Intel 3 reaches production this year, and it is the first node Intel built not just to make its own chips but to sell to other companies. Sierra Forest and Granite Rapids, the new Xeons, are the launch vehicles, following Intel 4 into the EUV era, but the more important audience is not Intel’s own product teams. It is the outside customers the foundry has been promising for three years, and Intel 3 is the first node engineered with them in mind, with the density libraries and design kits a foundry customer needs rather than just the ones Intel’s own designers use.
A foundry node has to do things an internal node never does: ship documented design rules a stranger’s engineers can follow, run on the standard EDA tools the rest of the industry uses instead of Intel’s in-house flow, hold yields predictable enough that a customer can plan a launch around them, and offer the kind of hand-holding Intel never bothered with when its only customer was Intel. Intel 3 is the first node where the company actually did that work, the unglamorous plumbing that turns a process into a product other people can buy. It is built for servers and the data-center silicon where the money increasingly sits, and it is meant to make Intel Foundry, for the first time, a credible place to send a design.
Intel 7, Intel 4, and Intel 3 have all shipped now, more or less on the schedule Gelsinger laid out, which is something almost nobody believed possible when he stood up and announced five nodes in four years. Three of five, landed close enough to on time that the credibility Intel badly needed is finally starting to bank. The part the scorecard hides is that these three are the catch-up nodes. They close the distance to where TSMC and Samsung already are; they do not pass anyone. The two nodes that are supposed to actually take the lead are still out ahead on the calendar, unshipped, and they are the truly hard ones.
Customers are watching Intel 3, not signing for it. The pitch reads well on paper, an American-built alternative to TSMC, courted by a government desperate to reshore chip manufacturing, on a roadmap climbing faster than anyone expected, but the leading-edge work still goes to TSMC, which keeps building everyone’s most advanced silicon with a reliability Intel has not earned back. A real node to sell was always necessary, and never close to sufficient on its own. The handful of foundry commitments Intel has been able to tout are aimed at 18A, a node still on the horizon, not at Intel 3. The question Intel 3 cannot answer is whether the big fabless names will actually trust Intel to build their chips, after a decade of watching it miss. No process node answers that. The only thing that will is a name on a contract for silicon Intel can build now, and so far there is not one.
Building five nodes in four years while standing up a foundry from scratch burns capital Intel no longer throws off the way it once did. The product business that is supposed to bankroll the whole effort is losing margin to AMD and watching its data-center share leak away, with AMD’s server share now climbed past a quarter and still rising, so the cash to fund the bet is getting thinner exactly as the bet gets more expensive. And the customers who would eventually justify it have not signed, which means Intel is spending now against revenue that is still hypothetical. Meanwhile TSMC does not slow down to let any of this resolve. Intel 3 is real, and that counts. It lands on a balance sheet that needs much more than one node going right.
Intel 3 will get less credit than it earns, sitting as it does next to the flashier leadership nodes still to come. It is the one that quietly proves Intel can build for somebody other than itself, which is the entire premise of the foundry. Perso, I think the real verdict was always going to land a node or two out, on 20A and 18A, where Intel has to beat the foundries at something instead of just matching them, and Intel 3’s actual job is to keep the comeback on schedule and the foundry breathing long enough to get there. It does that job well. What it does not do is win Intel a single customer or a dollar of foundry profit; that bill comes due later, on nodes that have not shipped yet, against a competitor that is not waiting around. Competent and on time, and still nowhere near enough. That is the most anyone at Intel can fairly say about it right now.