Reversal of fortune in the electronics supply chain
Image credit: Science Photo Library
Semiconductor availability in 2022 has been dire. Delving into the historical data reveals it is both predictable and short-term, but will it change the nature of global supply and demand?
The supply-chain crunch of the past year quickly became so bad that some firms even bought other companies’ appliances to strip them for the specialised semiconductor components they contained. It was a situation the industry had not experienced in a quarter of a century.
The years of sluggish economic growth since the 2008 crisis lured many into the false sense of security that the chipmaking sector could pretty much make whatever customers wanted. But, on the way out of the sudden Covid-lockdown slump, buyers were surprised that the industry was not in a position to fill gaps in their inventories. For long-term observers of the market, the supply crunch was not unprecedented but the resumption of normal behaviour.
“Semiconductors was, is and will be a cyclical industry. At the same time, the industry is kind of bad at predicting things,” Charles Shi, senior analyst at financier Needham & Company, explained at the Design Automation Conference in July.
By that point, Shi was looking at a situation where supply does not just catch up with demand but overtakes it. And it would likely come at a point where fab capacity built by TSMC, Intel and other large companies finally comes onstream and adds even more supply to what exists, creating the potential for a glut. The clues came just a month earlier.
Bill McClean, president of IC Insights, points out that the sales trajectory for the second half of this year looks as though “someone flipped a switch to the off position for the memory market, beginning in June”.
Change has spread across the industry since, according to Malcolm Penn, president of analyst firm Future Horizons. Using figures from the World Semiconductor Trade Statistics group, he argued at the start of August: “The 17th market downturn has now well and truly started.”
Penn pointed to projections from GlobalFoundries, number-four foundry worldwide. The company warned it expects to see its capacity utilisation – the degree to which its equipment is being used for orders and not left idle – fall in the second half of the year, not least because many suppliers have been frantically putting in new production lines to cope with the demand spike of 2021. It is a similar pattern at Samsung, though the leading foundry TSMC is currently reporting a stronger order book. However, the Taiwanese supplier has a higher proportion of leading-edge production than its competitors and those lines rarely run far below capacity even during major slumps.
“Reduced unit shipments will quickly translate into a sharp cutback in new orders, taking the pressure off fab capacity and eventually shorter lead times, just as the first wave of increased capacity is coming online,” Penn notes.
The question is, once this downturn passes, what will be the long-term effect of the recent supply crunch on attitudes among electronics buyers when you couple those effects with an expected period of deglobalisation? Globalisation helped suppress prices. From the boom of 1995 until late last decade, average selling prices (ASPs) of silicon bumped down the curve from a peak of $3 to just $1, seemingly fulfilling another of Gordon Moore’s laws.
“No-one liked this one, so it’s nowhere near as famous,” says Penn. “But he said the long-term trend for ASPs is a dollar. Though ASPs rose sharply in the past boom year they have already given up 70 per cent of those gains: wiped out in a few months.”
In an interview with the Brookings Institute earlier this year, Morris Chang, the former chairman of TSMC, described how globalisation not only made that business possible but drove down overall costs. At the point where Chang took over as chairman of Taiwan’s Industrial Technology Research Institute (ITRI) at the end of the 1980s, the island was well behind Japan and the US in semiconductor technology. Another weakness was a lack of design expertise on the island. “But there was one big strength: our yield was very good. By being a foundry, we kind of skipped over the weaknesses and used our strength. A foundry did not need strength in design. The customers do the design and just let the foundry manufacture.”
Chang founded TSMC as a way of capitalising on that imbalance in strengths, recognising that chip suppliers would seize on cheaper production, often no matter where it is. Over the following 30 years, TSMC gradually moved up the technology ladder to the point where, after Intel stumbled at the end of the last decade, it became the de facto leader in advanced silicon. Though South Korean manufacturer Samsung has maintained pace with TSMC, most of the customers for leading-edge silicon tend to pick the Taiwanese option first, partly because the company has a reputation for delivering. A second reason is that they are not competing directly with a pure-play foundry: a factor that makes Intel’s and Samsung’s foundry offerings less attractive. That sense of security is a vital characteristic of an industry that remains paranoid about its intellectual property (IP) even though most companies engaged in making chips need to trust multiple contract suppliers.
