Deep Dive — Foxconn wants 50% of AI servers at a 6% margin

Foxconn chairman Young Liu wants half of the world's AI server market. His company already has about 40% of it, he told CNBC's Christine Tan this week, and he expects to be at roughly 45% next year and "up to 50 percent" within a couple of years. His production capacity, he said, is fully booked for at least the next year and a half, in an interview with CNBC.

The share number is the one that will get quoted. It is also the least interesting number in the interview, because Foxconn's gross margin last quarter was 6.12% and its operating margin was 3.75%. Gaining ten points of a market measured in hundreds of billions of dollars, at those percentages, is a very different business from the one the headline implies: it is a volume business whose profit grows only as fast as the volume does, while the companies whose parts Foxconn screws into the rack keep the pricing power.

That gap — between share of the market and share of the money — is what this story is actually about. Foxconn is already one of the largest companies on earth, with NT$8.1 trillion in 2025 revenue (about $260 billion), roughly 900,000 employees at peak season and more than 240 campuses across 24 countries. It is now winning the AI build-out on exactly the terms it won the iPhone era: someone else designs the expensive part, Foxconn builds it beautifully and cheaply, and the margin stays small.

The quarter that shows the transformation is done

Foxconn's third-quarter revenue was NT$3.03 trillion, about $95.4 billion, up 47% year on year and well ahead of an LSEG SmartEstimate of NT$2.83 trillion, Reuters reported. September alone brought in NT$1.16 trillion, up 38% and the first month in the company's history above NT$1 trillion. Growth came from cloud and networking products — the division that holds AI servers — alongside the iPhone cycle.

The mix shift has already happened. In the first quarter, cloud and networking was nearly half of revenue; by the second, it was past half for the first time, which is a sentence that would have read as a misprint three years ago for a company whose identity was "the people who assemble your phone." Liu now says AI infrastructure will be 60% to 65% of revenue within three to five years, excluding edge devices.

The profit line did not make the same jump. Second-quarter revenue of NT$2.53 trillion was up 41%, and net profit of NT$60 billion was up 35% — good, but slower. Gross profit was NT$154.5 billion, an operating profit of NT$94.8 billion, a net margin of 2.37%, and earnings per share of NT$4.27, per the company's own second-quarter release. Return on equity for the first half was 6.21%, against a 12% target it describes as half-met.

Six percent, next to Samsung's seventy-two

Put Foxconn's quarter beside the one we covered on Wednesday. Samsung's chip division sold 147.1 trillion won of memory in July to September and earned 106.5 trillion won of operating profit on it — a margin above 70%, on parts that end up inside the same racks Foxconn assembles. That is the same physical object, the AI server, viewed from opposite ends of the value chain: the memory supplier prices scarcity, the assembler prices labour and logistics.

This is not a criticism of Foxconn, and Liu is not pretending otherwise. It is the structural fact that decides what "50% of the market" is worth. When the scarce inputs are GPUs and high-bandwidth memory — and they are, which is why 2027 DRAM and HBM capacity sold out months ago — the assembler is a price-taker on its bill of materials and a price-taker on its labour. It can win share by being the most reliable builder at the lowest cost. It cannot win share by charging more.

Which is why the sentence that should get more attention is not the 50% but the one about moving up the supply chain. "What we are going to do is we're going to move up the supply chain," Liu told CNBC, describing a US expansion aimed at higher-margin products first: AI server racks and data centre equipment. Read that as a margin strategy wearing a growth strategy's clothes. The content Foxconn makes itself — liquid-cooling manifolds, connectors, power shelves, the optics in a switch — carries a real gross margin. The GPU and the memory it bolts them to carry almost none.

There is evidence this is already working at the component level. Foxconn said its co-packaged optics switches entered mass production in the third quarter with shipments expected in the tens of thousands of units this year, and that revenue from 800G and faster switches should double in 2026. Those are Foxconn-designed products, not assemblies of someone else's bill of materials. The company showed the next step at its own trade show in Taipei on Friday: a "Token Factory" zone with a compute tray designed to run without cables, hoses or fans, its Vera Rubin rack integration, and sub-micron component work, part of a record 321 exhibits.

Why more share will not produce more pricing power

An assembler's moat is capacity and reliability, and both can be bought. Foxconn's answer to demand is capital spending, up more than 30% this year after NT$174 billion in 2025, with new capacity in Texas, Wisconsin, Ohio, California, Mexico, Vietnam and Taiwan. Its competitors are doing the same thing with the same customers: Quanta, Wistron and Inventec all build racks for the same handful of cloud providers, who multi-source on purpose. In that structure, "fully booked for eighteen months" is a statement about the order book, not about price. Memory can be sold out and raise prices because a fab takes three years to build. A rack line can be stood up in a year inside a building Foxconn already owns.

