How long can Intel justify having their own fabs?

For decades, Intel pretty much continuously had the most advanced fabs in the world.  This especially gave them an advantage over AMD, and before that, also Cyrix and VIA.  (VIA is still around, but doesn't provide the same sort of competition for Intel that AMD does.)  If AMD could do something in their fabs at one point, Intel could do it a year sooner, leaving AMD always a year or so behind.  To eliminate the fab advantage, you'd have to compare an AMD CPU at one point in time to what Intel had a year earlier.  That's not how customers operate, nor should they.

That's now gone.  With Intel's 10 nm process node being a catastrophic disaster, TSMC and Samsung are both now ahead of Intel for the first time ever.  That could just be a temporary hiccup before Intel reclaims their process node leadership.

But Intel's process node advantage may not be coming back.  It used to be not that expensive to operate a cutting edge fab.  Buy some hardware from this company and that, put it all together, and start cranking out chips.  There used to be a lot of companies that did this.  Yes, I'm glossing over a lot of complications, but 40 years ago, it didn't cost billions of dollars to bring up a new process node.

The problem is that every process node is more expensive than the last.  One estimate is that the cost of bringing up a new process node doubles every 4 years.  That has led one company after another to decide that they can't justify the cost and stop trying to keep up.  Keep running the old fabs on old process nodes, but don't bother with the investment to build anything new.

If your own chips are 20% of the world's volume and you're one of 20 cutting edge fabs, you can amortize that cost of bringing up a new fab over a lot more chips than your competitors.  You can justify added investment to be the first at everything.  If your own chips are 20% of the world's volume and you're one of 3 cutting edge fabs, you have to amortize that investment over fewer chips than your competitors.  That's not a good position to be in.

One solution to this is that you don't just fabricate your own chips.  You fabricate those of a bunch of other companies', too.  Get enough customers and you don't even need to directly sell your own chips to the public, as there's plenty of money to be had by fabbing chips for others.  Nvidia, AMD, Xilinx, Apple, and Bitmain may not have enough volume on their own to justify having their own fabs, but if you're TSMC building chips for all of them plus dozens of other companies, that's plenty of volume to justify the investment that it takes to stay at the cutting edge.

Intel tried to go that route in addition to building their own chips, and it seems to have flopped.  Their difficulties on 10 nm surely didn't help, but it's far from clear that it would have been successful even if 10 nm were right on schedule.

For one thing, they've traditionally been able to design their process nodes partially around the chips that they needed to make.  If all that a new process node needs to be able to handle is x86 CPUs and perhaps chipsets, that's not a very diverse set of chips.  If it needs to handle enormous behemoth chips from Nvidia and Xilinx, cell phone SoCs, SSD controllers, IoT devices, radios, and a bunch of other things, that's a lot harder.  Now, that doesn't all go on exactly the same process node.  TSMC and Samsung have a variety of process nodes optimized for different applications.  But Intel didn't have to do that, which helped them to be first to a given node.

For another, if you're fabricating a lot of your own chips, rivals creating competing chips have good reason to mistrust your fab.  If AMD hires TSMC to fabricate chips, they can be confident that TSMC will do what they can to make the process node work well for AMD.  TSMC has no reason to do otherwise, as they want AMD's future business.  Similarly if AMD hires Samsung.  But if AMD hires Intel?  Intel benefits enormously if AMD's CPUs are junk one generation, and AMD surely doesn't want to give Intel the ability to make that happen.

But it's not just AMD.  Xilinx is in the same situation.  Nvidia is certainly competing with Intel for data center compute products, and soon more directly once Intel launches discrete GPUs.  Apple would have reason to at least be wary with how hard Intel has tried to push into cell phones in the past.  Remember how hard Apple tried to switch from Samsung to TSMC when Samsung's Galaxy phones became the major rival to the iPhone?  And it's not just those few companies; Intel is in a lot of markets and is at minimum a threat to enter a lot more.  There really aren't that many companies that need a cutting edge process node and can be confident that Intel won't try to create a competing chip.

With TSMC, there are no such worries.  TSMC's ideal outcome in every single market would be that all the competitors in an industry get their chips produced at TSMC, the chips all turn out great, and they all feel the need to hire TSMC again for the next generation.  That's not quite true of Samsung, which produces their own cell phone SoCs and SSD controllers, among other things.  But it's a lot closer to being true of Samsung than it is for Intel.  Most of Samsung's products aren't produced on the sort of logic nodes where they're competing with TSMC and Intel.  Various vendors could be confident that Samsung isn't going to design their own high performance CPU, discrete GPU, FPGA, or a lot of other categories of products and compete with their own product.

Intel does have the dominant market share in x86 CPUs, which is a pretty big market.  But that has long been driven by their having superior products to AMD, which was in turn because they had superior process nodes.  If their process node advantage is gone, will they maintain the dominant market share?  If they don't, then they have to amortize their process node development costs over that much fewer chips.

At that point, Intel won't have enough of their own chips to justify the process node development costs, and will find it very difficult to bring in other foundry customers to add volume.  If that happens, then what?  That's the situation that led Global Foundries to abandon the leading edge recently, and before that, IBM to sell off their fabs, after AMD did so, after a lot of other companies abandoned the leading edge over the course of decades.

It seems to me that Intel is in a precarious position because of that.  The problem is that Intel needs both their fabs and their CPU design to be great to survive; other companies only need one or the other.  Maybe they'll be cranking out a bunch of awesome chips on a "7 nm" process node that is better than competitors' "5 nm" by the middle of 2021, and it will look like they've reclaimed their lead.  But even that will only offer a temporary reprieve.

And if Intel's fabs abandon the leading edge, what then comes of the company?  Why should they still be worth more than a hypothetical merger of, say, AMD, Xilinx, Realtek, and Global Foundries?