The iPhone Assembler Now Earns Most of Its Money From AI. Its Own CEO Says the Limit Isn't His Factories

Hon Hai Technology Group, the contract manufacturer most people know as Foxconn and as the company that assembles the iPhone, reported second-quarter 2026 results on August 12. The headline figures were records. The more consequential disclosure came in what its chief executive said about next year.

The quarter

The 51% figure is the structural fact in that table. A company defined in the public mind by consumer electronics assembly now takes the majority of its revenue from cloud and networking products, driven by volume growth in AI racks. Operating profit rising 68% while gross margin slipped slightly to 6.12% tells you how that happened: this is a scale business, and the profit came from volume and operating leverage rather than from richer pricing per unit.

The growth is not slowing yet

AI rack shipments grew more than threefold in the first half of 2026. For the full year the company forecasts AI server rack shipments more than doubling from last year, and expects shipments this quarter to grow by a high double-digit percentage sequentially — that is, compared with the previous quarter rather than the year-ago one.

Not every line is expanding. Computing product revenue is expected to decline slightly this quarter, and the stated reason is memory supply constraints. That is the same shortage visible elsewhere in the industry, arriving here as a limit on what can be built rather than as a demand problem.

The ceiling the company named for itself

The most useful thing for anyone modelling 2027 came from rotating chief executive Michael Chiang, who identified the binding constraint on next-generation AI server racks as TSMC's CoWoS advanced packaging capacity — explicitly not factory floor space.

CoWoS stands for chip-on-wafer-on-substrate. It is the packaging step that mounts processor dies and high-bandwidth memory together on a single substrate, and it is required for the accelerators that go into these racks. Only limited capacity for it exists, and it sits at TSMC rather than at the companies assembling the finished systems.

Chiang's framing is unusually candid for a manufacturer: he is saying Foxconn could build more racks than it will be allowed to, because the parts will not exist. The variable that decides 2027 volumes is therefore how much of TSMC's CoWoS allocation Nvidia can secure, measured against competing claims from AMD, Broadcom, Google's TPU programme and Amazon's Trainium.

That reframes what a reader should watch. Foxconn's own capital expenditure — NT$80.9 billion in the first half — tells you about its intent, not about its output ceiling. The number that governs the ceiling is set one step upstream, by a supplier, and split among buyers who are all competing for the same allocation.