wildcat.english

translated from: Wildcat no. 110, autumn 2022

»The Summer of Semis«

Because so much had happened, we wanted to continue the interview on the "chip crisis" (Wildcat 108) and met twice in mid-July to do so. Then the events came thick and fast! We updated the already finished article several times - and froze it at the status of 8/10/22. It's a compelling story about the importance of technological development in a key industry of contemporary capitalism, its limitations, and its devastating consequences, up to and including war. Here's the short version:

After the 2007/8 crisis, Apple had made a highly profitable business model out of having Silicon Valley design and Taiwan chips (TSMC) assembled by a Taiwanese company (Foxconn) in China and sold around the world (even in China, the iPhone ranks #1 with 22 percent market share). Highly paid developers design the product, factory workers in Asia assemble it. The industrial costs are incurred in Asia, and the profits flow to shareholders, mostly American funds. In the medium term, however, the industrial capabilities also accrue in Asia: TSMC guaranteed "technology leadership" and is now the world's third largest semiconductor manufacturer after Intel and Samsung, and by far the largest contract manufacturer (foundry) for fabless companies like AMD, Apple, Qualcomm, Nvidia, etc. The U.S. has lost its leadership in industrial manufacturing of solar panels, batteries ... and just chips to Asia.

Foxconn and TSMC are products of a tremendously fast industrialization of Taiwan. Besides South Korea, it is the only country that has not been caught in the middle income trap, with one-third of its workforce working in industry. Wages there are now reaching U.S. levels, and the island's ecological resources are running out.

China would be only too happy to copy the development model - and President Biden the industrial capabilities. This is another reason why not only the chip industry, but also Taiwan in particular, is becoming a battleground for geopolitical conflict. ²

With "friendshoring," the U.S. wants to block technological development in China. They declare the strength of the previous model, its global production chains and sales channels, to be its weakness ("the Ukraine war has taught us that we must never again become dependent!"). They are pushing the construction of chip factories "in the West." Since a current fab costs at least $15 billion, huge sums must be mobilized. (For comparison, a state of the art car factory for 150,000 to 200,000 cars a year costs a little over a billion).

Meanwhile, the chip industry is slipping into crisis (sharp drop in sales). The "Apple model" is probably exhausted ecologically and in terms of wage development; a new one is not in sight. The geostrategic initiative of the USA would increase the ecological footprint of the chip industry many times over if the planned fabs actually start production. Can this planet even bear so many chips?

The global subsidy race

The EU also wants to catch up in the semiconductor industry. The EU Chips Act ³ aims to "double its share of global production to 20 percent by 2030" with 43 billion euros. That's by volume (one trillion microchips are produced annually worldwide, ten percent of them in the EU). By price, sales in Europe are about 40 billion euros, and more than $600 billion worldwide.

In mid-March, it was announced that Intel plans to build two "Ångström-era" chip factories in Magdeburg. In our last conversation (Wildcat 108, mid-June 2021), you had answered the question "How realistic are new chip fabs in the low nanometer range in the FRG?": "I don't think the EU will put up ten billion - that's the figure Intel gave..." Apparently, not quite that much was needed. The first of the two Fabs is expected to cost $17 billion, about 40 percent of which will come from government pots: 2.7 billion euros had already been budgeted in the 2022 federal budget, and additional EU subsidies of 4.1 billion euros are expected to flow in 2024. What role will Magdeburg play for the European semiconductor industry?

Intel had announced at the beginning of the year that it would invest $88 billion in Europe to reduce its dependence on "Asian chip producers"; for the future, they are already advertising that they will be able to supply "everything from a single source" in the USA and Europe. They already have important development departments in Europe: in Ireland and also in Munich. Thus, Magdeburg is a strategic decision. Intel wants to build a new plant for packaging processes in northern Italy - planned investment five billion, subsidy from the Italian state 40 percent. No one is playing poker as hard for subsidies as Intel. In addition, the state is expected to provide lax approval procedures, cheap electricity - which is also used for green washing - and drinking water at a ridiculous price, as we have seen with Tesla in Brandenburg.

