Historically, semiconductor manufacturing has been a brutal industry yet Intel has outperformed its peers by a wide margin. Here’s what I think is going on…
In semiconductor manufacturing, there is also an incentive to invest in cutting-edge manufacturing as new technology can cram more transistors onto a piece of silicon, resulting in lower operating costs. It is cheaper to invest in cutting-edge manufacturing when there are large volumes behind it to drive economies of scale. This dynamic encourages companies to produce in higher and higher volumes (in terms of transistor count). If a company does not heavily re-invest into new manufacturing, they risk being uncompetitive against other companies with lower operating costs. The economies of scale favour the #1 market share leader. All the other companies are vying for that #1 spot because otherwise they risk falling behind in the technology race and being priced out of the market.
It can create a brutal dynamic where at least some of the companies in the field are driving prices down to gain more volume. Andy Grove writes in his book Only the Paranoid Survive that one of Intel’s Japanese competitors had a memo that instructed its sales force to price 10% below competitors. Here’s the crazy thing about the memory business. One idea floated around in Intel was to significantly increase production to drive down costs and make Intel’s memory business profitable again. This is an industry where some participants think about increasing supply when the industry is in oversupply. Intel eventually exited the memory business because it couldn’t compete anymore. Andy Grove admits that he dragged his heels in exiting the memory business and that there was strong internal resistance to doing so. Since then, the memory business hasn’t been that great as many other manufacturers have gone into bankruptcy or encountered other financial trouble (e.g. Hynix, Rambus, Elpida). Memory has not been a great business despite the huge growth of computing.
What makes Intel different?
- One factor is that Intel has only one major competitor (AMD) in the x86 CPU market. There is only one other company that can initiate a price war against Intel. (In the past, Intel had essentially initiated a price war against AMD when it was paying computer manufacturers not to use AMD products. Intel was later sued for this practice and Intel paid a significant sum of money to AMD.) Other CPU manufacturers have a hard time competing against Intel and AMD due to the entrenchment of the x86 instruction set in the desktop PC. The entrenchment of the x86 instruction set is likely less of a moat today.
- The disparity between Intel and AMD’s market share is self-reinforcing as Intel enjoys greater economies of scale and can afford to wage repeated price wars against AMD.
- The CPU business provides greater rewards to manufacturers that have the latest fabrication technology. Both memory and CPUs get faster with newer manufacturing technology (e.g. when the process size shrinks). However, the speed of memory (e.g. memory bandwidth, memory latency) has little effect on an overall system’s performance whereas the CPU’s speed has a very large effect on overall performance. Computer manufacturers usually draw more attention to the speed of the CPU and the amount of memory than the memory’s speed/bandwidth/latency. Because having the latest manufacturing technology is more rewarding for CPUs, Intel has enjoyed greater scale advantages in CPUs than in memory (see #2).
On the other hand, maybe Intel’s moat is not as durable as I think it is. Intel tried to enter many other markets and did not do so well despite its manufacturing advantage. It used to make a discrete video card (the i740 was released in 1998), failed in the FPGA market, has seen many failures in the smartphone/tablet SoC market so far, and killed off its RISC processor (i860, released in 1989). Its Itanium processor (based on a VLIW architecture that was supposed to be better than RISC) has had limited success. Intel has a hard time breaking into new markets despite its manufacturing advantage. It is likely that Intel has trouble with the software and design aspects of its new products. Some semiconductor markets require software (such as compilers and in the case of FPGAs design tools) to support a manufacturer’s designs. Intel does not have any obvious advantage in design or software. Like all semiconductor design companies, it has a history of many failures (though of course Intel has had successes).
What Intel could be doing to protect its moat
Perhaps an alternate strategy for Intel would be to become a contract manufacturer for other semiconductor design companies. This would enable Intel to have a higher market share of semiconductor manufacturing. It would put Intel in direct competition with other contract fab companies such as TSMC (which is the largest contract fab). Intel currently does some contract fab work for some startup FPGA companies (who have almost no market share) but otherwise has a very insignificant contract fab business. Nvidia’s CEO suggests that Intel should use its fabs to build chips for other companies:
“Why not be a foundry for all the mobile companies?” Huang told a group of reporters Wednesday evening. “There’s no shame in that.”
Intel’s response: thanks but no thanks.
“We have a small nascent foundry business, but our focus with our SOCs (systems on a chip) is really on Intel based platforms,” said Intel spokesman Jon Carvill. “Our process technology is a huge advantage going forward in 2012 and 2013, so our focus at this time is on building Intel products, not on building products for our competitors.”
It seems to me that Intel is being ambitious (greedy?) as it wants to use its manufacturing advantage to help its SoC design business gain a toehold in smartphones and eventually gain dominant market share. If the x86 instruction set does eventually dominate smartphones and tablets, then Intel’s design costs would be lower as it can use variations on designs for its other low-power processor products (i.e. the Atom line used in netbooks and laptops and embedded applications). Whereas if Intel lets the ARM ecosystem proliferate, it cannot do this. I’m not sure if Intel’s decision is a great one since the computer industry has been trending away from vertical integration and towards companies that dominate a horizontal segment. Intel may be better off focusing on its semiconductor manufacturing business where it is strongest. Intel’s design business is decent (enough) but it often does not win. There is a small chance that tablets and smartphones could fuel the growth for a large non-Intel contract manufacturing base (though currently the smartphone SoCs market is a fraction of Intel’s revenues). If that non-Intel manufacturing base has technology comparable to Intel, then suddenly AMD can become a very dangerous competitor to Intel.
Then again, maybe Intel is confident that it will dominate the smartphone and tablet markets. So far Intel’s smartphones are competitive according to the reviews (roughly middle-of-the-pack in power consumption and performance) so Intel is on the right track. The situation may change in the future as Intel starts shifting its smartphone SoCs onto its best fabs, as so far it has been using its older technology (32nm) to build its smartphone SoCs. Intel’s current part (Z2460 Medfield) is built on a 32nm process while competitors are on 32nm (e.g. the iPhone 5 SoC) and 28nm (e.g. Qualcomm’s Snapdragon S4). Next year, Intel will migrate its smartphone SoCs to 22nm while the rest of the industry will still be on 28nm. Intel is currently shipping 22nm parts with finFET (and high-k/metal gate) technology… it may be 2/3 years before TSMC catches up to Intel. On the software side, the future will get better for Intel. Intel-based Android smartphones are very slow at a small number of games as they have to emulate the ARM instruction set. This will go away in the future as software developers compile an x86 version for Intel processors to avoid emulation (the developers will need to spend some time testing their software on an x86 phone… but they will likely do that anyways).
*Disclosure: Long Intel.