NEW TAIPEI CITY, 13 AUG 2026 — For the first time in its history, Foxconn earned more from AI servers than from iPhones and every other consumer device it makes, combined. Cloud and networking products were 51 per cent of second-quarter revenue, against 29 per cent for consumer electronics.
The company known for assembling phones in Shenzhen is now, by revenue, an AI infrastructure business.
The quarter
Two things in that table deserve more than a glance.
Operating profit rose 68 per cent on revenue up 41 per cent. The gap between those two figures is margin improving, rather than volume simply piling up. For a contract manufacturer — a business model defined by taking a small cut of somebody else's product — that is the harder half of the achievement.
But net profit grew only 35 per cent, slower than operating profit and slower than the revenue line it sits under. Something below the operating line absorbed the difference: tax, currency, or non-operating items. Foxconn reports in New Taiwan dollars while selling into a US-dollar AI market, so exchange movement is the obvious candidate, and we could not establish from the published figures which it was. That detail matters before anyone treats 68 per cent as the new run rate.
And the absolute margins remain what they have always been. An operating margin of 3.75 per cent is a record for this company and would be a crisis for almost any other technology firm. Assembling AI racks pays better than assembling phones. The underlying economics have not changed category.
Where the ceiling moved
The most consequential thing said this week was not a number. Rotating chief executive Michael Chiang identified what limits AI server output in 2027, and it is not Foxconn.
It is CoWoS — TSMC's advanced packaging process, which bonds logic dies and high-bandwidth memory onto a single substrate. Floor space is not the limit here, and neither is labour or component supply generally. One process at one supplier is.
The bottleneck has moved. For two decades the constraint on electronics volume was manufacturing capacity, and manufacturing capacity was Foxconn's answer to everything: more floor space, more lines, more people. That constraint has been solved so thoroughly that it has stopped being the binding one.
Advanced packaging cannot be scaled the same way. It requires specific tooling with long lead times at a small number of sites, and TSMC allocates it across every customer building accelerators. Whoever is buying AI capacity in 2027 is queuing behind that allocation, whatever their supplier says about floor space.
Foxconn's own guidance sits on top of it: AI rack shipments are expected to more than double for the full year, with Vera Rubin racks entering mass production this quarter and high double-digit growth on-quarter and on-year for cloud and networking.
What a rack is, and why it pays better
This all hinges on the word rack.
Foxconn is not shipping accelerator boards for somebody else to install. NVIDIA's current generation is delivered as a rack-scale system — the Vera Rubin NVL72, 72 GPU packages wired as one machine, where the VR200 carries 288GB of HBM4 and roughly 50 petaflops of FP4 compute, about 3.3 times the throughput of the B300 it replaces.
Selling that as a unit means the integrator owns the power distribution, the liquid cooling loop, the internal fabric and the burn-in testing of a machine that draws more than most small offices. Populating a phone chassis is a far simpler job, and the difference shows up in the margin.
That widened scope is what put operating profit 27 points ahead of revenue. Foxconn is selling more of the job, at much the same rate. The value that used to sit with whoever assembled the datacentre has moved into the box, and Foxconn is now the one assembling the box.
It also explains why the third quarter is the one to watch rather than this one. Vera Rubin racks enter mass production now, which is when a new platform either yields well or does not.
What this looks like from ASEAN
Two things follow for anyone in the region planning capacity.
The first is that hardware lead times are now set upstream of the vendor you are negotiating with. A quote for 2027 delivery depends on a packaging allocation that neither you nor your integrator controls. Ask what the quote assumes about it — a supplier who cannot answer is passing you a risk without pricing it.
The second is a pattern we have seen three times this month alone. Indonesia's Zankore is buying a gigawatt of capacity whose delivery depends on GB300-class supply. OVHcloud is raising prices by up to 87 per cent because memory has repriced sixfold. And now the manufacturer with the most AI server capacity on earth says the limit is somebody else's packaging line.
Each company is describing the same problem from a different angle. The AI buildout has run into physical limits that sit upstream of whoever is selling you the hardware.
The concentration question
For a company this size, that 51 per cent figure is worth a moment's thought.
Foxconn spent thirty years being criticised for depending on a single customer. The iPhone made it and exposed it, and diversification has been the strategic story for a decade. That diversification has arrived. Its new segment takes demand from a handful of hyperscalers, buying accelerators designed by one company and packaged by one foundry.
It is not obvious that this is less concentrated than a single phone customer. The customer list is different and slightly longer. The dependency underneath it may be narrower.
Consumer electronics did not shrink; the rest of the company grew past it. At 29 per cent of a much larger revenue base it remains enormous, and it is the part with the seasonal peak still ahead of it in the third quarter.
What to watch
Whether the gap between operating and net profit growth persists. One quarter of currency movement is noise; two quarters is a structural feature of selling dollar-denominated hardware from a Taiwan dollar cost base.
Whether CoWoS allocation gets disclosed by anybody. Every AI capacity forecast for 2027 now rests on a number that no participant in the chain publishes.
And finally, whether the operating margin can hold above 3.5 per cent once AI racks become routine work. The current premium partly reflects scarcity — being able to deliver at all. Contract manufacturing has a long history of competing such premiums away. Racks may be hard enough to resist that. They may not be.