If the EU Chips Act is ratified in time, Intel "hopes" to break ground in Magdeburg in 2023, when volume production would start in 2027 with the "Intel 3" process (supposed to be 5 nanometers); the fab is already prepared for "Intel 20A". TSMC has been supplying 5 nanometer chips since 2019, 3 nanometer production starts in early 2023, 2 nanometer is announced for 2026 - so the backlog is considerable.

Intel also wants to expand its manufacturing capacities in the USA and is planning two fabs in Ohio for $20 billion. This could also turn into 100 billion for a total of eight fabs if enough government dough flows in. On July 28, the US Congress passed the Chips and Science Act to promote chip production in the USA. The total volume of $280 billion includes a 25 percent tax rebate for companies building new facilities or expanding existing ones. Fifty-two billion dollars will go as direct subsidies; more than 20 percent of this, eleven billion dollars, will go to research and workforce development. The bulk will go to building fabs and will benefit Intel in particular. TSMC has also just completed the shell of a 5-nanometer fab in Arizona; from 2024, 20,000 wafers per month are to be produced there. And Samsung is building a 5-nanometer fab in Texas, which is scheduled to go into operation at the end of 2024.

How far behind is Intel technologically?

Intel's roadmap is Intel 7→ Intel 4→ Intel 3→ Intel 20A→ Intel 18A. Intel 3 would be Magdeburg; but to date they are having great difficulty ramping up Intel 4 (supposed to correspond to something between 7 and 5 nanometers). As said, 3 nanometer prototypes are already available at TSMC and Samsung; larger quantities are available in 2023.

Chips with nodes smaller than ten nanometers accounted for only two percent of global production in the first half of 2022, chips of 14 nanometers and above for 95 percent. Where are smaller node sizes necessary at all? And what is this Ångström planned for Magdeburg?

As described in Wildcat 108, "nanometer" is no longer a measurable quantity; already with the 45-nanometer node, the FET gate length decoupled from scaled-down structures. ⁴ But what TSMC and Samsung market as 5 and 3 nanometers, respectively, continues to be the reference size for the respective process advances. The bigger the problems in technological progress become, the more creative the designations, Intel in front: After Intel 3, the "A family" is to come: first Intel 20A, then Intel 18A. The A is supposed to allude to the unit of measurement "Ångström" (1 Å = 0.1 nm!). Paul Alcorn of Tom's Hardware etched, "If you can't beat them, change the name."

Chips with nodes smaller than ten nanometers are used for smartphones, high performance computing (HPC), 6G, autonomous driving, AI accelerators, etc. Node sizes of 10-45 nanometers account for half of production; they are standard for MCU (microcontroller), standard MPU (microprocessor), "consumer Internet-of-Things", low-cost PCs and laptops, for white goods, cars, medical technology, SmartHome (Alexa), SmartWatch and game consoles. Processes larger than 90 nanometers still account for 38 percent of the chips manufactured and are used in particular in measurement and control technology, machine control and robotics.

Is there similar pressure in the car or appliance industry as there is in SmartPhones and server farms towards smaller nodes?

For cars and industry, robust processes are far more critical; technically, there is no pressure to go to node sizes smaller than 14 nanometers - except for AI accelerators in both areas, which need 7 nanometers and smaller. Going to 7 nanometers requires different technology (manufacturing and physical design of transistors). But going from 28 nanometers to 14 nanometers might well be a good idea for cars, industry and medical technology.

You've mentioned "AI accelerators" several times now. Artificial intelligence" is being advertised or threatened everywhere right now. What is meant by this is "machine learning" - and that requires not only a lot of data (big data), but also a lot of energy. An oft-cited 2019 University of Massachusetts study (Strubell et al., 2019) concluded that training a single "artificial intelligence" emits as much CO₂ as five cars in their entire life cycle. However, in the last two years, the average energy consumption to train an AI model has increased 18000-fold! AI is thus one of the biggest climate killers… ⁵

Yes, and I don't understand at all the current hype in the science feuilleton "AI in the fight against climate change" from a factual point of view: fighting climate change means drastic and rapid reduction of the release of carbon dioxide and methane. Period. At most, the AI models help insurance companies work through possible disaster scenarios, because that's exactly the reduction that isn't happening. And, of course, all energy demand met from fossil resources accelerates climate change.

