The Caravan and the Cluster: a 1340 Instruction Manual,
Read as a Guide to Underwriting the AI Buildout

In one line: The camels never made anyone rich. The paperwork did. That was true on the Silk Road in 1340 and I think it is the right prior for the AI trade in 2026.

Abstract. Around 1340 Francesco Pegolotti, an employee of the Bardi bank, compiled La Pratica della Mercatura, a working manual for the overland China trade, from the reports of returning merchants. Its numbers survive: a worked example carrying 25,000 florins of cargo, total transport cost of roughly 1–2% of cargo value, a standard financing contract splitting profit 75/25 in favor of capital, and a realistic net return historians put near 10% for more than a year of mortal risk. Read carefully, the manual describes a business that was never really about moving silk. The caravan was a bank in motion: value converted from cloth to silver to state paper, margin skimmed at each conversion, the best economics reserved for merchants holding an imperial license, and the whole route exposed to single officials whose decisions could erase it. I map ten of the manual’s structural lessons onto the 2026 AI buildout and find that six of them correspond to forecasts I have already priced and published in the Forecast Wing. The correspondence is not decoration. It is a claim about where the load-bearing joints of this cycle are.

Keywords: Pegolotti, Silk Road, commenda, ortoq, AI capex, value chain, financing structure, forecast book.

1. The manual

Pegolotti never made the trip. He wrote at a desk in Florence, compiling what returning merchants told him: what to pack, whom to hire, what each stage cost, what a pound of silk fetched at the far end. The manual’s most famous sentence, that the road to Cathay was “perfectly safe, whether by day or by night,” was written by a man who never tested it. Hold that thought; it returns in section 6.

The numbers he recorded are the useful part. The worked example assumes cargo worth 25,000 gold florins, more than three hundred thousand days of a laborer’s wages. Moving that fortune from the Black Sea to Beijing cost 60 to 80 sommi of silver, call it 300 to 400 florins. Transport was 1 to 2 percent of cargo value. Moving a fortune across Asia was cheap. Having the fortune was the hard part, and almost nobody did, which is why the trade ran on other people’s capital.

2. The structure, not the silk

Three structural facts define the business, and none of them is about camels.

The financing. The standard contract, the commenda, put up the investor’s capital against the merchant’s year of travel and split profit 75/25 in capital’s favor. Genoa’s notaries recorded more than eight thousand of these. The trader who walked to China kept a quarter.

The relay. Almost nobody carried goods from China to Italy. Cargo hopped from oasis to oasis, sold and resold, each trader working the stretch he knew. The famous hundredfold markup on spices existed, but by the time silk reached a Genoese dock a dozen hands had taken a dozen cuts. The legend’s margin belonged to twelve different people.

The conversions. Pegolotti’s actual instruction was to travel light: carry linen as far as Urgench, sell it, convert everything into silver bars, carry the silver east, surrender it at the Chinese border for the Khan’s paper currency, and spend the paper on silk at fixed rates. The merchant ran the same value through three monetary forms and took profit at each conversion. The cargo was the costume. The trade was moving money between currency zones.

3. The license and the fine print

The best business on the road was partnership with power. Under the ortoq arrangement a Mongol prince advanced capital, and the merchant carried a paiza, a stamped metal tablet that worked as passport, credit line, and armed escort in one, opening the empire’s relay stations across five thousand kilometers. The Polo family traded under exactly this system.

Then the fine print. In the early decades the prince ate losses up to his invested capital, an equity structure. By the late 1200s the same silver and the same paiza had quietly become a loan: if the venture failed, the debt was the merchant’s, personally, without limit. Most merchants signed anyway. Same asset, same route, different loss-bearer, and therefore a different trade entirely.

4. What broke it

Two clauses ended ventures. A dead ruler dissolved the roads until a successor took power; trade waited on politics. And in 1218 the governor of Otrar looked at the richest convoy he had ever seen, four hundred fifty merchants carrying state capital under state protection, declared them spies, and executed nearly all of them. The strongest backer on the planet could not make the road safe, and the response, the Mongol invasion of Khwarazm, could not bring the caravan back. The risk unit was not the state. It was one official.

The route itself died the same way, from two directions at once. The plague travelled the same roads as the silk, the Yuan dynasty fell and China closed, and Italian cities responded by planting mulberry trees and raising their own silkworms. Within a generation the greatest trade route on earth was an asset class that no longer existed. Nobody on it ever called it the Silk Road; the name was coined by a German geographer in 1877, five centuries after the last caravan unloaded.

5. Who actually got rich

A good year paid a realistic net near 10 percent, after the duties, the drivers, the interpreter, and capital’s 75. The honest description of the asset was not a steady yield but a violent average of ruins and jackpots. The winners, the Karimi spice merchants at the road’s western end, were not braver and did not own better camels. They ran family agents in every major city, held state partnerships, earned as much from lending and exchange as from goods, and rerouted when a road died. The losers were single-venture men: all capital, one road, one role.

6. The 2026 reading

Each structural lesson above has a present-tense form, and six of them are already priced in my public forecast book (the Forecast Wing, fourteen pre-registered forecasts with binding resolution criteria).

The relay. The AI value chain is a relay: lithography, fabs, chips, power, cloud, models, applications. The headline “AI margin” is being divided along it exactly as the silk markup was. Any underwriting that books the full markup to one layer is counting seventy-five-to-one returns before the relay takes its cuts. My software forecasts (F13, F14) price the layer where I think the meter breaks.

The paiza. Export licenses, sovereign compute deals, and government equity stakes are the stamped tablet: partnership with power as the best business on the route. This is the thesis of my slate, intelligence abundant and permission scarce, and it is priced directly in F1 and F12, and in F5’s test of whose license the Chinese fab sector ultimately needs.

The fine print. The ortoq flip, equity quietly becoming personal debt, is the question to ask of every financed GPU fleet, vendor-financing loop, and collateralized compute contract: who eats the loss if the venture fails? The same cluster under equity and under debt is two different trades. F2 prices the balance-sheet version of this (tariffs financed as working capital, recognized at the refinancing wall); F7 tests whether the capex itself keeps compounding.

The caravansaries. The road worked because watering stops sat one camel-day apart. Infrastructure spacing, not merchant courage, set throughput, and the binding constraint on this cycle is the electrical equivalent: interconnection queues and turbine lead times. F8 prices it at its most observable point, the PJM capacity auction.

The dragoman. Pegolotti’s highest-paid hire was the interpreter: “a good interpreter’s wages will cost you less than what a bad one loses you.” The 2026 equivalent is the translation layer between models and domains, the integration and evaluation work, where scarcity value sits while raw capability commoditizes.

The mulberry trees. Import substitution ended the route’s demand side within a generation. The parties being disintermediated today are planting mulberries now; F5 prices whether the Chinese fab sector’s trees mature.

7. Limitations, and one methodological note

This is an essay, not a statistical result. Nothing here passed a gate, and no historical analogy resolves a forecast; the fourteen forecasts resolve on their own binding criteria between 2027 and 2030. The mapping earns its place for one reason: six independent structural features of a documented seven-century-old trade correspond to joints I had already priced before reading the manual. That is weak evidence that the slate is testing the load-bearing structure of this cycle rather than its headlines, and it is exactly as weak as stated.

The methodological note is Pegolotti himself. The man who wrote “perfectly safe” never left Florence, and the men who tested his sentence told a different story. Most guides to the AI trade are written at desks. This museum exists to be the other kind of document: the record of what the road actually did to the ideas I sent down it.