Who is developing the chips of the future? RELOADED



Technology and Geopolitics

Julia Hess, Project Manager 
Jan-Peter Kleinhans, Project Director

Artificial Intelligence & Data Science

Laurenz Hemmen, Data Scientist
Lisa Koeritz, Student Assistant (Visualizations)


Executive Summary 

Semiconductor manufacturing has received considerable attention in recent years. New fab announcements from leading semiconductor suppliers, such as Intel, Samsung and TSMC, receive a lot of media coverage, and the groundbreaking events of greenfield investments are regularly attended by heads of state. However, what is often underestimated is the constant pressure to innovate to remain competitive. The semiconductor industry has one of the highest research and development (R&D) margins—companies spend upward of 16% of their revenue on R&D. The chip industry needs innovation across the entire value chain, from wafer fabrication to chemicals, chip design and tools, to produce today’s cutting-edge processors that are at the heart of advances in machine learning, such as ChatGPT. Understanding the distribution of research power across industry, academia and technology organizations (RTOs) can support policy makers in developing strategies to strengthen a region’s semiconductor ecosystem in the long term.  

Two years ago, we began analyzing this underexposed topic with our first publication on the origin of research power in the semiconductor industry. Based on our takeaways from that paper, we now aim to further explore three questions:  

  1. How did the distribution of research power between academia and industry change from 1995 to 2022?  
  2. How much research is conducted in collaboration between institutions and countries?  
  3. How important are EU’s RTOs? 

We analyzed 28 years of invited papers to the three leading semiconductor conferences as a proxy for “research power”. Of course, research power or innovative capacity is much more than the research contributed to academic conferences. At the very least, patents and investments must also be considered. Our analysis of 17,518 papers constitutes a significant piece of the puzzle and offers important insights into the research dynamics in the chip industry.  

There are three key takeaways.  

First, academia took over. Interestingly, industry seems less eager today than 20 years ago to showcase their research at international conferences. The share of papers with industry contributions has declined substantially since the early 2000s. 

Second, international collaboration seems to have plateaued. Since around 2010, global research collaboration has stagnated and even declined in some cases, although there was no decrease in the overall number of contributed papers. 

Third, the EU’s RTOs are a key driver of the EU’s research power. They contributed 36% of the EU’s total papers in 2022, twice as many as EU companies. Imec in Belgium and CEA-Leti in France are, by far, the leading RTOs in the EU. We also observe that RTOs increasingly are collaborating with companies outside the EU, as paper contributions from EU companies have decreased substantially since the early 2000s. 


Making chips is hard work. Researching the architecture of future chips and their production processes is even harder: it takes 18 times more researchers today than in the 1970s to double the transistor density of chips every two years (Moore’s Law). Nonetheless, our society and industry today depend substantially more on these innovative breakthroughs in semiconductor technologies than in the 1970s. We are increasingly in need of energy-efficient power semiconductors for electric vehicles, radio frequency chips for our mobile networks, and artificial intelligence (AI) accelerators to create pictures of astronaut cats, to name just a few. Therefore, the semiconductor industry has one of the highest research and development (R&D) expenditures of all industries—with companies spending an average of 18% of their sales on R&D—second only to the pharmaceutical and biotech industries. Research is at the heart of the semiconductor industry. Yet, policymakers, and their subsidy packages, as well as the media and public discourse, seem to be overly focused on where and by whom chips are built and much less on who develops them. This paper attempts to illuminate what we believe to be a blind spot. 

In our semiconductor data briefs, through quantitative analysis, we put the spotlight on segments of the global semiconductor value chain that are typically not at the center of attention of most policy makers. Earlier this year, we analyzed global semiconductor startup funding and the geography of investment flows. In 2021, we examined the geography of global semiconductor research for the first time. This paper further deepens our analysis. The 2021 publication analyzed the data on a country level to understand (1) how research power shifted geographically over time and (2) which countries collaborated the most with each other.  

In this paper, we expanded our dataset to explore three questions. First, how research power is distributed between academia and private companies—globally as well as per country—since 1995. Second, which academic institutions, research and technology organizations (RTOs), and private companies have coauthored the most conference papers since 1995 and whether we can observe changes over time. Third, whether EU universities and RTOs mainly collaborate with the European Union (EU) industry or with companies outside of the EU.  

