Intel’s Foundry arm is picking up the customers SemiAccurate told you about last year, but is this a good thing? On the surface it is, but some of the problems that sank previous attempts at becoming a foundry lurk in the shadows once again.
SemiAccurate has a history of bringing you the news of Intel’s process highs and lows months or years before anyone else. We say this not to pat ourselves on the back, just to give you some background on our sources. Remember the Intel 10nm meltdown that started with Cannon Lake? SemiAccurate had exclusive news of that in 2016, and followed that with several technical articles on what exactly was going wrong. Then there were the problems for customers, take Ericsson’s near death experience for example. As with everything else, we got some bits wrong too, but that does happen now and again. The point being we know Intel’s process tech and roadmap pretty well.
While it is out of the scope of this piece, suffice it to say the near decade of Intel’s wandering in the process wilderness started at the 22nm node and continued until Pat Gelsinger fixed things during his tenure as CEO. Intel didn’t have a technical problem, they had a cultural problem, the technical problems were only a symptom of the internal rot. That is what Pat changed, or at least made lots of headway towards the goal. Instead of being off the rails, process tech was on the right track and finally moving forward.
Another thing Gelsinger did was make a separate foundry unit so Intel could finally take on external customers with their fabs and related process tech. Some called this Foundry 2.0 but it is really 3.0 even though you could make a case for 4.0. Lets just stick with 3.0 for the current iteration that begins with the 18ap node. (Note: 18a is an internal process that was finalized before the foundry arm was formed and is not really suitable for external customers. This is not a reflection of the process, just what offerings and support were developed.) So 18ap can service external customers but where do you make the wafers, in location and capacity terms? This was a perennial problem for the company that caused a lot of friction for Foundry 1.0 and 2.0, a problem Gelsinger was not about to repeat. (Note: Yes we know)
There were a few major problems that killed almost every customer effort on previous foundry forays, yield, IP, and dedicated capacity, plus a few other ancillary issues. Before Gelsinger, Intel buried these problems, no one admitted to them much less addressed them. This lead to severe problems not being fixed, bad management has it’s consequences as does lack of board oversight. Spending billions to hide the Snow Ridge debacle works as long as you have billions to burn. By the late 2010’s, AMD was eating Intel’s lunch and things looked dire. They couldn’t continue as is, and no one tried to catch those spears until Pat stepped in and did just that, up front, early in his tenure. It didn’t fix things but unquestionably set Intel on the right path.
So with most of the process problems finally being addressed, that left IP and capacity. The IP issue was thorny, Intel had a culture of NIH that nearly sunk them during the dark days of the 10nm process failure. They double tracked most products, one for internal nodes and one for TSMC. This worked well but was a tad costly, doing two full designs wasn’t feasible, especially if the tools used for the internal layout were not compatible with the ones used for TSMC, you had to do two complete designs from scratch. If the staggering cost of that wasn’t enough, what do you think the opportunity cost of having half your engineers do work that would never be used was? This was changed before Gelsinger became CEO by moving from internal Intel developed tools to industry standard ones. From what SemiAccurate hears, this was a very messy internal ‘debate’.
With the designers using standard tools and workflows, much of the two track process could be collapsed into a single project, something we think literally saved the company. This also lead to another critical advance, the ability to use external IP. Prior to this, a foundry customer who wanted something, say a particular SERDES, that Intel didn’t make for their own chips was told, “Have fun making it yourself”. If that same customer went to TSMC and asked the same questions, they were given 73 options that were pre-rolled and ready with the click of a mouse. Going Intel drove up the cost of any project and dragged out timelines because basic plumbing blocks needed to be made from scratch. Remember how Intel signed Cisco, Panasonic, and others for previous foundry rollouts? We ask for no particular reason.
So with that mountain climbed, the only real one left was capacity. On nodes like 18a, anything that uses EUV really, a fab costs in the high billions of dollars or low 10s of billions to make. The physical building, the shell, is the cheap part, the tools are the expensive bit. Worse yet the delivery times are a tad longer than people might expect, think many years, often longer than the ~3 years it takes to build and equip a fab. If you want to build a fab in the EUV era, you need to plan way way out or you will be waiting a lot longer than you might prefer for your care package from ASML among others.
There is no substitute for capacity in the foundry business, you can either make the product a customer wants in the time they need it or you can’t. Intel wasn’t very good at this part of the customer relationship in the past but for Foundry 3.0, they were out in front of things and out in front publicly. This meant not only building out capacity in a serious way, but also building out advanced packaging capacity in proportion to it. The advanced packaging side of things is a much quicker job, the fabs themselves lagged. Intel’s advanced packaging business and customer portfolio is a bright spot on their balance sheets right now, the plan was for foundry to do the same.
Unfortunately for Intel, fabs aren’t cheap to build as we discussed earlier. Building many fabs costs many times as much, and Intel was doing just that. People accused Gelsinger of, to paraphrase one financial wonk, “spending money like a drunken sailor”. SemiAccurate isn’t a financial outlet so we won’t opine on this angle other than to say Intel wasn’t in the greatest financial shape in the first few years of Gelsinger’s tenure as CEO. This wasn’t a problem of his making, he just exposed the skeletons that were overflowing the stuffed closets from prior CEOs and allowed by a board that seems to think oversight is something you do when trying to look over a wall. Exposing things and cleaning them up is painful, just because a CEO had the spine to do it where his predecessors didn’t doesn’t mean it was his fault.