In the case of "AI," it's not just about the power consumption of the server farms required for this. Even the production of a chip weighing two grams requires more than 70 grams of chemicals, 1.5 kilograms of fossil fuels and 35 liters of water… ⁶

... the energy requirements of the new Bosch Fab in Dresden are comparable to the needs of a small town, explained the Drewag network boss. This is in the lower range; a 3-nanometer fab needs 6.3 terawatt hours per year (the total worldwide electricity demand is 25,000 terawatt hours). That also becomes a problem financially, because it accounts for 30 percent of total fab operating costs, according to McKinsey.

In addition, a modern fab uses up to 45 million liters of water per day, much of it "ultra pure water," which makes recycling difficult. The recycling rate today is 40 to 70 percent, and 80 percent for new fabs being built. Technically possible (but expensive) would be 90 percent - which would still correspond to 4.5 million liters of fresh water per day.

The importance of industrial manufacturing in geopolitics

The U.S. has chosen the electronics/semiconductor industry as a battleground for its geostrategic struggle with China. This was obvious in terms of power politics, since China depends on certain key technologies that the U.S. and its allies - the EU, Japan and South Korea - control. But how insane is it to want to draw sharp dividing lines through sanctions into what is arguably the most globally interconnected industry? Can you please briefly outline the production process again?

First, the wafers are produced in a complex process (see Wildcat 108); then they are cut into pieces and (sometimes several of them) installed in a package; this process is called "packaging". Afterwards, printed circuit boards are assembled with these chips, which are finally installed in the end device (SmartPhone, computer, motor control, ...). The chip design is done by other companies, many of which are located in the USA. All manufacturing steps (wafer, packaging, assembly, final assembly) are distributed worldwide. And what makes things even more complex: in each of the steps, in turn, subcontractors are needed.

For example, only the Dutch company ASML can currently supply lithography machines in the "extreme ultraviolet range" (EUV). To assemble such a machine, costing 150 million euros, ASML in turn relies on 5000 suppliers; among them Zeiss for the optics. Tremendous efforts are needed to cut this complex global network into independent circles of "friendly states"!

On June 14, WirtschaftsWoche warned of the paradoxical effect of friendshoring: "If China were to prepare for an invasion of Taiwan, it would begin to reduce its dependence on Western economies," and it is precisely this process that "Western friendshoring is inadvertently driving."

The Chips and Science Act also calls for friendshoring; subsidies are only available to companies that do not manufacture in or export to countries that pose a "threat to the national security" of the U.S. ⁷ This applies to anything under 28 nanometers, and is aimed directly at TSMC, whose Nanjing fab, which opened in 2018, was announced for 16-nanometer chips... "threat to national security" can easily be extended to other countries, by the way!

Friendshoring certainly helps to dominate one's allies - but possibly only harms China in the short term. A good example is the sanctions policy against Huawei. ⁸ In 2019, the U.S. blacklisted the company, banning TSMC from selling them chips in the 7-nanometer range (and had its chief financial officer arrested in Canada in 2020 for violating Iran sanctions). In order for SMIC, China's largest chipmaker, to produce 7-nanometer chips themselves, they would need the ASML EUV lithography machines I just mentioned. (SMIC is a major customer of ASML, having purchased 81 previous-generation DUV machines in 2021 alone). However, the EUV machines use software developed in ASML's U.S. subsidiary (American intellectual property content) and fall under U.S. sanctions. That's why Huawei has slipped to fourth place as a handset producer - but remains the world leader in telecom equipment with a 31 percent share, twice as much as its nearest rivals Nokia and Ericsson. Meanwhile, the federal government is considering a ban on Huawei components in network infrastructure - even older components may be rolled out again in the face of opposition from network equipment suppliers.

In mid-July 2022, the South China Morning Post reported that the U.S. government wants to expand its export ban on ASML from EUV to DUV machines. The background to this, it said, was that China had managed to use technical tricks to go below 14 nanometers even with DUV. However, the American intellectual property content in DUV machines is not high enough to support the U.S. government's request. The real loser is TSMC.

Meanwhile, how far along is China in trying to build chip manufacturing entirely with its own equipment? How many years behind are they?