The goal of this research is to inform various policy discussions in the EU and abroad, from innovation and research policy to industrial policy and state aid regulation. For policymakers to meaningfully strengthen academic research, the activities of RTOs, and the research power of private companies, different policy interventions will be necessary. Our analysis provides a quantitative foundation for these discussions.  

The first section of this paper introduces the data and provides a first assessment of the distribution of research power between academia and industry in different countries. The second section presents interactive charts to explore how research power between academia and companies, as well as collaboration between the two, shifted over time in different countries. The third section describes the results of quantifying the research power of leading companies and academic institutions based on paper contributions. The last section is a deep dive into the research power of the EU’s RTO ecosystem. 

As with all our data briefs, we collaborated with SNV’s Data Science Unit to analyze the data. We invite you to explore the data yourself through the three interactive charts in the two Chapters “Academic and Industrial Research Power“ and “Top 20 in Industry and Academia”. For questions regarding our methodology and the data science aspects of this analysis, please reach out to Laurenz Hemmen. As for questions regarding the topic of R&D in the chip industry and the semiconductor value chain in general, Jan-Peter Kleinhans and Julia Hess would be delighted to hear from you.  

Understanding the Data

Chart 1: Invited papers to IEDM, ISSCC, and VLSI

We analyzed 28 years (1995–2022) of invited papers from three leading global semiconductor conferences: the International Electron Devices Meeting (IEDM), the International Solid-State Circuits Conference (ISSCC), and the Symposia on VLSI Technology and Circuits (VLSI). In total, our dataset consists of 6299 IEDM papers, 6036 ISSCC papers, and 5183 VLSI papers. Before 2022, papers invited to the two VLSI conferences (green and blue bars) were disclosed separately. Since 2022, VLSI Circuits and VLSI Technology have been officially recognized together as VLSI Technology (green bar). The following charts do not differentiate between the two. For more background on those conferences, please have a look at our first data brief, Who is developing the chips of the future?, from 2021.  

Since 2004, the three conferences together have invited, on average, 657 papers per year with relatively limited fluctuation over time. Additionally, the proportion of paper contributions per conference has remained roughly the same over the entire time span. Although we think that counting invited IEDM, ISSCC, and VLSI papers is a meaningful proxy for research power because of their leading global role in semiconductor research, it is important to understand that (1) there are many more semiconductor technology conferences, and (2) most research is not published in the form of conference papers. To better understand the limitations of our research, please see the FAQ at the end of the paper.  


Chart 2: Papers with foreign collaborations

What you see:

Chart 2 shows the total number of papers (1995–2022) coauthored by one or several domestic institutions (yellow) or in participation with at least one foreign institution (orange). In our analysis, we either affiliate a country based on the affiliations listed in the paper or based on the company’s headquarters. This chart is based on the country stated as part of the author’s affiliation (country on paper). For example, a paper coauthored by Author A (Intel, Jerusalem, Israel) and Author B (Intel, Hillsboro, United States) would count as a foreign collaboration for the US and Israel. Affiliations with Hong Kong are counted as China’s in our analysis.


What it means:

As expected (based on our first analysis), the United States (US), Japan (JP), and the European Union (EU) contributed by far the most papers to the three conferences. What comes as a surprise is the relatively small share of foreign collaborations. The chart shows that most invited papers from 1995–2022 were not based on collaboration between institutions from two or more countries. The share of foreign collaborations is especially small for Japan (JP) and South Korea (KR). This is surprising for a value chain that heavily depends on the transnational division of labor. Notably, Switzerland (CH), Singapore (SG), the United Kingdom (UK), and Canada (CA) have a relatively high share of papers with foreign research collaborations. In the Chapter “Academic and Industrial Research Power“, we delve deeper into how the role of international collaborations has changed over time. 

Chart 3: Share of paper contributions by industry versus academia

What you see:

To better understand where a country’s research power comes from—industry or academia (including RTOs)—we also assessed the relative paper contributions per country based on the authors’ affiliations. Again, we examined only the total number of papers invited to the three conferences in the last 28 years. A more detailed look at the chronological course of this development can be found in the Chapter “Academic and Industrial Research Power“.

For example, a paper coauthored by Author A (University of California, Berkeley, CA) and Author B (Hanyang University, Seoul, Korea) would count as an academic paper for the United States (US) and South Korea (KR).


