Foxconn Now Makes More From AI Servers Than From iPhones

Foxconn reported second-quarter net profit of NT$59.97 billion, roughly $1.86 billion, up 35% year over year from NT$44.4 billion and ahead of the LSEG consensus of NT$58.8 billion. It is a record second quarter for the company. But the number that reorganizes how you should think about Foxconn is buried in the segment table.

The cloud-and-networking division — the business that builds AI servers and rack systems — accounted for 51% of total revenue, up from 48% in the first quarter. It has never crossed 50% in a single quarter before. Consumer electronics, the segment containing the iPhone assembly business that made Foxconn a household name in supply-chain circles, came in at 29%.

The world's largest contract electronics manufacturer now derives more revenue from AI infrastructure than from consumer devices. Operating profit in the server segment rose 68%, considerably faster than revenue, which means the mix shift is not just larger — it is richer.

What actually flipped

It is worth being precise about what happened, because "Foxconn pivots from Apple to AI" is a headline that flattens something more interesting.

Foxconn did not lose the iPhone business. Consumer electronics revenue is not collapsing. What happened is that a second business grew past it — from effectively nothing four years ago to the majority of a company that books well over $200 billion in annual revenue. The AI server line did not replace the phone line. It was built alongside it, at a velocity that would be implausible for any manufacturer without Foxconn's existing footprint in tooling, labor, and Taiwanese supply-chain proximity.

That velocity is also why the margin structure changed. Contract assembly of consumer hardware is a famously thin business — single-digit operating margins, and Apple holds the pricing pen. AI rack integration is not thin in the same way. Foxconn is not just screwing components together; it is doing thermal design, power delivery, liquid cooling loops, and system-level validation on machines where a single rack can cost more than a small apartment building. The 68% operating profit growth against 51% revenue share is what that difference looks like on a P&L.

The ceiling has a name

Chairman Young Liu, asked what constrains 2027, did not say GPUs. He said CoWoS.

Chip-on-Wafer-on-Substrate is TSMC's advanced packaging process — the step that bonds logic dies to high-bandwidth memory stacks on a silicon interposer. It is the operation that turns a Blackwell or Rubin die into something that can actually be mounted on a board. And it is the narrowest point in the entire AI hardware pipeline.

This is a genuinely important thing for a systems integrator to say out loud. Foxconn's own capacity is not the limit. TSMC's wafer starts are not the limit either — TSMC is ramping 2nm toward 100,000 monthly wafers by year-end, pulling 3nm to 180,000 wafer starts a month two to three months earlier than planned, and putting A16 into production in the second half of this year. The silicon is coming. What decides how many finished accelerators reach a rack is how many packages TSMC can assemble, and packaging capacity expands on a slower, more capital-intensive clock than front-end fabs.

Every AI capex forecast implicitly assumes this bottleneck loosens. Foxconn just told you that the company physically building the racks does not assume it loosens fast.

Why this shows up in your portfolio, not just your supply chain

Foxconn is the closest thing the market has to an honest instrument for AI buildout volume. TSMC's monthly revenue tells you about wafers. Nvidia's guidance tells you about bookings, which are a promise. Foxconn's segment mix tells you about systems that shipped — physical racks that went into physical buildings and got plugged into physical power.

At 51%, that instrument is now saying the buildout is not decelerating. Foxconn guided to AI rack shipments growing several-fold across 2026. Against a backdrop where the Magnificent Seven shed trillions in market value earlier this year on capex-reckoning fears, and where data-center power siting has become a live political fight in a half-dozen US states, the manufacturer's order book is stubbornly, boringly full.

The risk in that picture is concentration, and it is real. A business that is 51% AI servers is 51% exposed to a customer set of maybe a dozen hyperscalers and neoclouds, several of which are financing their purchases with structures — vendor guarantees, third-party capital platforms, rent backstops — that did not exist two years ago. If the financing wobbles, the order book does not degrade gracefully; it stops.

The quiet reframe

For fifteen years, the analytical shorthand was that Foxconn's fortunes tracked Apple's. That relationship is now a minority of the business. The company that assembles your phone has become, primarily, the company that assembles the machines that will eventually make phone assembly a smaller part of what anyone cares about.

Watch the Q3 print for two things: whether the segment share holds above 50% rather than oscillating, and whether Liu revises the CoWoS commentary. The first tells you the mix shift is structural. The second tells you whether 2027 is supply-constrained or demand-constrained — and those are very different years.