The Chinese government is investing massively in the independence of its own chip sector: developing new processes, building new manufacturing capacities and (basic) research. To this end, resources are also being diverted from other tech sectors such as social media and online retail. In 2021, 16 percent of global chip production (this is reported as the total area of wafers produced) came from China; 19 percent is forecast for 2022. However, about half of this is produced in fabs owned and controlled by non-Chinese companies. ¹⁰ The leaders in chip production are still South Korea with 15 percent and Taiwan with more than 60 percent of the global foundry market.

That is the quantitative side; the qualitative is more difficult to assess. The propaganda of the CCP proclaims successes, the competition outside China points to failures. As everywhere where a lot of money flows, there are always reports of unfulfilled plans, unforeseen bankruptcies, corruption and corresponding accusations against fired managers in China. ¹¹

At the end of July, the Taipei Times also reported that SMIC will be able to manufacture 7-nanometer chips as early as 2021; this was also quoted immediately by Heise. ¹² SMIC calls these processes "N+1" and "N+2". "N+1" is also referred to by SMIC as "8nm" or "early 7nm". Rough estimates based on accessible information show it to be about midway between TSMC's 14nm and 7nm. But obviously the people poached from TSMC brought with them the knowledge of how to go below 14 nanometers without EUV. "N+2" (7 nanometers) should be available in 2023.

Both processes use indigenous devices based on more advanced DUV. This is probably the background why the USA now wants to enforce not to supply any more 14nm equipment to China either.

China's focus is clearly on securing critical infrastructure despite all the U.S. sanctions; at the heart of this is the "3-5-2 program" to replace all imported computers (especially in data centers) with Made in China equipment (30 percent in 2020, another 50 percent in 2021, another 20 percent in 2022 ... hence "3-5-2"). In doing so, China is not simply copying what is no longer available to the country, but is relying on the creativity of a mass of highly skilled technicians (many of them trained in the U.S., Taiwan or South Korea) to find quick and workable solutions. Overall, China is likely to suffer more from the general crisis developments in its financial and real estate sectors and from the zero-covid lockdowns than from the U.S. sanctions policy.

The U.S. trade war was intended to prevent China from becoming independent in cutting-edge technology and, in the end, actually helped the development of China's IT industry? The Asia Times commented on the attempted export ban on DUV equipment: Once again, the US tried to close the barn door after the horse is already out… ¹³

This assessment is consistent. Although even Huawei is complying with U.S. sanctions and not supplying Russia, something like a Eurasian bloc could also emerge in the semiconductor industry in the medium term. German companies such as Merck are also involved in the lithography sector. And China still has a few cards up its sleeve, such as import bans for Apple and others, export bans for rare earths, or production bans for US tech companies at Foxconn China (Intel and Apple). Apple, for example, is increasingly moving the final assembly of its devices to India and Vietnam, even if the justification for this is currently still the Covid lockdowns. ¹⁴

... which, after all, fits into the strategy of the USA. Shortly before Pelosi's August 3 "hazard trip" to Taiwan, the "U.S.-Japan Competitiveness and Resilience (CoRe) Partnership," a strategic cooperation between the U.S. and Japan on chip production, was announced. It had been ratified on July 29 during Biden's visit to Japan. Can you assess its prospects for success?

According to an August 1 Asia Times article, they want to "bring 2-nanometer chips to production readiness" by 2025 to reduce their joint dependence on TSMC. ¹⁵ They are probably building on the process IBM unveiled in May 2021, but which is not yet ready for industrial production. TSMC also plans to start producing 2-nanometer chips in 2025, but is already planning the fab to do so. The Asia Times commented that the U.S. and Japan are "three or more generations behind TSMC."

At the core of the venture are Japanese and U.S. companies such as IBM, Intel, Canon, Tokyo Electron and others; in November 2021, for example, Japan's JSR Corporation had completed the purchase of U.S. company Inpria, the world leader in photoresists for EUV, both close industry partners of Intel. The technological gap (especially with TSMC) and Intel's difficulties in stabilizing their fab processes force the US to locate TSMC and Samsung fabs - but now it has become public that they have been working for some time to "hedge" against (a failure of) TSMC or Taiwan.