What it means:

Although the research power is relatively evenly distributed for many countries (South Korea (KR), the United States (US), Taiwan (TW), and the United Kingdom (UK)), there are some outliers. Most of Japan’s papers (JP) were coauthored by Japanese industry, with academia accounting for less than 30% of Japan’s invited papers. At the other end of the spectrum are Singapore (SG) and China (CN), where industry contributed to less than 20% of invited papers. Of course, we can only guess why the distribution of research power between academia and industry varies between countries. One reason could be different attitudes or priorities toward presenting research at international conferences. Another reason could be the simple lack of competitive industry: Although Singapore plays an important role in the global semiconductor value chain, this is mainly due to foreign companies investing in manufacturing in Singapore. The high importance of China’s academic semiconductor research could point to the fact that Chinese companies are still trying to catch up with the global cutting-edge, as China’s role in the semiconductor ecosystem only grew significantly in the last ten years, resulting in less R&D investments.

Chart 4: EU’s research power in detail

What you see:

Chart 4 shows that most of the EU’s invited papers (as an indicator of overall research power) came from five member states—Belgium (BE), France (FR), Germany (DE), Netherlands (NL), and Italy (IT). The yellow/orange bar shows the distribution of research power between industry (yellow) and academia (orange), based on the authors’ country of affiliation. The lower bar for each country shows the same distribution, industry (blue) and academia (green), but based on the author’s institution’s headquarters location. For example, a paper coauthored by an author from Texas Instruments, Freising, Germany, would count as an industry paper for Germany in the yellow/orange bars but would not be counted for Germany in the blue/green bars because Texas Instruments is headquartered in the US. This example also explains why the yellow/orange (country on paper) and blue/green (headquarters location) bars do not add up to the same number of papers. 


What it means:

A few member states stand out: Belgium’s (BE) research power comes almost exclusively from academia. This is mainly due to the country’s Interuniversity Microelectronics Center (imec), the EU’s most competitive semiconductor RTO. The picture looks very different for Germany (DE), where industry accounts for roughly 60% of invited conference papers. Noteworthy is also Italy’s and Belgium’s substantially lower share of industry papers based on headquarters locations; this means that for Italy and Belgium, most papers coauthored by industry were from companies headquartered outside of Italy and Belgium.  

It is important to reflect on the different effects of assessing a country’s research power based either on the country in the authors’ affiliation versus the authors’ institution’s headquarters. The reason behind this is that a country can be an important location for conducting semiconductor research (many paper contributions based on author affiliations), but the origin of that research power can lie within a foreign company (many paper contributions based on headquarters location). Thus, comparing these two different paths of country affiliation can, for example, indicate the relevance of the EU as an R&D hub for foreign companies.  

The extent to which this difference can have a concrete impact is shown by the special case of STMicroelectronics in France, Italy, and Switzerland. The substantial drop in total papers based on headquarters location (shorter blue/green bar versus yellow/orange bar) in both France and Italy can be mainly attributed to STMicroelectronics, France’s largest semiconductor company. STMicroelectronics has most of its R&D and production in France and Italy, but its legal headquarters is in Switzerland, which is not part of the EU and thus does not show up in the blue/green bar. The company accounts for 301 French industry papers (yellow bar) and 139 Italian industry papers.  

Importantly, the charts show the aggregated number of papers over the last 28 years (1995–2022). As will be shown in the next section, the distribution of research power between academia and industry substantially shifts over time.

Academic and Industrial Research Power

Two key ingredients for the high innovative pace and efficiency in the highly specialized semiconductor value chain are knowledge intensiveness and strong collaboration between companies active in different process steps. The R&D of new manufacturing processes and materials builds on close partnerships throughout the value chain and academic institutions, for example, between companies designing chips, fabs and RTOs, equipment and chemical suppliers, and academia for basic research. How has the level of collaboration between academia and industry changed over time? The following section sheds light on countries’ patterns regarding academic and industrial research and collaboration time.  

Chart 5: How did the level of collaboration change?

Share of papers with...

What you see:

The three figures above show different forms of research collaboration and how they evolved over time.  

On the left, the figure shows the share of papers that are based on collaboration between two or more countries. In the middle, you can see the share of papers written by two or more different institutions (this graph does not indicate the type of institution). On the right, the graph displays the share of “mixed” (academia/industry) collaborations over time. For example, in 2000, 15% of all papers in our sample were written in collaboration between academia and industry. The remaining 85% were attributed solely to industrial or academic research on one paper. For context, 397 universities, 308 companies, and 11 RTOs accounted for more than 90% of the affiliations for all 17,518 invited papers.  


