The hardware behind the future


Apple and Nvidia Weigh Sending 2028 Chip Work to Intel’s Fabs

A DigiTimes report puts tariffs, politics and capacity, not process technology, behind talks over Intel 18A and 14A production and EMIB packaging in New Mexico.

Apple and Nvidia Weigh Sending 2028 Chip Work to Intel’s Fabs

Every so often this industry serves up a plot twist a screenwriter would be fired for proposing. Apple spent five years walking away from Intel silicon and finished the job in 2022. Nvidia has bought its manufacturing from TSMC for as long as most of us have been writing about it. Now DigiTimes reports, as relayed by Tom’s Hardware, that both companies are weighing whether to have Intel build some of their chips in 2028.

Neither company has confirmed the talks, and Tom's Hardware, which relayed the report, says so plainly. What is described is not a defection. It is a hedge, and the reasoning has more to do with maps than with transistors.

Tom's Hardware lists the pressures as geopolitical concerns, political pressure from the U.S. government, potential tariffs imposed on semiconductors produced outside the United States, and capacity constraints. Put a levy on imported silicon and a Mac processor made in Arizona stops being a sentimental idea and becomes an accounting one.

Apple's contribution would reportedly be its entry-level M-series processors, the parts inside the cheaper MacBooks and iPads. Those are the sensible ones to experiment with. They use small dies, cheap packaging, and they tolerate a little variation in yield and clock speed, which is another way of saying nobody notices if one runs a fraction cooler than the next. The A-series chips in iPhones and the bigger M Pro and M Max parts are not in the conversation.

The processes under discussion are Intel 18A, or more precisely 18A-P, and Intel 14A. Whether Intel will have enough of that capacity spare for outside customers in 2028 is unknown. It would also have to give Apple a clean way to move a design drawn for TSMC's rules across to its own without surrendering performance or battery life in the translation, and to do the packaging in America too, since U.S. policy is aimed at the silicon rather than the finished laptop.

Nvidia's side of the story is stranger. According to DigiTimes, it is looking at having Intel make some of the I/O dies for its Feynman generation of AI accelerators, and at packaging about a quarter of those Feynman GPUs at Intel's New Mexico facilities using EMIB, a technique that buries a small silicon bridge in the package substrate so neighbouring chiplets can talk at high speed without a full interposer underneath them.

Here the physics turns awkward. Tom's Hardware expects a Feynman GPU to draw five to six kilowatts, more than a kitchen circuit will carry, and at those currents you cannot sensibly push sub-one-volt power across a motherboard without losing much of it to heat on the way. The answer is to move the voltage regulator inside the package, feed it around 1.8 volts, and step down at the last possible moment. EMIB can sit beside such a regulator but not truly embed one, which Tom's Hardware argues is not enough for an accelerator this hungry.

Intel does have Foveros, which stacks dedicated power tiles underneath the logic and regulates very quickly. It is a different philosophy rather than a swap for TSMC's CoWoS-L interposer, and redesigning a reticle-sized AI GPU to suit a second packaging house is not a small favour to ask.

The likelier near-term prize, in that reading, is packaging Nvidia's Vera CPUs or building custom Xeon parts, neither of which has to survive six kilowatts. Nvidia has already put five billion dollars into Intel, so the relationship exists on paper. Turning it into wafers is a 2028 question, and 2028 is a long way from a report nobody will confirm.

Leave a Reply