That said, Intel was in dire financial shape while spending billions and billions, and had even more committed on an ongoing basis. They needed fabs and fabs cost lots of money so you need to spend vast sums to get from point A to B in this game. These two realities ran head on into each other. Publicly. And the fallout was handled about as badly as one could imagine. Little or no effort was seemingly made to explain to the financial world why this spend was necessary. There were some platitudes like, ‘we will fill them’ or, ‘I would rather have too much than too little’, all paraphrased.
For some reason the financial world didn’t like that answer. Intel said outright they had clients, real ones, for the foundry wing and these buildouts were necessary. When asked questions like, “can you name a customer”, Intel did the hot shoe dance in public. This was interpreted as the company being a tad looser with the truth than investors liked and the stock was punished. Intel was in a bind, they had real customers, big customers, and the numbers they touted were quite real. Unfortunately they also had contracts with said big, real customers that said they could not name those customers.
Luckily SemiAccurate wasn’t bound by said contracts. We directly named Apple as a customer over a year before it became public. We also directly named Nvidia and Microsoft in that piece, those two were not public yet but we know they are customers. In early 2025, over a year before the Terafab was announced, SemiAccurate told you about Elon Musk’s intent to buy Intel. This did not come to pass even if the intent was there, instead the two companies have effectively joined forces to make a fab in Texas. That means Intel has signed Apple, Nvidia, Microsoft, and SpaceX/Telsa/whatever to make chips at or with Intel’s foundry services. With four ‘minor’ customers like this, can big customers be far behind? (Authors note: That was sarcasm)
Unfortunately while SemiAccurate was screaming about these customers, Intel wasn’t and it shattered their reputation among the financial world. Money was going out in large quantities and coming in the door in far smaller quantities. The customers were real, large, signed, and the capacity needed to service them was being built. Then one day, all of a sudden, Pat Gelsinger was on the unemployment line. SemiAccurate thought this was rank stupidity, basically par for the course for Intel’s board.
A few months later, Lip-Bu Tan was appointed Intel CEO. Our view at the time was if he keeps things on the same path as Gelsinger laid out, Intel could turn out OK. Looking back, Lip-Bu did just that, and placated Wall Street in a big way. The stock is up, investors are happy, and Intel is selling everything they can make. Server CPUs are in short supply, the consumer business is doing great, and things are looking up. One of the main ways this was accomplished was to cut things, employees and capital spending being the two big ones. While we feel terrible for the employees let go, the spending, specifically spending on the fabs, is what we will focus on for this piece.
Gelsinger had plans for many new fabs at many new sites, massive expansion of advanced packaging facilities, and all the ancillary bits that go along with it. The tools to fill those fabs were on order, Intel’s place in line was secure, and things were looking good on the technical side. The process development pipeline was back on track after a decade of abject failures, and birds sang songs in the trees of Santa Clara. Everything was right. Unless you looked at the balance sheets under Pat.
Many have said Intel could not continue spending like they did. Many others have argued that they could have if messaged right. Either way when Gelsinger was shown the door and Lip-Bu Tan given the hot seat, then spending was cut drastically and quickly. Again, some would argue that this saved the company and the collective Wall Street seemed to agree. Intel’s stock went up, the whispers went from dire predictions of corporate death to warm smiles, all was good again. But was it really?
The reason we bring this up is simple, Gelsinger’s main premise for building out capacity at a staggering pace was that he saw it as being needed about the time the new fabs would come on line. When you don’t have capacity, building it is ~3 years out if you have the tools available. The critical tool pipeline is a tad constrained at the moment so that 3 year timeline may be optimistic if you are not in line before the first shovel hits the ground. In short if you don’t have it, you can’t get capacity in a time frame that matters.
Gelsinger saw this and was doing something about it. Lip-Bu saw the numbers and did something about it. Those fabs and that capacity were halted, again this being a good or bad thing is your call. What you can’t argue is that right now Intel is at capacity, they have said so repeatedly on multiple quarterly calls. If they could make more, they would sell more, it isn’t a yield thing, it is a demand thing. If you look at the time frames of making a fab, that capacity that Pat was building out would have come on line roughly when Intel hit those capacity ceilings. What a coincidence, eh?
So as this is written, the industry is at capacity. TSMC is sold out, Samsung may have a process that yields someday, and Intel is at capacity. This means high prices for what is available and also not enough is available. Any more would mean big bucks for the company that can provide wafers. Intel would be there but Lip-Bu canceled several of those fabs. Additionally Intel might also have gone bankrupt if those fabs weren’t cancelled. Was it the right decision? Both visions are correct for their overarching direction but both have very apparent down sides. What was visionary leadership and what was a critical failure is for you to debate. Have fun.S|A
Charlie Demerjian
Latest posts by Charlie Demerjian (see all)
- Did Intel’s Cost Cutting Bite Them? - Oct 14, 2026
- When Is Nvidia releasing N1X/Spark Systems? - Oct 2, 2026
- Qualcomm Finally Fixes Their PC CPU Line! - Oct 1, 2026
- What Is Intel’s Diamond Rapids Server CPU? - Sep 11, 2026
- ARM’s 2026 ‘Tech’ Day Is A Self-Inflicted Wound - Sep 7, 2026