What it means:

The charts on the left and in the middle show a clear increase from 1995 to 2010. Since then, both trends have settled at a relatively constant level, and the upward trend has not continued in the last twelve years. 

For collaboration between academia and industry (chart on the right), the picture looks different. Whereas there was a steady increase until 2012, since then, there has been a decrease in collaborations between academia and industry.  

In summary, although collaboration across all three dimensions (between countries, institutions, and academia/industry) is higher today than in 1995, collaboration across countries and institutions seems to stagnate since 2010. The shrinking share of collaboration between industry and academia (graph on the right) can have many reasons. One explanation is that invited papers from industry peaked in 2003 and have declined since then. Since around 2010, more papers have been coauthored by academia than by industry (see interactive Chart 6). In other words, companies are writing fewer papers; thus, there is less chance of collaborating with academia. The data do not reveal, however, why the semiconductor industry seems to be less interested in publishing their research at these conferences. 

Interactive Chart 6: Who is driving semiconductor research in each country?

What you see:

Chart 6 is an interactive chart to explore our data on paper contributions by academia and industry globally and per country. The interactive chart is preset to show the distribution of Japan’s research power between industry (yellow) and academia (orange) over time by counting their paper submissions to all three conferences (IEDM, ISSCC, and VLSI). Please note that you can toggle between the country identification by an organization’s headquarters or the country stated on paper. 


What it means:

We chose Japan as the default setting for this chart because the country is a particularly good example of how economic developments and political initiatives are reflected in the progression of the two curves. As we already pointed out in the first edition of “Who develops the chips of the future?“ published in 2021, Japan’s research power, particularly industrial research, has fallen significantly over the last 28 years.  

This is related to Japan’s declining role in the semiconductor industry over the same period. Nevertheless, the country still plays an indispensable role, especially in the supply of wafers, chemicals, and gases, as well as manufacturing equipment and several other segments of the global semiconductor value chain.  

In contrast to the steadily shrinking role of the industry’s research power, Japan’s academic research continued to increase from 1995 to 2013 and only started declining afterwards. One possible explanation is that in the 1990s, the Japanese government led several initiatives, such as the “ASKA” Project and the “MIRAI” Project. The goal was two-fold: to push R&D for the design and manufacturing of advanced semiconductors and to support companies in adopting new business models based on specialization and vertical integration. However, academic research alone was not able to outbalance the lacking revenues of Japanese semiconductor companies, which directly led to a decrease in private R&D spending.  

In the last two years, the government has started a new effort to attract domestic semiconductor manufacturing, including $7.7bn of funding, and is also driving forward international collaboration in semiconductor R&D. Recently, a memorandum of cooperation (MoC) was signed between the envisioned advanced semiconductor manufacturing base “Rapidus” and imec, the EU’s leading semiconductor RTO, to work on “next-generation semiconductors”.  

Interactive Chart 7: Which country is leading in academic/industry research?

What you see:

Chart 7 is an interactive chart for comparing academic or industrial research power between different countries. You can toggle between the country identification by an organization’s headquarters or the country stated on paper. The chart is preset to give an overview of the development of academic research power measured in paper submissions from China (CN), the European Union (EU), Japan (JP), and the United States (US) over the last 28 years. 


What it means:

In general, the role of academic research has gained relevance over the last 28 years. This upward trend is evident for every country except Japan. The US, the EU, and China significantly expanded their academic research power from 1995 to 2022.  

The pole position of the US in academic semiconductor research: At no point in the last 28 years has any other country contributed more papers from academia than the US. The chart shows that the US currently publishes approximately 1.5 times more papers than in 1995. This success strongly correlates with the establishment of funding from the “Defense Advanced Research Projects Agency” (DARPA) as the centerpiece of federal funding for semiconductor research. The idea was to initiate academic research that ultimately stimulates industrial R&D, leading to commercial innovation. It has proven very successful and has led to several technological breakthroughs in semiconductor manufacturing. However, in the last ten years, there has appeared to be a stagnation in academic paper submissions from the US. Here, recent regulations, such as the “CHIPS and Science Act” and the establishment of seven new “University Microelectronics Research Centers” under the “JUMP 2.0” initiative, come into play. The latter is being led by the Semiconductor Research Corporation (SRC) in collaboration with the DARPA and comes with $250mn of funding for centers at universities such as Columbia or Pennsylvania State.

