Purana Archer (ancientarcher.delete@this.gmail.com) on October 1, 2013 5:52 am wrote:
> As for capex, Intel will spend $11bn in 2013 with $2bn reserved for 450mm trial (not production, just trial).
> TSMC will spend $10.5bn and Samsung $7-8bn. GloFo will spend $5bn. The ARM ecosystem spends significantly
> more and all these players are making money. Just look at TSMC's P&L. So what is the outlook for Intel's
> edge in process node? How long will it sustain with these guys bearing down on it at such a rate...


Capex != investment in developing future nodes. What do you think the comparison between Intel and TSMC capex numbers would be if Intel had a single leading edge fab that they were building and equipping for 14nm, versus TSMC if they had a dozen fabs they were building and equipping for 28nm? TSMC's capex would be much larger than Intel's, but it wouldn't indicate that TSMC is catching up, only that they are running more silicon than Intel.

Capex numbers are not too meaningful, what we'd need to know is how much Intel spends per year to develop new process nodes, versus what TSMC spends. Then consider that since TSMC is behind they'd have to spend at much higher rate than Intel to catch up - because they'd need to try to do in the next two years what Intel has done in the past two years plus what Intel will do in the next two years. Theoretically that would mean double Intel's spending, but since R&D doesn't work in a linear fashion it would surely require much more than double.

However, Intel's "process lead" may not matter that much in practice. Consider that TSMC's goals are different than Intel's. Intel reportedly sacrifices transistor density in its design rules. This means they required complications like immersion or double patterning a process generation later than everyone else, which is partially responsible for their process lead. A foundry like TSMC wouldn't want to sacrifice density, because that raises the per transistor price. Intel is willing to take a hit in its per transistor price because it sells its transistors with the x86 near-monopoly premium. TSMC sells its transistors at the foundry industry standard markup. Because they have competition, while Intel mostly does not.

Look at Apple's A7, fabbed on Samsung's 28nm process. It has significantly better transistor density than Haswell on Intel's 22nm process (dual core A7 102 sq mm "over a billion" transistors, i7-4650U dual core Haswell GT2 181 sq mm 1.3 billion transistors, 15W TDP) They aren't identical chips by any means, but both are dominated by the GPU, have similar SRAM sizes, etc. One can quibble over how equal the two really are, but it is an interesting comparison when looking at foundry economics. Foundries may be equal to or even beating Intel where it matters for them - transistor density. If so, that's a problem for Intel, because it doesn't want to sell transistors at the same price foundries do, it wants to sell them at the same premium it sells x86 transistors.