In fact, Pelosi met with TSMC bosses "to discuss the Chips Act as well as assistance in training engineers in the U.S.," according to the official word. They will have had a lot to talk about! After all, historically, the U.S. had made sure that the CCP's "One China Policy" was accepted worldwide, because at that time they wanted to pit China against the Soviet Union. In 1971, Taiwan lost its representation in the UN so that Nixon could meet with Mao in early 1972. In return, Mao talked about postponing reunification, possibly for a hundred years. In 1979, the U.S. cut diplomatic contact with Taiwan, etc. Today, the time has obviously come to take on Russia and China at the same time, and so it has to perform a balancing act vis-à-vis Taiwan. ¹⁶ The company that the U.S. needs the most suffers the most from its sanctions; so far, TSMC makes 20 percent of its sales with China. How many more years will "the West" need to do without Taiwan's chip industry?

If we look at the whole process - chip design, wafer, packaging, assembly, final assembly - from the U.S. perspective, the design is not a big problem. The biggest challenge is the production of the smaller nodes in the single-digit nanometer range. TSMC and Samsung are the leaders here, followed by Intel (USA) and just behind them China (SMIC). For this production, a highly qualified, young and motivated workforce is needed, which first has to be trained in Europe and the USA, will be more expensive and will also be difficult to motivate for the twelve-hour shifts common in Asia. TSMC is experiencing this right now with their new fab in the US. According to the Semiconductor Manufacturers Association (SEMI), there will be a shortage of up to 750,000 workers in the industry in the U.S. over the next few years; Sanjay Malhotra of SEMI sums up the problem: "You can have a fab, you can fill it with equipment, but if you don't have the people to run it, what's the point?"

Qualcomm had to experience how difficult it is to produce anything usable at all with new processes when it came to 7-nanometer processors (for smartphones): For months, the scrap rate was 100 percent until the first functioning chips finally came out of the factory. Samsung also has a yield of just 35 percent for the new 4-nanometer Snapdragon8 processors nine months after the start of production, the rest is scrap.

Fabs for older processes in Germany and Europe are something else. They run quite stably and can therefore be sufficiently automated (e.g., also the new Bosch fab in Dresden). Driving industrial processes forward technologically, on the other hand, requires the constant intervention of female workers in factory shifts around the clock!

Speaking of which, how has the situation at TSMC in Taiwan evolved?

The number of employees has increased by another 25 percent from 2019 to 2021, to more than 65,000. After rather moderate wage increases of three to five percent over the past five years, TSMC has increased salaries by 20 percent in 2021 - avowedly to retain and recruit qualified workers. They now range from $77,500 to $136,000 per year, on par with the wages offered to applicants for U.S. fabs under construction ($63,000 to $155,000). Conversely, this also partly explains the problems of finding qualified workers in the USA.

All available resources into the chip industry?

Meanwhile, the chip industry crisis is unfolding: the inflation-driven decline in consumer electronics is affecting all sectors: TVs, PCs, smartphones, home appliances... For the first time since 1976, chip sales declined on a month-to-month basis in a June…

Last year, nearly one in five people bought a new SmartPhone (1.39 billion units sold worldwide in 2021). That explains the current record profits - but also shows that they are hard to sustain. The chip industry is in a cyclical downturn. Memory chips are considered the industry's early warning system because they are needed in all sectors. And memory chip makers such as Micron and Hynix are reporting sharply shrinking sales for the third quarter in a row. Meanwhile, all chipmakers are struggling with extended order cancellations, with stable demand coming only from data centers and the auto industry. TSMC raised prices after its record second-quarter profit in 2022 - and cut capital spending at the same time. One reason was delivery delays of fab equipment, but the main reason is that customers' inventories are full; they will order less. ¹⁷

The chip industry has always known cycles of two to three years. But the current downturn is the most severe in more than a decade. Add to that technical problems: The ramp-up of the 3-nanometer process at TSMC has been delayed because Intel simply can't finish its next but one generation of processors - the graphics part of which they want TSMC to manufacture - and has canceled the wafers ordered for 2023. For the first time since Moore's Law was formulated, chip prices are rising.