EU catching up? The EU’s paper count rose from 46 papers in 2000 to 126 papers in 2010. As we will see in the Chapter “Deep Dive: Research Power of EU RTOs”, one key success factor is the EU’s RTOs—imec, CEA-Leti, and Fraunhofer—all of which have a long history of semiconductor research. The Fraunhofer Society is predominantly working on fundamental and evolutionary innovation and collaborates with small and medium-sized companies. CEA-Leti, which belongs to CEA, the French Commission for Alternative Energies and Atomic Energy, also focuses on fundamental research that is then applied in partnership with companies such as STMicroelectronics or SoitecImec is the largest RTO, evolving from its initial expertise in lithography to innovating many more areas in the semiconductor ecosystem. The idea of “co-development”—considering every process step that is impacted by technological innovation—is the fundament of imec’s work.  

China is on the fast track: The most impressive gain in academic research power is attributable to China. Starting with no role in academic research, 86 papers with Chinese contributions were invited to the three conferences in 2022. This puts China almost on par with invited papers from the EU, which published 90 last year. In the case of ISSCC 2023, most invited papers came from China. For the first time, it overtook the US. One reason for China’s increase in academic research power is its pursuit of self-sufficiency, driven by tremendous amounts of government funding that flow into semiconductor R&D. However, the Chinese industry’s research contributions paint a different picture. Until recently, a few papers coauthored by the Chinese industry were invited to the conferences. This again comes back to the fact that China did not play a role in the semiconductor value chain until it started to invest heavily in building a domestic semiconductor ecosystem more than a decade ago. 

A side note: For most of the countries mentioned, the trends of contributed papers are very different when comparing industry and academia.

Top 20 in Industry and Academia

Interactive Chart 8: Top 20 academic and industrial contributors



What you see:

Chart 8 is an interactive chart of the top 20 institutions from academia and industry based on the total number of invited conference papers from 1995–2022. This means that these are not the top 20 institutions today but the top 20 based on their cumulative research in the last 28 years. Academia here includes universities, academic institutions, and RTOs. From the 40 institutions that can be selected from the chart, we preselected imec, TSMC, and KU Leuven.  


What it means:

All three organizations—imec, KU Leuven, and TSMC—started from almost no invited papers in 1995 to 41 (imec), 37 (TSMC), and 19 (KU Leuven) invited papers in 2022 alone. The EU’s leading semiconductor RTO imec is headquartered in Leuven, Belgium, so it makes sense that it has very close research collaborations with the university KU Leuven. The Taiwanese contract manufacturer TSMC has also had long-standing research cooperation with imec since at least 2005. With its increasing dominance in cutting-edge logic semiconductor manufacturing, the company works closely with imec to develop next-generation semiconductor manufacturing processes. This clear upward trend in the research power of imec and KU Leuven, as well as TSMC, highlights two points: First, the growing importance of the EU’s RTOs in the global semiconductor ecosystem. Second, the trend that TSMC, and by extension Taiwan, are not mere manufacturing hubs but key drivers of semiconductor R&D. 

Deep Dive: Research Power of EU RTOs

Chart 9: EU’s research power – RTOs, academia, and industry

What you see:

Unlike the previous charts, in which we include RTOs as a subcategory of academia, here they are shown separately. The aim is to differentiate between academic and RTO contributions and to measure their respective relevance. Chart 9 compares the research power of the EU’s two leading RTOs – imec (orange) and CEA (yellow), with the accumulated research power of other EU RTOs combined (green), EU academia (pink), and EU industry (blue). On the right, you see the accumulated paper count for each entity from 1995 to 2022. In the FAQ section, you can find a list of the RTOs that we have subsumed as “other EU RTOs.” Except for Fraunhofer Society (62 invited papers), Centre Suisse d’Electronique et de Microtechnique (CSEM) (29 invited papers), and the Netherland Organisation for Applied Scientific Research (TNO) (16 invited papers), other RTOs submitted six or fewer papers during the period under review. 


What it means:

Over the past 28 years, the dominance in semiconductor research within the EU has transitioned from industry to academia and RTOs. Up until 2005, both industrial and academic papers constantly increased. A turning point is also obvious: industry’s paper contributions started decreasing, while academic (excluding RTOs) research reached a plateau, ranging around an average of 80 papers per year.  

Paper contributions from imec have also started growing significantly. From 2005 to 2010, imec doubled its published papers from 20 to 40 per year. CEA’s growth period has already started in 2000 and has leveled off around 20 papers per year since 2010.  

