LONDON – The world's leading foundry chip maker Taiwan Semiconductor Manufacturing Co. Ltd. is considering operating single-customer wafer fabs, according to chairman and CEO Morris Chang.
Chang, speaking to analysts on a conference call to discuss the company's second quarter financial results, said that the market is tending to produce fewer higher volume customers and some are so large they need their own dedicated fabs. This is despite the fact that, as a foundry TSMC, has risen on its ability to serve many customers from a single line.
"I think that they are going to be larger customers, and now it makes complete sense to dedicate a whole fab to just one customer and hold that – to hold fabs in fact to just one customer," Chang was reported as saying in a Seeking Alpha transcript of the question and answer session in the conference call.
Chang said TSMC would retain the ability to serve many customers but the presence of large customers that are getting bigger means "it makes sense that we dedicate a whole fab or even more than a whole fab to just one customer."
Qualcomm is one such very large customer. It has had problems in recent months getting as much supply of 28-nm CMOS from TSMC as it would like.
Chang did not mention Qualcomm explicitly but said that Taichung, where TSMC has its Fab 15, will be the source of the majority of TSMC's 28-nm CMOS whereas Tainan, home to Fab 14, will be the source of the majority of 20-nm planar CMOS and 16-nm FinFET CMOS.
For older technologies, it was easy to keep a 2nd source but for advanced nodes from 28, it does not make sense for many fabless companies to keep a 2nd source due to economical reason. Cost of engineering resources and NRE will easily add up to multi-million dollars per project.
I think that Samsung's Austin FAB makes chips for Apple. I have heard about use it or lose it, but customers doesn't like the idea. They don't want to put down the capital. Specially when they can choose between TSMC, GF, UMC & Samsung.
It always makes sense to keep a 2nd source.
I think this is not new idea. Back in 2011 it was widely rumored that samsung was building apple's chips at a dedicated fab in the US, so a single customer fab may already exist, and trying to copy samsung may be what put the idea into Morris Chang head.
Like in all business, TSMC needs to plan its capacity according to customers' and its own forecast. The good thing about foundry model is each individual customer does not need to forecast accurately as long as the foundry can forecast a rough aggregate demand to plan for its expansion. In this model, customers don't need to risk any capital. The foundry is taking all the financial risk for building the fab because it believes (with some customer inputs) the demand will be there when the fab is on line 18 months after fab construction.
Qualcomm complains TSMC didn't have enough capacity when it couldn't forecast the demand of its own business.
TSMC of course can commit a whole foundry to one customer as long as that customer commits up front to buy, say 90% of the capacity, lots of wafers over 3 or more years. Or, the customer can chip in part or all of the fab cost. The fab will only use the fab equipments for its own designs, instead of meeting the needs for 20 other customers. The benefit is better process optimization of the processes, lower cost of the fab, and faster to production.
The whole idea is simple. The issue is if there will be any customer willing to commit up front for TSMC to remove that fab the chance to make any profit from other customers.
Foundries have been doing this for eons. The process(es) which they offered to big customers are optimized to their so called spec. Adjustments gets done everyday and every products. so this is no new news.
Exactly. Agree with your thoughts here, grzy. This, to me, is almost going back to the future. If a big chip company like Qualcomm wants its own fab, they should build their own fab. Period.
But by putting this idea into Morris Chang's head, they are making this TSMC's problem. (fab utilization fluctuation, optimization/customization of a process techonlogy, etc.)
I agree with peter.clarke. Whow knows who forced and who gave Morris Chang such idea to think that single customer fab would be benefitial. I strongly believe that Apple and/or Qualcomm had their word in it.
Other thing to note is that TMSC runs this monopoly of the chip manufacturing pretty much, so they can do whatever they want. Whow are they going to lose their business from?
Of course it may be that Apple is saying a dedicated fab along the lines mentioned by DMcCunney is part of the terms of conditions of making Apple processors.
And that is what is putting some of these thoughts to top-of-mind with Morris Chang.
Whether it make sense depends upon the details of the deal.
For the customer, the question is "Can we get enough components to meet our production requirements?" No one wants to leave money on the table because they simply can't build as many devices as their customers want to buy.
For TMSC, the question is "Can we in fact meet the customer's needs?" Depending upon the volume the customer requires, the answer may be "Yes, if we dedicate a fab to them."
The question is what "dedicated" means. Current leading edge fabs are fantastically expensive, to the point where very few companies have their own. And the only way you make money is to have the fab producing in volume all the time.
So if I'm TMSC, I'm deciding whether to dedicate a fab, and which fab to dedicate, based on the customer order. It would have to be a hard commitment by someone like Apple to get me to do it. And even then, "dedicated" in practice will mean "The customer has first dibs on capacity, and the fab will meet their orders before it does anything else. But if the fab has any leftover capacity *after* filling the customer's order, it will be allocated to other production. It is too expensive to have any capacity idle."
I think a joint venture where a major customer like Apple puts up a chunk of the financing needed to *build* a fab is a possibility, if volume demands are great enough a new fab has to be built. But initially, I see this as simply a way of allocating existing capacity to reassure major customers that they'll get what they order. If I were TMSC, I'd think very hard about whether I'd do a joint venture to build new capacity, and leave myself that indebted to that customer. If the volume I'm handling is great enough I *need* to expand, I might decide I could do so on my own.
David Patterson, known for his pioneering research that led to RAID, clusters and more, is part of a team at UC Berkeley that recently made its RISC-V processor architecture an open source hardware offering. We talk with Patterson and one of his colleagues behind the effort about the opportunities they see, what new kinds of designs they hope to enable and what it means for today’s commercial processor giants such as Intel, ARM and Imagination Technologies.