Fears over IP security and other issues have restricted China’s abilities to follow Taiwan into the semiconductor elite. Though SMIC has demonstrated it can deliver a process technology similar to TSMC’s four-year-old 7nm node, the Big Fund that financed other chipmakers and technology companies has been mired in allegations of executive corruption. In the meantime, starting in 2016, sanctions on China have steadily been ramping up.
In the wake of Beijing’s decision to conduct military exercises off the coast of Taiwan, the US has stepped up its attempts to choke off supply of technology to China in a variety of ways. The Biden administration not only invoked long-standing legislation on advanced devices that the government deems can be used in weaponry by banning sales of Nvidia GPUs but sought to clamp down on China’s ability to make its own equivalents, probably restricting SMIC’s ability to increase capacity for its advanced processes.
The recently passed Chips Act and the parallel act drafted by the European Union seek to bring wafer manufacturing back onshore and reduce the dependency of western countries on Far Eastern factories. But Chang says the US and countries like it have different attributes from those exploited by Taiwan three decades ago. “There’s a lack of manufacturing talents to begin with. I don’t really think it’s a bad thing for the US, actually. But it’s a bad thing for trying to do semiconductor manufacturing in the US.”
Chang uses the example of a little-known, relatively small fab that TSMC has operated on the border between Oregon and Washington since the late 1990s. Over that time, a cost difference between production in the US and Taiwan has remained more or less constant. “The same product, the Oregon cost, is about 50 per cent more than the Taiwan cost. The Oregon product is still profitable, although not nearly as profitable as the Taiwan product,” Chang claims.
Other chipmakers have claimed greater success in transplanting fab experience. Whereas TSMC has focused on building gigafabs on its home territory, Intel has over the years taken advantage of inward-investment initiatives around the world, putting fabs into Ireland and Israel and soon Germany, coupled with a factory-management strategy called Copy Exactly that it introduced 25 years ago. As its name suggests, the policy benchmarks factories against each other and anything that reduces cost or improves yield in one is translated to the others wherever possible.
However, Penn points out a number of factors make it hard to transplant complex factories like these into new locations. “It’s partly because of infrastructure, it’s partly because the smaller scale these projects have and also the learning curve. In the interim, they will be more expensive to operate. That’s what’s so attractive about the whole consolidation we have seen in the past.”
However, as with any technology-focused industry, clustering is a major factor in determining success for a region. Once an area becomes known as a centre, it attracts better-quality talent, not least because those people can move easily between employers. They are far less likely to relocate to another continent without significant incentives. With greater use of reshoring, the dominant clusters can be expected to reduce in size compared to a collection of smaller regional clusters that need to compete for a finite talent pool. If you add the geopolitical factor, in which local champions become more important to nearby customers who can no longer buy from a global leader, the tendency will be towards more duplication of effort.
Boston Consulting Group has estimated that regionalising the semiconductor sector so that each was responsible for investment in progress would add a trillion dollars to the cost of R&D globally and make the resulting chips approximately 1.5x more expensive to make. Whether there is a full regionalisation depends on the geopolitical landscape. It could easily take a decade to distribute the production centred on Taiwan to other destinations. And there is currently little appetite among the CEOs of electronics companies to support the transition unless they have military customers. However, in the face of growing international competition where governments will favour locally owned business, they may become significantly more paranoid. That could lead to deeper changes in the way companies approach their supply chain.
Management consultants have not been slow to argue that customers need to alter how they buy from suppliers and avoid being caught out again by sudden shortages (see box). But Penn doubts that the industry will see long-term changes. “Human nature says it will revert to normal once this all passes. Building deep relationships in that way can’t practically be done because of the nature of the industry and the way in which so much is outsourced,” he says.
Because many components filter through a pipeline of wafer fabs, such as those operated by TSMC, and packaging plants dotted mostly around South Asia, and then into a network of distributors, visibility for most customers is limited. Only the largest and those who insist on custom silicon get a chance to build those deeper relationships.
