Semiconductor Series, Part II : Foundries
Refer back to Part I: Raw Materials and Wafer Fabrication Equipment (WFE) for a quick refresher on the players at the very beginning of the semiconductor food chain. Part II of the semiconductor series is going to be about the companies and the factories that make them.
In the olden days, those who designed the chips were also the ones who made them: the fabled IDM model. Since then, industry titans like AMD have either sold or spun-off their foundries, and other players like Texas Instruments and NXP have elected to go with “fab-lite” models. The foundry and chip designer spectrum is largely divided into 3 sections: pure-play foundry, fab-lite, and fabless. Some quick housekeeping, the word foundry and fab-short for fabrication plant- are largely interchangeable, and the word IDM stands for Integrated Device Manufacturer, meaning a company that designs and fabs their own chips.
Starting from the left hand-side with pure-play foundry companies, Taiwan Semiconductor Manufacturing Company (TMSC) is the objective leader of the pack. 90% of the world’s advanced chips come out of TSMC. (Source) The chips that go into phones, laptops, VR headsets, datacenters, supercomputers etc. are likely made in a TSMC fab. These fabs are composed of the latest and greatest wafer fabrication equipment (WFE) made by the big 5: Applied Materials, ASML, KLA-Tencor, Lam Research and Tokyo Electron. The job of the foundry, TSMC, is to combine that magical WFE into a functioning assembly line that can churn out the latest chips.
This is an extremely challenging process that even the equipment suppliers themselves do not know how to handle. The foundry is like the structural engineer of a 100+ story skyscraper and the conductor of a 1000 piece orchestra. It must find out how to architect transistors and interconnects such that speed, efficiency, and thermals are optimized, while also tuning every piece of equipment precisely so it operates within a tolerance measured by nanometers.
This is why the job of the foundry is so vital to both its customers, and the nations upon which its facilities lie upon. There are currently only 3 countries with foundries considered bleeding-edge operated by domestic players: Taiwan, Korea, and the United States. Taiwan has TSMC, Korea has Samsung, and America has Intel. These 3 companies all have the capability to manufacture what would be considered the most advanced chips.
Each generation, or “node,” is marked with a name defined in nanometers. The latest node to enter mass production has been the 3nm node. 3nm doesn’t actually mean each transistor is 3nm large, it is just the name of a single generation of node. Similar to how the iPhone goes from iPhone 14 to 15, nodes will jump from 5nm to 3nm and aren’t an indication of transistor size.
Aside from the big 3 foundries TSMC, Samsung, and Intel, the manufacturing capability of competitors drops drastically. The 4th most competitive foundry is China’s SMIC, Semiconductor Manufacturing International Corporation, which is technologically years behind the frontier. To put it into perspective, SMIC’s 7nm generation node is their latest actual node to enter mass production. Meanwhile, TSMC’s 7nm node entered mass production nearly 6 years earlier. To quantify that in terms of performance, we can compare the performance of smartphone chips. Apple’s most recent iPhone chip on TSMC’s 3nm process is >2x as fast as their last iPhone chip on 7nm.
Smartphone CPU performance is likely to be a more accurate representation of performance improvements solely attributable to process node improvements, as smartphone chip design is in a far more mature state than GPU design-where architectural and algorithmic improvements are increasing performance 2x every few months. That being said, using the smartphone chip benchmark, a given chip based on TSMC 3nm is likely to be either twice as fast, or twice as power efficient as a given chip produced from the latest SMIC fab.
At a datacenter level, TSMC process node customers can build out twice the compute capability for the same price, power envelope, and real estate footprint. This advantage is absolutely crucial in a time where the artificial intelligence industry is being driven by logarithmic scaling laws, and where the size of datacenters is being capped by local power draw limitations. This means that you must increase compute resources exponentially for linear increases in model performance, but you only have a finite allotment of power your datacenter can draw. In the age of AI, power efficiency doesn’t just matter from a TCO point of view, but it is now the main driving force behind how much compute we can throw into training a model. These power limitations imposed by the local utility are the reason why big tech companies have opted to go the private route and invest in or partner with small modular nuclear reactor companies-another topic we have covered in the past!
Shifting back to foundry, it is evident that there is a large difference between a great foundry and a good one. Major supply chains like that of the iPhone or a Hopper GPU are entirely reliant on fabs in Taiwan for manufacturing their chips. Being a country with cutting edge fabs on your territory is crucial from a national security perspective because it insulates your domestic innovation from global conflicts. If Taiwan’s facilities are compromised by an earthquake or a military operation, US chip designers like Nvidia can still turn to TSMC’s USA facilities, or even Intel to continue to pump out competitive chips.
Moving onto fab-lite, there are companies like Texas Instruments (now less so), NXP Semi, and Infineon that employ this business model. Keeping on the bleeding edge of semiconductor fabrication is challenging from both a human and financial capital lens. Developing the next generation of process node is getting increasingly expensive as TSMC’s 5nm node cost twice as much to develop as their 7nm node did. (Source) Without mentioning R&D, the CapEx bill can get into the $20-30 billion mark for just one facility. This is why, for companies whose core business does not revolve around cutting edge nodes, it makes sense to go fab-lite. These companies can still be considered IDMs because they still design and manufacture some of their own chips to an extent. But, any chip they sell that is designed on an advanced node will be manufactured by a third party like TSMC. The fundamental reason this operating model works is due to how process nodes age. The bulk of Texas Instruments’ products are manufactured on process nodes in the 45nm-130nm node, (Source) process nodes that initially entered mass production 2 decades ago. Even though the next generations of process nodes are getting more expensive to develop, those same process nodes will eventually drop in price and will be able to be made more economically. Over time, fabs can always figure out how to decrease the defect density on a wafer. Here is how it generally trends overtime:
As you can see, fabs consistently improve at mitigating manufacturing defects, which increases yield per wafer, and improves the cost basis. Not only this, but WFE based on previous gen technology is far cheaper to buy off equipment suppliers than when they were cutting edge. This means that for a given node, the cost for building a fab, and subsequently producing that node gets cheaper with time. A great example of the fab-lite business model is with Texas Instruments’ purchase of Micron’s Lehi Fab, which produces 65nm and 45nm chips. They bought the fab for just $900 million in 2021, while it likely cost roughly ~$4 billion to build. This $4 billion figure comes from the plant’s initial $2.5 billion construction (Source), plus an additional $1.5 billion investment in the early 2010s (Source). This is why IDMs whose products do not need to be manufactured on the latest process node prefer the fab-lite model. With fab-lite companies, it’s not mission critical to be on the latest node. They can save themselves the capex and outsource to TSMC while they wait for the technology to inevitably become cheap enough to invest in.
This fab-lite model is the strategy that Texas Instruments, NXP Semi, Infineon, STMicro, and many others have committed to. Their products are not in the latest smartphones or the newest Nvidia GPUs, but instead inside your cars, scooters, internet routers, refrigerators, laundry machines, and electrical equipment. These applications don’t need the most power efficient or speedy chip on the market! When was the last time you felt the user interface on your microwave wasn’t snappy enough?
