The Shenzhen Receipt

The receipt is folded into eighths, soft at the creases from a week in a shirt pocket. It is printed on the kind of carbonless paper that has gone slightly grey at the edges, in the way of all paper that has spent any length of time in a humid city. There is a column of part numbers, a column of quantities, a column of unit prices in renminbi, and at the bottom a hand-written total, in green pen, that does not quite match the printed subtotal. The discrepancy is, I am told, the matter of tea. The seller and I shared a cup of tea, and the price of the tea has been deducted from the total, and the entire transaction has been documented in a way that would make any auditor weep but that is, in the world it belongs to, entirely standard.

I am writing this with the receipt on the desk beside me. The receipt is from a stall in the old part of Huaqiangbei, Shenzhen, the electronics market that runs for several city blocks and that contains, at any given moment, roughly every semiconductor part ever produced. The receipt covers four trays of STM32 microcontrollers, total weight 1.3 kilograms, which translates, after some sorting, to about eleven hundred individual chips. The chips were sold to me by weight, not by part number, on the explicit understanding that some fraction of them would not work.

The receipt is for my client, who is rebuilding a control system for a glass-bottle inspection machine that has been in service since 2004. The original chip on the control board is no longer in production, has not been in distribution for nine years, and exists, in the global supply chain, only in the grey market of brokers, recyclers, and people in Shenzhen who sort old boards by weight. I have spent the last week in Shenzhen, on behalf of the client, finding the chips and bringing them home.

This essay is about that week.

The two supply chains

There are, in the consumer's mind, two electronics supply chains. The first is the official one. It runs through Mouser and Digi-Key and Arrow, and it ships in reels of three thousand, on tape, with full traceability, with date codes, with a paper trail that satisfies the most demanding industrial auditor. The unit prices are clear. The lead times are published. The chips, when they arrive, are exactly what was ordered, in the exact quantity, with serial numbers that can be looked up in the manufacturer's database.

The second supply chain is the one nobody writes about. It runs through Huaqiangbei, through similar markets in Shenzhen and Shanghai and a half-dozen other Chinese cities, through small brokers in Singapore and Penang, through workshops in Mexico and the Czech Republic where end-of-life electronics get disassembled and sorted, through eBay and Taobao and the entirely informal network of people who know people. The unit prices are negotiable. The lead times are: whenever the shipment from the Foxconn recycling yard arrives. The chips, when they arrive, are mostly what was ordered, in approximately the right quantity, with date codes that have, in some cases, been polished off the surface and replaced with new ones.

The official supply chain is the one that gets written about in trade journals. The unofficial supply chain is, by volume, the one that keeps the existing installed base of industrial electronics alive.

What is actually in the trays

The trays I bought in Huaqiangbei contained, I estimate, about eleven hundred STM32F103 microcontrollers. The chip is a workhorse: introduced in 2007, used in approximately every cheap industrial controller built in the following decade, still in steady demand from a long tail of products whose original designs were locked in fifteen years ago. ST still produces the F103, in principle, but on lead times that make the F103-dependent industrial-controller market, in practice, dependent on whatever the brokers can find.

I sorted the chips, the morning after I bought them, on a folding table in a guesthouse near the market. The sorting procedure was, in essence: look at the chip under a loupe, decide whether the marking is genuine, the package shows signs of having been desoldered from a previous life, the leads are bent or oxidised, and the date code is plausible. About sixty per cent of the chips passed visual inspection. The remaining forty per cent went into a separate tray, to be sold back to a different broker, at a lower price, who would in turn sell them to somebody whose tolerance for cosmetic flaws was higher than mine.

Of the sixty per cent that passed visual inspection, I tested every chip electrically. I built a small jig, on a piece of perfboard, that could programme each chip with a brief firmware that exercised its peripherals — a UART loopback, an I²C ping, an ADC reading of a reference voltage, a small SPI transaction — and report success or failure over a serial cable to my laptop. The jig took a chip through its paces in about forty seconds. I tested chips for two days. About eighty per cent of the chips that had passed visual inspection passed electrical testing. About fifteen per cent had one or more peripherals that did not work but were otherwise functional. About five per cent did not respond at all.

The arithmetic, when I did it on the back of an envelope: of the eleven hundred chips I had bought, about five hundred and thirty were fully usable. The all-in cost per usable chip, including shipping and the cost of my own time, was about ninety US cents. The official-channel cost for the same chip, at the time of writing, is around eight US dollars when it can be found at all, on lead times of twenty-six weeks.

Why this works

The industrial-controller market that depends on these grey-market chips is, in one sense, perfectly rational. The original equipment was specified, in 2004, with a thirty-year design life. The chips that are needed to keep it running, in 2026, are simply not produced any more in the volumes that the official supply chain can serve. The market has, accordingly, produced an unofficial supply chain that does serve them, at the cost of giving up traceability, date codes, and the warm certainty that the chip in your hand was made yesterday in a clean room rather than recovered last year from a Korean television.

The market also produces a particular kind of engineer, who is comfortable working in this environment, and a particular kind of firmware, which is written defensively against the assumption that some fraction of the chips will have undocumented errata. You write your peripherals so that if a peripheral does not initialise, the chip can still boot, log the failure, and continue operating with reduced functionality. You design your boards so that the most failure-prone parts — the SPI flash, the Ethernet PHY, the crystal oscillator — can be replaced in the field without specialised tools. You assume that the chip on the bench in front of you might, in some small respect, lie about its identity, and you write code that can handle the lie.

This is, I think, a discipline that the official supply chain does not teach. The engineer who has only ever ordered chips from Mouser, on reels of three thousand, has no reason to believe that a chip might not be exactly what its marking says it is. The engineer who has spent a week sorting trays in Huaqiangbei knows better. The knowledge is, in some sense, a kind of competence that the official world does not have a name for.

The receipt

The receipt in my pocket is a small piece of paper, but it represents a piece of infrastructure that is, in aggregate, vast. The glass-bottle inspection machine that my client runs is one of perhaps ten thousand machines, across India and South-East Asia, that depend on a thinning supply of obsolete STM32 chips, that depend in turn on a Shenzhen market, that depends in turn on a global flow of end-of-life electronics, that depends in turn on the recycling and sorting workshops that keep the whole thing moving. None of this is documented. None of it appears on the manufacturer's website. The price discovery happens, mostly, in person, over tea.

I am going home tomorrow. I will declare the chips, at the airport, as samples for engineering evaluation, which is true. I will ship them, when I get home, to my client's facility in Mumbai, where the bottle-inspection machine has been sitting idle for two weeks waiting for a replacement controller board. The board, when I have it, will outlast the machine. The chip on the board will outlast the board. The receipt, folded in my shirt pocket, will be filed in a manila folder and forgotten, until someone, in some future audit, asks where the chips came from.

The answer to that question is, in the most literal sense, on a folded piece of carbonless paper, soft at the creases from a week in a shirt pocket, in a guesthouse near Huaqiangbei.