The silicon guilds and their dubious fortunes¶

Or how to separate the genuinely rich from the merely loud

The chip market resembles nothing so much as a particularly rowdy Ankh-Morpork tavern on payday: everyone is shouting about their prospects, coins are changing hands at alarming speeds, and somewhere in the corner a few sober wizards are quietly counting the real money. The cast rarely changes; only the numbers do.

The one printing money¶

At the top sits a near-monopolist of AI accelerators, enjoying margins that would make a medieval guild master weep with envy and a software moat that keeps developers loyal long after rivals match the hardware. The position has two structural weaknesses that no quarterly result removes. First, customer concentration: a large share of revenue comes from a handful of hyperscalers, each of whom resents the dependence and is designing its own chips. Being the only umbrella seller in Ankh-Morpork is splendid right up until everyone realises they can stay indoors. Second, scarcity cuts both ways: waiting lists sustain premium prices, and it brings to mind Cut-Me-Own-Throat Dibbler’s sausage cart, where the queue is always longest when there are only three sausages left, and somehow there are always only three sausages left.

The understudy¶

Below sits the perennial challenger, technically impressive and structurally second. Its hardware can beat the monopolist’s on paper, particularly on memory, and it enjoys the one advantage the leader can never claim: it is not the leader, and every large customer wants a second supplier for leverage alone. What it lacks is the software ecosystem, and a lovely new tavern with excellent ale still fails if all your friends drink at the other place. The understudy’s share climbs when the leader’s customers are angriest, which is a business model of a sort, though not one to build a guild on.

The fallen incumbent¶

Then there is the former master of the guild, attempting one of the most ambitious turnarounds in the trade’s history: rebuilding leading-edge manufacturing while its core businesses erode and its foundry loses billions waiting for external customers who keep not arriving. Spending fortunes on capacity nobody has yet booked is either a heroic long game or a very expensive hobby, and which of the two it is will be known only afterwards. The uncomfortable exhibit: its own most competitive AI chips have been manufactured by the rival foundry, which is rather like the Alchemists’ Guild inventing something useful and having to ask the Artificers to build it because their own workshop is on fire.

The landlord¶

And underneath everyone sits the architecture licensor, collecting modest rent on nearly every chip in the world. Design once, licence many times, collect royalties: an elegant model that yields steady billions rather than spectacular tens of billions, because owning the patent on hinges makes everyone pay you for hinges without ever making you as rich as the people building entire buildings. The landlord’s risk is its tenants: a majority of revenue comes from a few large customers, who increasingly design their own chips, making the landlord simultaneously more essential and more bypassable.

What unites the cast is that nobody truly knows where the market is heading, and current valuations price in rather more certainty than the participants themselves possess. As The Patrician might observe while the guild masters squabble: the thing about bubbles is not that they burst. It is that they always do, and nobody ever thinks theirs will.

The clerk’s brief¶

January 2026: The year is counted, and one guild crosses a line¶

Gartner’s preliminary results, published January 2026, put 2025 semiconductor revenue at 793 billion dollars, up 21 per cent, with Nvidia the first vendor ever past 100 billion dollars in annual chip sales and forecasts pointing to a trillion-dollar market by 2028. One guild member now books roughly an eighth of the entire trade. The clerks note that the last time a single institution held that share of anything in Ankh-Morpork, it was the Patrician, and he at least publishes no forecasts.

August 2025: A tariff with a door in it¶

In August 2025 the US president announced a 100 per cent tariff on imported semiconductors, with an exemption for companies building fabs in the United States; TSMC’s pledged US investment stood at 165 billion dollars, and Apple added another 100 billion to its domestic commitments at the same podium. The clerks read the design plainly: the tariff is not meant to be paid, it is meant to be escaped, and the price of escape is a fab. Protection rackets, they note, have operated on this principle for centuries, generally with less paperwork.

May 2025: The diffusion rule dies before it lives¶

On 13 May 2025 the US Commerce Department rescinded the AI diffusion rule two days before it would have taken effect, replacing tiered country limits on AI chips with guidance, including a warning against use of Huawei’s advanced processors anywhere in the world. Export control policy reversed within a single administration cycle while fab timelines run in decades. The clerks observe that the players must place ten-year bets under rules with a shelf life of months, and that this asymmetry, more than any tariff, is what the guild accounts keep paying for.

The guild accounts¶

The accounts have kept one shape through every entry: the money pools at the monopolist and its toolmakers, the challenger gains share in proportion to customer resentment, the fallen incumbent burns capital waiting for foundry customers, and the landlord collects rent throughout. What changed across 2025 and into 2026 is the scale and the politics: a single vendor past 100 billion dollars as of January 2026, and market access now negotiated fab by fab under tariff and export regimes that reverse faster than a fab can be roofed. The clerks’ standing assessment: the fortunes remain real, the moats remain rentable, and the sober wizards in the corner have started counting in geography as well as gold.