The green line shows that the research power of other EU RTOs is comparatively low. In summary, when we refer to the EU’s well-known RTOs leading the way in semiconductor research, we are in fact referring to imec and CEA (and to a much lesser degree Fraunhofer, which here, however, is part of the green “other RTOs” category). To illustrate this point, in 2019 (and every year since), imec alone contributed to more conference papers (48) than the rest of the entire EU semiconductor industry (26). 

A key question is whether EU academia will follow in the footsteps of its Japanese counterpart—declining in research power due to a lack of industrial research power (see interactive Chart 6). It could also be that EU RTOs provide a counterbalance to this trend through increased R&D collaboration with semiconductor companies outside the EU. For the implementation of the EU Chips Act, attention should be paid to strengthening the research power of the EU’s semiconductor industry as well as academic institutions to counteract their downward trends. 

Chart 10: Who collaborates the most with EU RTOs? 

What you see:

Top bar chart (number of papers): Each country has two bars (fully colored and semi-colored) that both show the total number of industry papers coauthored with EU RTOs. The difference is the country to which the company is attributed: either the country given in the paper (fully colored bar) or the company’s headquarters location (semi-colored bar). For example, we will use a conference paper coauthored by Author A (Globalfoundries, Dresden, Germany) and Author B (CEA-Léti, Grenoble, France). Option 1: Globalfoundries’ country given in the paper is Germany; thus, this paper would be counted in Germany’s fully colored bar. Option 2: Since Globalfoundries is headquartered in the US, this paper would be counted in the semi-colored bar from the US. 

Bottom bar chart (share of industry papers): The bottom bar chart shows the share of collaborations with EU RTOs as a percentage of all industry papers from a country. For example, in the 28 years from 1995 to 2022, roughly 10% of conference papers coauthored by German companies were written in collaboration with an EU RTO.  


What it means:

Not surprisingly, French, Belgian, German, Dutch, and Swiss industries are heavily collaborating with EU RTOs. Among the leading collaborators are companies from the US and Japan and, to a lesser degree, from Taiwan and Korea. For the EU’s semiconductor ecosystem, collaboration with industry outside the EU is a double-edged sword. On the one hand, enhancing international and academia–industry collaboration is an important success factor for constant innovation in this transnational value chain. Foreign companies that significantly invest in R&D activities also strengthen the EU’s role in the global semiconductor ecosystem as a whole. On the other hand, semiconductor research by foreign companies in collaboration with EU RTOs also poses the risk of the transfer and commercialization of intellectual property outside of the EU.  

The bottom bar chart shows that for companies outside the EU, collaborations with EU RTOs only play a minor role. Less than 5% of all papers invited from the South Korean (KR), Taiwanese (TW), and American (US) industries were coauthored with EU RTOs. However, this takeaway only refers to the total paper count for the last 28 years. The next chart presents an analysis of the development over time. 

Especially in the EU, there are big differences between the two country affiliation types (semi- and fully colored bars differ in length). The following are some possible explanations. 

France (FR): The French industry coauthored 190 or 43% (51 or 47% based on headquarters) of its papers with EU RTOs. This is the highest share of collaboration between industry and EU RTOs in our sample. The steep drop in papers (190 to 51) when accounting for headquarters location is due to the special case of STMicroelectronics introduced in the Chapter “Deep Dive: Research Power of EU RTOs”. It is one of the EU’s largest semiconductor companies and headquartered in Switzerland but with most of its production and R&D in France and Italy (in 1987 French “Thompson Semiconductors” and Italian “SGS Microelettronica” merged to STMicroelectronics). The high share of collaboration with EU RTOs is due to CEA-Leti and STMicroelectronics, both located in Grenoble, France. STMicroelectronics is also the reason for the huge discrepancy between the two bars for Switzerland (CH) —the country itself has a comparatively small semiconductor industry.  

Belgium (BE): Belgium industry coauthored 120 or 55% (2 or 25% based on headquarters) of its papers with EU RTOs. The steepest drop in papers (compared to every other country)—from 120 to 2—when considering headquarters location is likely because of the leading role of imec in attracting foreign companies, within and outside of the EU, to collaborate in semiconductor research in Belgium. Additionally, Belgium has a comparatively small domestic semiconductor industry. Another indicator for this is that only two papers were coauthored by industry headquartered in Belgium in collaboration with EU RTOs. These two papers, in turn, accounted for 25% of all papers coauthored by Belgian industry. 

