an address. A rack is min(accelerators, stacks ÷ 12, turbines × 8, transformers × 2), and it earns $100,000 a day standing. That one line is the whole of it: the scarcest input decides, and buying more of whatever you already have most of buys nothing at all. Installing is one-way — hardware scores at 30% scrap on the last day, so a turbine you fit is a turbine you did not sell.
• One branch a day. Everybody who opens this console on the same date is dealt the same thirty days — the same boards, the same wire, the same customs rolls. Prices are drawn from the seed, not from the clock, so reloading is not a reroll and a bad board stays bad. That makes a score mean something: it is not how lucky you were, it is what you did with a month somebody else also got. Free run draws a private branch instead, which nobody can compare against and which is the honest place to experiment.

Ranks

• THE BUILDOUT 12 racks standing — money alone never gets here
• OFFTAKE KING $45,000,000 and up
• PRIME CONTRACTOR $15,000,000 and up
• PROCUREMENT DESK $5,000,000 and up
• FREIGHT BROKER $1,000,000 and up
• PALLET RAT anything above water
• LIQUIDATED zero or worse