Gordon Moore had formulated his "law" in 1965, according to which semiconductors would be able to double the performance of chips every two years at approximately the same production costs. In 1968, he founded Intel, which was to become the market leader and top dog in the chip industry for many years. "Moore's Law" was the basis of the "digital revolution" and the benchmark for strategic investment decisions - as long as the performance of the chips grew accordingly. Today, Intel in particular seems to be at the center of the crisis, there is a lack of money, technological development is stagnating, competitors are passing by…

... in 2021, Samsung was ahead of Intel for the first time with more than 83 billion US dollars in sales. When Intel now reported a 22 percent drop in sales in July, a loss of 500 million US dollars in the second quarter and a sales expectation of only 65 to 68 billion for the full year, the share prices collapsed and Intel was overtaken in market capitalization (also for the first time) - by AMD, who are also highly profitable! Intel has been in crisis for a long time, the booming demand during the pandemic had concealed that. But now it's getting down to the wire: the PC/laptop division minus 25 percent, "data centers and AI" minus 16 percent; that means the core business is collapsing, and at the end of July 2022, Intel has given up its last own production of memory chips (Optane). The outlook is bleak: the next processor generation was supposed to come in early 2021, now it looks like early 2023; ¹⁸ the generation after next has been postponed to early 2024; planned investments have been cut from $27 billion to $23 billion - while Intel would have to expand investments in order not to be completely left behind! Comments from the industry are already talking about "Intel's Kodak moment".

"For resilience reasons", new chip factories are to be built everywhere. By the end of 2024, China wants to build 31 new fabs, Taiwan 19, the U.S. 12. Can this get the chip industry out of the crisis? Or does the reference to Kodak mean the end of a technology like analog film back then?

If I take the discussions at the Design Automation Conference in mid-July together with a presentation by TSMC in early August, "Kodak" means the limits of a technology. Mii, TSMC's vice president, wanted to showcase the bright prospects of ultra-high density integrated circuits. With 6G (6G is communications infrastructure plus data centers) would come a "digital data boom for HPC," because an Internet user produces 1.5 GB of data daily, but a self-driving car produces four terabytes, and a "smart factory" even a petabyte! For its part, the DAC discussed the exploding energy consumption that comes with this development. For example, power consumption to train "artificial intelligences" has grown exponentially, and with 6G, 70 to 80 percent of the total energy demand occurs at the edge, i.e., the end device. The billions of devices that make up the "Internet of Things" will soon consume more energy than we can produce, said one conference participant. If there really were 100 million cars on the road with AI, the grids would fly apart (another participant). Some companies in the AI industry promise to reduce their energy needs to one hundredth (!!) in the next five years - DAC's summary: "That's not even close to enough!" We have entered an "era of inefficiency," said another participant.

When Mii then proudly declares that the transition from 7-nanometer nodes to 5 nanometers brought an increase in logic gates per unit area by a factor of 1.83, and the transition from 5 nanometers to 3 brought another 1.6, his own numbers show that even TSMC can no longer follow "Moore's Law" - with a falling trend. At the same time, manufacturing costs and the creation costs of the factories increase significantly from process step to process step!

Moore had formulated his "law" in 1965 to keep doubting investors of his then company in line. Intel has been failing to fulfill it for 15 years - and has big problems with its investors: Even in the red quarter, they pushed through a further increase in dividends paid out, to $1.5 billion. Intel CEO Gelsinger, at 179 million in annual salary the world's highest-paid executive, is substantially preoccupied with getting money from investors. But his narrative is no longer "progress," but "strategic resilience," because that's where most of the government dough can be made. Are we watching geopolitics and overaccumulation, exploitation and ecology collide right now?

In 2021, the number of 300mm wafer fabs increased by 14 to 153. Ten fabs are scheduled to open this year, 2023: 13, 2024 ten. In 2025/26, another 17. By 2026, more than 200 such production lines would be in operation. ²⁰ The buildup of duplicate supply chains in the U.S. and Europe is leading to overinvestment into the long-term downturn.

According to Bloomberg, the semiconductor industry will consume 20 percent of global energy demand in 2030. Add to that the enormous consumption of drinking water! In ecological terms, the chip industry's hunger for resources is a whole different ball game from the auto industry. Almost all available resources - energy, water, labor - would have to be thrown into its growth. But nobody can explain to me why every household lamp has to contain a computer with many chips and possibly even an Internet connection! Neither chips nor data streams are free!

The question is not so much whether this capitalist cycle will come to an end, but whether it will choke on its over-accumulation before it ruins the earth.