Chart 11: EU academia/RTO collaboration with industry

What you see:

Chart 11 shows the 5-year moving average of invited papers. In contrast to previous charts, it does not differentiate between academic institutions and RTOs. Solid lines and dashed lines show different country attributions: affiliation on paper (solid line) or company headquarters location (dashed line). The gray lines show the EU industry’s total paper contributions. For example, a paper with a coauthor from Texas Instruments, Freising, Germany, would show up in the solid gray line (affiliation given in the paper is Germany) but would not show up on the dashed gray line because Texas Instruments’ headquarters is in the US.  

The yellow lines show collaborations between EU academia/RTOs with EU industry. The orange lines show collaborations between EU academia/RTOs with non-EU industry.  

Example 1: A paper coauthored by Author A (imec - Holst Centre, Eindhoven, The Netherlands) and Author B (Renesas Electronics, Tokyo, Japanwould show up in both orange lines because Renesas is headquartered in Japan, and Author B’s affiliation in the paper is Japan. 

Example 2: A paper coauthored by Author A (imec, Leuven, Belgium) and Author B (HiSilicon, Leuven, Belgium) would show up in the yellow solid line (EU industry by country on paper) and the orange dashed line (non-EU industry by company’s headquarters location) because HiSilicon is headquartered in China. 


What it means:

Declining research power of EU industry since 2007: As we have seen in previous charts, invited papers coauthored by EU industry (as a proxy for research power) have declined since around 2007 (gray lines). The yellow lines (academia/RTO collaboration with the EU industry) are thus limited by the gray lines. This also explains why the yellow lines follow the gray lines in their decline since 2010: Since the research power of the EU’s semiconductor industry is declining, naturally there are fewer and fewer opportunities for research collaborations between the EU industry and EU academia/RTOs (yellow lines). 

Importance of research collaboration with non-EU industry: Since 2011, the dashed orange line (research collaborations with non-EU industry based on headquarters) has been higher than the dashed yellow line (research collaborations with EU industry based on headquarters). This shows that since 2011, EU academia/RTOs have had more research collaborations with companies headquartered outside of the EU than with companies headquartered in the EU. 

Overall decline in research collaborations since 2016: Chart 5 shows that research collaborations between academia and industry have been declining since around 2013 globally. This can also be seen in Chart 11. Research collaborations with non-EU industries have also declined since around 2016 (orange lines). This does not seem to be a development unique to the EU’s academia and RTOs but a global trend, as Chart 5 already indicated. However, Chart 9 shows that paper contributions from EU academia (excluding RTOs) stagnated since 2016, which could be one reason why research collaborations with non-EU industry have also declined since then (Chart 11, orange lines). 


This research is based on a substantial amount of data with many insights and interesting information artifacts. Here, we would like to point out three broader themes that we noticed throughout the entire analysis.  

First, while the global semiconductor value chain is deeply rooted in the transnational division of labor, the same does not seem to be the case regarding research collaboration. No country controls the entire production stack, from chip design and IP to wafer fabrication, raw materials, chemicals, manufacturing equipment, and, lastly, back-end manufacturing. We had assumed that semiconductor research is equally transnational—but this does not seem to be the case, at least not to the same degree as semiconductor production. First, research collaboration across countries and between organizations seems to have been relatively stagnant since 2010. Second, research collaboration between companies and academia (including RTOs) has been declining since 2010 (Chart 5). Although assessing the various reasons for this stagnation and decline is beyond the scope of this paper, this interesting phenomenon deserves to be researched in more depth in the future.  

Second, the distribution of research power between industry and academia looks different for different countries (interactive Chart 6). In the US, research power has been evenly distributed between industry and academia since around 2007. In 2007, the same was true for the EU, but since then, the distribution of research power has shifted from industry to academia (including RTOs), and the latter accounts for 75% of current paper contributions. In that regard, China’s distribution of research power today looks very similar to that of the EU—predominantly driven by academia but with quickly increasing industry participation in recent years. Further, with the intensifying US–China technology rivalry, the Chinese industry might shy away from showing off its research advances in international conferences.  

