There have been some important developers in the semiconductor market this week that have unsettled the semiconductor world. The news was that Global Foundries has exited the market for 7nm production. This may seem a small matter, a company rationalizing its product roadmap, but when thought through, it leads to some unsettling consequences. This news has implications far beyond the world of semiconductors, it will ripple through all of electronics, then all tech, and we would argue is part of a bigger geopolitical issue. Small press release, big News.
Let’s start with the basics. Most ‘chip’ companies do not manufacture their own products. Companies like Qualcomm or AMD or Silicon Labs design chips, but they do now own ‘fabs’, which is what we call semiconductor factories. Once upon a time all chip companies owned their own fabs. The saying was (and we literally heard people say this), “Real Men own fabs”. That manufacturing philosophy now sounds just as awkward and misguided as the sexist tone of the quote. Starting about 20 years ago, the cost of building a new fab got so expensive that most chip companies moved to a ‘fabless’ model. Chip design companies outsourced their manufacturing to third parties, known as ‘foundries’.
This model worked incredibly well for years. It freed the fabless companies from the capex required to build fabs and instead focus on core competencies around their own specialized products. However, the pressures on the foundries remained intense. It became an endless arms race to keep advancing manufacturing technology. The common wisdom on this is what we call “Moore’s Law”, which is the notion that semiconductor capacity doubles every 18 months, but usually skipped over was the corollary that the cost of those fabs also doubles. For instance, the largest foundry TSMC has had several years where their capex exceeded their revenue. The result was a gradual winnowing of foundries. There are still hundreds of fabs out there, but there are essentially only four companies left today that can produce chips using the most advanced manufacturing technology – the 10nm process. These companies are Taiwan Semi (TSMC), Intel, Samsung, and Global Foundries.
The reality is Moore’s Law has been slowing down for years. That 18 months we mentioned above is probably closer to three years, and the 10nm process has proven to be a bear. Intel appears to be stuck getting their 10nm process to work, hence the delay we talked about in our earlier post (and here). And the next node – 7nm – looks to bring even more headaches.
This brings us to Global Foundries. They announced that they are discontinuing their 7nm process. And since every node builds on the one before, this means they are giving up on 5nm and are essentially out of the race. This is a smart move for GloFo, or at least the smartest move given their limited options. They have been losing money for years, largely due to the cost of competing in this Arms Race for process technology. They can now focus on profitably supplying a range of specialty products using trailing edge (older, larger) nodes. But make no mistake, this move was driven by necessity and it will lead to a drastic reduction in their strategic options over the long term.
Intel is stuck at 10nm, and will be for at least a few years. Samsung has mastered 10nm, but there are some serious concerns about their timeline (and maybe their ability) for getting to 7nm. And when we say concerns, we mean they have lost two of their largest external customers because of these concerns. This leaves TSMC as the only company capable of producing 7nm chips currently, and they are likely to maintain that position well into the 2020’s.
Impacts on the Tech Industry
For the past forty years, the technology industry has been built on the back of (the good part of) Moore’s Law. Computers and phones and networks kept getting faster and better performing. This now looks set to change. It is clear that we are going to get stuck on 7nm for a long time. TSMC has no incentive to rush beyond 7nm. Effective monopoly profits must sound good to a company that had all those capex bills we discussed above.
To be fair, there are still a lot of ways that chip designers can improve their products without moving a node ahead. For the next few years, they will focus on ‘architecture’ improvements, winnowing out unnecessary parts of their earlier designs. Expect to hear chip designers talk a lot about how they are using ‘Machine Learning on their chips, for instance. We can also expect to see more special purpose products, chips that only perform a few special-purpose features like Neural Networks as opposed to general purpose processors. There are many ways for chip companies to push performance. But it is also seems likely, that performance improvements for many things will slow down.