Similarly, the focus on local production is also likely to be limited. Much of the funding for local fab production in the US has been driven by concerns about securing access to complex silicon for its defence industry. The issues Russia now faces in trying to obtain production of even fairly basic semiconductors to replace the systems used in its invasion of Ukraine have underlined the strategic importance of guaranteeing supplies. However, once the defence interests are taken care of, the likelihood of production moving away from Taiwan and Asia is low. “And when the market does go into a downturn, watch the politicians run,” Penn notes. “Why invest when you can buy more capacity, more cheaply elsewhere? There is not a single end customer that really wants [reshoring], apart from the DoD.
“However, this doesn’t change the fact that there has been too much concentration in one area. Pressuring TSMC to build factories elsewhere was overdue, though you can understand why Chang was so against it. From a geopolitical point of view, we need a degree of diversity. And not just for that reason. It’s also in an earthquake zone, and earthquakes in the region are not occasional events. But, having said that, TSMC did such an outstanding job that was never really seen as a problem.”
In the wake of China’s increasingly aggressive stance on reunification with Taiwan, pressure may ramp up again on geographic diversity for wafer fabs and for packaging. But history suggests that, short of moves to a full invasion, the pressure will dissipate as costs become more important.
The semiconductor industry presents unusual challenges to supply-chain management, which make it harder to predict when shortages will hit.
The core problem is time. It takes months for an individual chip to make it from a wafer being sliced from a silicon ingot to being packed into a tube or mounted on a reel ready to be shipped to an assembly plant. The customers’ orders will be shifting at a different rate, creating spikes and slumps in demand to which chipmakers cannot easily react. In the wake of the dot-com slump some suppliers diverted unfinished wafers to nitrogen storage at a point in manufacture where the chips onboard would not degrade before being returned to the fab to be finished. They have even fewer options when faced with a peak in demand: whatever is ordered cannot turn up in less than three months even with dedicated production.
In an environment like this, the obvious answer for most industries is to diversify supply – but that is not easy to achieve in electronics where often parts that fit a particular design only come from one chipmaker. Multiple-sourcing of complex devices was common up to the end of the 1980s but quickly died as the primary suppliers decided they were better off keeping designs to themselves.
“Understanding semiconductor constraints is key to solving them. In many cases, the issues are multiple tiers up in the supply chain. Therefore, companies should map out the multi-tier supplier network to understand the constraints,” says Koen Jacobs, vice president of industry solutions at o9 Solutions, which provides software that lets users build digital twins of their supply-chain environment.
In situations where key devices become unobtainable, Jacobs adds that the digital twin can be used for scenario planning to understand vulnerabilities in the supply chain. “Organisations can ask: what are the critical components? Are they single sourced, geographical risks, or something else? Based on those insights, companies can then act to reduce risk: this could be by finding alternative suppliers, or re-engineering products.”
It is feasible to swap out a microcontroller with an Arm core from one vendor for another but it is not a painless or quick process even if the software is designed to be portable across both or the PCBs designed to take either part. Doing this for highly customised system-on-chip devices, finding alternatives is often technically and commercially impractical: the only option is to make sure you are first in the queue for capacity.
The answer, at least for larger customers, is tighter integration with the primary chipmakers and fabs. This has worked well in the past, though it seems to be an easily forgotten lesson. The recent supply problems faced by carmakers were largely driven by their decision to cancel orders suddenly and try to restart them as the worst of the pandemic lockdowns passed.
In the 1995 supply crunch, investors encouraged the larger fabless companies to take part-shares in fabs. Xilinx, now part of AMD, bought into one of these consortiums organised by UMC, then its prime foundry. But the firm’s head of Europe argued at the time that the decision was driven mostly by the finance community. Xilinx’s policy of committing to long-term supply deals proved effective enough. It still works today.
“To address the shortage of capacity, the most important point is to communicate your forecast well in advance of your demand. This allows all partners to plan their capacity in the most efficient way,” says Gareth Jones, vice president of ASIC at design house Sondrel.
Sign up to the E&T News e-mail to get great stories like this delivered to your inbox every day.