GlobalFoundries Fab Expansion: Doing A Little Math
GlobalFoundries (GF) hosted U.S. Senate Majority Leader Charles Schumer and Commerce Secretary Gina Raimondo in Malta, NY for a major capacity expansion announcement yesterday. The company’s Fab 8 in Malta currently is the lead factory for producing its most advanced CMOS logic chips. CEO Tom Caulfield announced a $1 billion plan to increase capacity 43% within the current footprint, as well as build a second fab on adjacent land that would double their available cleanroom space, or “maybe a little bigger … roughly the space of a megafab,” in his words. Let’s do a little math and see what that tells us about semiconductor chip manufacturing.
The capacity expansion
GF has fabs spread around the world, as its name suggests, with facilities in Singapore, Dresden, Germany, Burlington, VT, and Malta. Malta produces 12 and 14 nm FinFET devices on 300 mm wafers (12 inches in diameter), with each wafer holding anywhere from a few hundred large chips like microprocessors to thousands of smaller chips. Fab 8’s current capacity is 350,000 wafer starts a year, and to get to that output rate the company has to date invested $15 billion in site acquisition, facilities construction, and fit-up with the specialized tools and materials handling that chip manufacturing requires.
The company indicated that $1 billion in new investment will buy an additional 150,000 wafer starts a year of capacity. To get a 43% increase in capacity with a 7% incremental investment represents huge leverage, and it highlights several interesting aspects of the scale economies in chip manufacturing:
There is a really high fixed cost in building such facilities. This comes from a combination of the need for extraordinary stability and freedom from vibrations (which means a lot of concrete and steel are used in construction), and a vast infrastructure for providing ultraclean air, water, and a complex range of chemicals. But they didn’t spent the majority of the $15 billion on the building, that was a small part of the overall cost.
The specialized tools inside the cleanroom are the major expense, with some – like the lithography tools – costing in excess of $100 million each, and you need multiples of them. Typically in a fab, you ensure that the bottleneck is at the most expensive tools; it’s okay to run the other tools at less than 100% utilization, and the spare capacity in those tools is where the leverage comes from.
What this tells us is that a 500,000 wafer starts per year (41.6 K per month) fab is much more capital efficient than a 350,000 wafer start per year fab. Of course the exact size will depend on the chip recipes and tools used to support those recipes. It’s not uncommon to have fabs run at 70,000 – 90,000 wafer starts per month. TSMC’s “Gigafabs” run at around 250,000 wafer starts per month, but those fabs are partitioned into phases. But this gives us a sense of economies of scale in semiconductor manufacturing.
MORE FOR YOU
The other thing that the GF announcement suggests is that the incremental capacity will be relatively quick to bolt on, and will mainly depend on how long it takes to receive the additional tools. In view of the current high demand for such tools, it may still take a while, but at least they don’t have to wait until the second fab is built next door. You could say that was great capacity planning in view of the current shortage, or if you were the CFO you might have bemoaned that capacity planning in the years you were underutilized. Timing is everything.
Rumors about Intel
The first question asked at the event concerned a Wall Street Journalreport last week that Intel was in discussions with GF’s owner, Abu Dhabi-based Mubadala Investment Co. to purchase the company. Under its new CEO Pat Gelsinger, Intel has stated its intention to ramp up its foundry business. While Intel has tried to enter the foundry business before, it is quite a different business model when you make chips for others than when you make them for yourself. Foundries like Taiwan Semiconductor Manufacturing Company (TSMC), Samsung, United Microelectronics Corp. (UMC), and GF have ecosystems of software tool suppliers, providers of intellectual property (IP) blocks, and many other complimentary asset and service providers. One of the challenges facing Intel is that these ecosystems take time to establish, and buying GF would speed up that process. Needless to say, Caulfield artfully evaded the question, but it is interesting to think about it.
Ironically, GF started as the manufacturing arm of Intel arch-rival AMD. After AMD spun-off GF in 2009, the company had to learn how to become a foundry itself. Of course, buying GF means Intel would have to get comfortable supporting AMD’s manufacturing needs. GF recently signed an agreement to supply AMD with $1.6 billion of wafers, using its 12 and 14 nm processes to make dies for AMD server chips. But that’s what being a foundry takes, a desire to make your customers successful, even if they are your competitor.
My students usually find discussing strategy hurts the lead less than doing the math of capacity planning. This might be a case where the opposite holds.