One theory we are working on now is the impact of ancillary products – memory, networking, etc. – used in the data center. Sales of these products are showing some signs of slowing as the large data center builders delay their upgrades of installed compute capacity, and thus need less of the adjacent products. Data center demand continues to grow as the cloud service providers build new facilities, but if CPUs advance more slowly, the cloud providers have little incentive to replace older products that sit on the same boards as those good-enough, fully depreciated CPUs.
Impacts on the Foundry Business
Over the next few years, the slowdown in semiconductor advances will gradually become more apparent, and this will put new pressure on the foundry model. We are reaching an important impasse. If this were an industry less central to the economy, conditions would eventually lead to consolidation and stabilize in some sort of low-growth status quo. However, it is clear that semiconductor technology has become an important battleground in trade politics.
China is acutely aware of its reliance on US chip companies to power its export economy. Some estimates have it that the Apple supply chain alone employs something like seven million people in China. Add in phones and PCs and all the rest of consumer electroinics, there are a lot workers to tied to these supply chains. All of these chips are assembled into working products inside China’s immense manufacturing capacity, but they are all designed by non-Chinese companies. The government there has been overtly subsiding the growth of a domestic semiconductor business. But this has shown only mixed results. While there a small number of promising chip design companies in China, we can say with a high degree of confidence, that Chinese chip design capacity is still very far behind the US. Moreover, there is no Chinese foundry at anywhere near the leading edge. Shanghai-based SMIC is the leader, and they are still struggling several nodes back.
Again, if this were a small industry, Chinese companies would be a major force for consolidation. However, the US government has made it clear for years that it will not allow Chinese capital to buy US chip companies. And those policies were in place long before the current administration. The mounting trade war only emphasizes them.
It seems likely that the US government is becoming very uncomfortable with the industry’s growing reliance on a single foundry. TSMC is based in Taiwan, and operates some older technology fabs in China. TSMC is now essentially Finland, domiciled in one political-industrial sphere but tightly coupled to the competing sphere. Assuming Intel can motivate itself to solve its process technology problems (by no means guaranteed), it has never been a player in the third party foundry business. They really do not have any external customers (see our previous note). So it is not clear that they can become a leading edge supplier to the broader industry. Samsung, which as noted, has its own process concerns, is in a similar position to TSMC, straddling two worlds.
Further complicating matters is Global Foundries’ ownership. If it is not clear now, it will become clear soon, that Global Foundries cannot exist indefinitely as a standalone entity. Its growth is dependent on finding niche, specialty products to produce. There are definitely markets for this, but by removing their future technology roadmap, they are capping their growth. Further, it is easier for other foundries to enter these specialty markets, which over time will lead to revenue and profit headwinds. At heart, the foundry business is immensely capital intensive, and GloFo will find it increasingly hard to maintain that investment from internal cashflows. Its owners have shown they are unwilling to subsidize investments indefinitely. GloFo is owned by the sovereign wealth fund of Abu Dhabi, and a big part of the motivation for them quitting 7nm was the estimated $10 billion to $20 billion the company needed and Abu Dhabi was unwell to provide. However, GloFo has one other wrinkle to consider. Several years ago they acquired IBM’s semis fabs. Crucially, this included a “Secure Fab” which is used to produce chips for the US Military. Part of the acquisition agreement included provisions that kept control of this unit under a separate Board, whose members are approved by the US Government. In a normal trade environment, this business could not be sold to a foreign owner, and in current conditions that holds double.
The two Asian foundries (TSMC and Samsung) cannot be sold, nor could they buy either of the two “US” foundries (Intel and Global Foundries). It is unlikely that Abu Dhabi would be willing to shell out billions to buy any of the other players. Which leaves Intel as the only natural buyer, and they currently have no CEO, and troubles of their own. Anyone else who wants to enter the foundry business has to be willing to not only buy Global Foundries, but invest a lot of money to re-start its leading edge processes. There is always speculation that Apple, Broadcom or Qualcomm is looking at GloFo, but this would mark a drastic shift for those companies, and each of them has their own individual strategic mis-fit issues.
In short, the industry is stuck at an impasse.