That asymmetry is the whole difference between the two halves of this supply chain, and it is why the memory makers' margins went vertical while Foxconn's stayed in the low sixes.

The 2027 revenue number will look better than the business

Here is the arithmetic that will confuse a lot of people next year. TrendForce expects AI server revenue to reach about $479 billion in 2026 and then jump more than 90% in 2027 to nearly $920 billion, while unit shipments grow about 36%, in its October forecast. Revenue growth of 90% on volume growth of 36% means roughly half of next year's dollar growth is price: denser racks, more memory per system, and higher component costs passed through.

For a component maker, that is margin. For an assembler, a large part of it is pass-through revenue with almost no incremental gross profit — the HBM inside the rack got more expensive, so the rack costs more, and Foxconn books the difference on the top line while earning the same few percent of assembly value. Foxconn's revenue can therefore compound at rates that look like a memory company's while its profit grows like a contract manufacturer's. The one metric that separates the two is gross margin per rack, and that is the number to watch in every Foxconn report from here.

"Overinvestment will kill the company"

Liu's most quotable line was a warning to his own industry: "Overinvestment will kill the company." He put Foxconn's AI spending at $3 billion to $5 billion a year, possibly more depending on how far vertical integration goes — a figure that sits below what the group already spends in total, since 2025 capital expenditure alone was NT$174 billion, about $5.5 billion, before this year's 30% increase. The gap suggests the number refers to a specific programme, most likely the US build-out. Either way, the caution is real, and it is unusual in a sector where the dominant rhetorical mode is to announce a bigger number than the last company.

It is also the correct worry. Foxconn's US expansion happens at American manufacturing costs, which is why Liu says the company will start with high-margin products and let scale bring costs down later. Wisconsin is the template: a further $569 million at Mount Pleasant and 1,374 jobs over four years, against up to $96 million in performance-based state credits and a total commitment of $1.2 billion in capital investment through 2029. The company has been in the US since 1985 and runs more than 40 facilities across 12 states, and last year agreed a design and US-manufacturing collaboration with OpenAI that carries no purchase commitments.

That is capacity built for policy reasons as much as for demand, arriving in 2027 and 2028, at a cost structure above Taiwan's, into a market whose growth rate is the single biggest assumption in the entire AI trade. If the hyperscalers' spending — $957 billion across nine companies this year, heading above $1.5 trillion next year on TrendForce's numbers — slows even to flat growth, the assembler with new American plants is the one holding the fixed costs.

Liu's answer to that is the honest one. Weaker demand from one customer can be offset by growth elsewhere, he said; an industry-wide slowdown is a much bigger problem. He also noted where the demand is not yet coming from: governments and enterprises, which have barely started buying. That is the reserve he is counting on, and it is also the part of the forecast nobody has tested.

What to watch

Foxconn reports full third-quarter results in November, and the only line that matters is gross margin — a mix shifting to AI infrastructure that pushes gross margin above 6.5% would mean the vertical-integration strategy is working; a mix shift that leaves it flat or lower means the company is buying revenue. October revenue in early November will show whether the September record was a peak or a step.

Then the Vera Rubin ramp. Foxconn's AI rack shipments are supposed to more than double this year, and the next generation lands in volume in the first half of 2027. Watch whether Foxconn's in-house content per rack — cooling, power, optics, connectors — rises with it, since that is the mechanism by which 50% share might finally pay better than 40% did.

And watch memory prices, for the counterintuitive reason. Rising HBM costs inflate Foxconn's revenue and compress nothing, because the cost passes through; falling memory prices would do the opposite to the top line while leaving the assembly margin intact. A quarter where Foxconn's revenue growth slows and its margin improves would be the best news in this story, and it will look like bad news in the headline.

The 50% target is achievable. Foxconn is the most capable contract manufacturer in the world, and the customers are already in the building. The question worth asking is what the company earns when it gets there — and on current numbers, the answer is a thin slice of a very large number. Anyone modelling Foxconn as an AI winner should model it as a toll booth on other people's scarcity, which is a fine business, and a very different one from the memory makers whose capacity sold out two years forward. We have written before about how the credit side of this build-out is already being tested — the neoclouds hit their own wall while the hardware orders kept coming. Foxconn's version of that test arrives when the capacity it is building in Wisconsin and Texas has to earn its cost of capital.

Is Foxconn's thin margin a warning about the AI build-out, or just what assembly has always paid? Write to us via our contact page.