Firstly it must be considered that circumstances are different now.
Back then foundries were about manufacturing and supply and the machine-readable interface between design and manufacturing was relatively straightforward and clean. You put in your masks and you got your chips back. It has become a lot more complex and iterative now with IP coming from multiple sources and at the leading-edge much of it is by definition not silicon-proven.
So now, I would argue, there are are two things to think hard about; manufacturing capacity and manufacturing technology.
Right now there is plenty of foundry capacity available; but not at the leading-edge 28-nm process node, exactly the place where there has been an arms race in processors for mobile applications.
That foundry undersupply at the leading edge is accentuated by the fact that fully-fabbed chip giant Intel Corp. is turning up the heat on the foundries – and through them on Qualcomm and others in the mobile device sector – with its launch of 22-nm FinFET silicon.
If Qualcomm and others were to team up and invest in a joint venture wafer fab, presumably with an experienced chip manufacturer – such as a TSMC, Samsung, Globalfoundries or UMC – it might help them with capacity at the 22-nm/20-nm node but only in 2013 or 2014. And they could potentially be outflanked in terms of process technology or any fabless chip companies that Intel my chose to work with.
Therefore the more important need is for Qualcomm those of its fabless peers who want to compete at the leading edge to collaborate to help the foundries make up ground on Intel in terms of process technology. And that requires a longer term view and a focus on how 14-nm silicon is going to be made and ramped into volume production. In effect, Qualcomm and its peers may be able to avoid the capital cost of owning their own wafer fabs but they can no longer ignore the cost and progress of process development R&D.
Qualcomm for one has said it does not deem it necessary to become an IDM to continue to thrive in the semiconductor business, but that it will likely get closer to its foundry partners. I interpret that to mean that Qualcomm and its leading-edge peers will have to spend R&D money in aggregate to help TSMC and one or two other foundries compete with Intel, but in return could get first claim on the capacity to manufacture the resulting leading edge process technology. And a joint venture fab part funded by the customers would be one way to account for it.
I wonder if circumstances are really as different today from what they were in the last joint venture backed fab episode?
Companies have been going fabless to begin with because *building* fabs is so enormously expensive that very few companies can afford to do it on their own, and it's likely a few that could simply see better returns on investment elsewhere. Joint ventures are fraught with inherent difficulties because of the issues of getting agreement from the partners on what is being done, what each is expected to put into it, and what each will get out of it, and *maintaining* that agreement as changes in the underlying economy require adjustments to the plans.
If I'm Qualcomm, I'm leaving money on the table because I can't get all the chips I want, but how *much* money is that? How much will it cost to build this joint venture fab? How long will it take me to recoup my investment? And what happens when the wheel turns as it always does and we move to a condition of oversupply again?
I may look at the numbers, and decide I'm better off *leaving* money on the table, because the amount of potential revenue I'm forgoing won't approach what it would cost me to build a fab, even with partners, to address capacity issues that are ultimately transitory.
I think that now is a good time to be an IDM. The fabless chip companies can't have it both ways. I agree with others that it is too late for them to catch Intel by the 14 nm node anyway. This problem of constrained capacity at the leading edge could go away in a couple of quarters between improving yields and more capacity coming on line. I don't think that now is the time to go back to a failed model - like committing to a JV if you are Qualcomm. I think there is more downside risk than upside potential there.
I think ultimately that Qualcomm and other chip companies would just as soon continue to outsource the whole enchilada to TSMC and others, rather than get in bed with TSMC in a JV. As you point out, Peter, owning a fab, or owning it in part, is not simply a matter of opening up a magic chip-building facility. There is a lot of work and know how that has to be done in terms of process development. All of that would have to come from TSMC or another foundry. In the end, it wouldn't be that much different from the situation that Qualcomm and others enjoy today--they pay money to foundries, the foundries provide the knowledge, capability and capacity to build their chips. Entering a JV would only mean that a fabless company was on the hook for the depreciation of the facility and equipment, which would be a liability as soon as oversupply conditions took hold. And oversupply is never that far off.
Not only that, but Qualcomm and its fabless brethen today enjoy the right to work with different foundries, play them off each other, and negotiate the best deal. Once they are a partner in the fab, that flexibility goes out the window.
Of course the advantage of such an arrangement, on the surface, would appear to be certainty of supply. Right about now, that sounds great. But as Qualcomm executives acknowledged the other day, at least part of the problem with the 28-nm capacity crunch is that demand for Qualcomm's 28-nm chips is simply greater than the company anticipated. If Qualcomm had been able to accurately predict what the demand would be, TSMC would have had the capacity available, I think.
For the last decade or so our Fabless Wonders have had a free ride on process R&D done at "credible" IDMs ( in my book after their HKMG fiasco IBM no longer qualifies ) siphoned out through various consortia and tool vendors.
They may now have $$$ on the table but they can't buy time. Intel started finFET work almost 8 years back. Hard to catch up even at 14 nm within 5 years even if the fabless wonders were now able to pry off enough process talent on behalf of their Foundries and went at it using the traditional chinese water torture approach to Physics ( a huge no. of experiments over a wide parameter space rather than targeting on narrow space with the benefit of device simulation ).