Third, the EU’s research power has substantially shifted and increasingly depends on RTOs, namely imec and CEA-Leti. These RTOs, because of decreasing research power by the EU industry, turn to foreign companies for research collaboration (Chart 11). This is a double-edged sword for the EU because the commercialization of that research mostly lies with the foreign company. However, without its RTOs, the EU would have substantially less research power in the global semiconductor ecosystem. One measure of success for the EU Chips Act over the next ten years will certainly be whether these downward trends can be reversed and whether the research power of the EU’s industry can be rejuvenated. 


Data Sources and Definitions 

What is the basis of our dataset?

Our dataset is based on the metadata of academic papers published at IEDM, ISSCC, and VLSI (Technology and Circuits) conferences between 1995 and 2022. These conferences were chosen as three longstanding and leading international semiconductor research conferences.

While the papers are published as closed access, IEEE, the publisher of all three conference proceedings, provides metadata for all conference papers via the IEEE Xplore API. For each paper, we collected the affiliations of all authors. To mitigate missing data from IEEE, we complemented affiliations by searching for each paper’s digital object identifier (DOI) on OpenAlex, a free open-source catalog of research papers.

We extracted the institution name and country from each affiliation and inferred its headquarters’ location and type (whether it was an academic institution or industry). We explain our methodology in more detail below.

Why did we supplement the IEEE data with data from OpenAlex?

We found that the metadata retrieved via the IEEE Xplore API contained parsing errors that sometimes resulted in missing or incomplete affiliations. This is disproportionally the case for papers with many affiliations and papers authored or coauthored by RTOs. To mitigate these errors, we extended the dataset obtained via IEEE Xplore with affiliations from OpenAlex. Our manual checks showed that both datasets contain missing or incomplete affiliations but complement each other. The resulting dataset contains redundant entries for many affiliations that are present in both IEEE and OpenAlex. As explained in “How did we count contributions?”, this does not change our paper counts.

How did we count contributions?

Our analysis aims to generate insights into the research strengths of different entities, namely countries, organizations, and sectors. We used “full credit” to count contributions per country or type of institution. This means that each country on a paper got one counted contribution, rather than fractionally splitting the contribution between authors—that is, it did not matter if there are five authors from a particular country or just one—in both cases, that country was counted just once. The same logic was applied when we counted institution types: if there was at least one academic affiliation on a paper, “academia” got one point, no matter how many authors from academic institutions contributed to the research.

In those parts of the analysis in which we treated the EU as a single entity, contributions from different member states were counted as one. For example, a paper written by three French and two German authors increased the EU paper count by one. Assigning full credit to each entity on a paper emphasizes the aspect of collaboration: a paper written by institutions in Switzerland, the US, and Belgium increased the total count of contributors by three, whereas a paper solely written in Germany only increased the count by one.

On the flip side, this means that some papers were weighted higher than others in the total count. Considering country contributions, the sum of the paper counts for all individual countries is larger than the actual sum of papers. Generally, it is important to bear in mind that our paper count varies depending on the aspect of interest—whether we look at countries, institutions, or types of institutions, or whether EU countries are bundled as one country.

Why can the sum of paper contributions exceed the total number of papers?

Papers are often coauthored by multiple institutions, countries, or institution types. Using full credit counting (as explained in “How do we count contributions?”), one paper can increase the count of different entities: A paper coauthored by the US and Japan increases the count for both countries. Hence, the sum of paper contributions for all countries will exceed the total number of papers.

Which countries are counted as the EU?

The label “EU” refers to EU-27, excluding the United Kingdom. This applies to all years starting in 1995, regardless of when certain member states joined the EU. Therefore, even if a country joined the EU after 1995, it was still considered part of the EU from 1995 onwards in this analysis. We made this choice to maintain a consistent definition of the “EU” label over time, allowing for any observed increases in paper count in our graphs to be interpreted as genuine increases of research productivity rather than a result of changes in EU membership.

Which paper contributions are counted as China?

The counts for China (CN) do not include Taiwanese (TW) papers. However, papers from Hong Kong and Macau are counted as contributions for China.

How are “academia” and “industry” labels defined?

In the context of this analysis, the label “academia” encompasses academic institutions like universities but also other government-funded public research, such as from RTOs (when it is not noted otherwise for the specific chart). The category “industry” refers to private companies. See “How did we identify institutions?” for more information on how these labels were assigned.

What is the difference between the source “country on paper” and “headquarters”?

Each affiliation string included in our dataset has a country name. This is the “country on paper” in which the actual work was done. An institution’s headquarters, by contrast, refers to the institution’s main location or legal seat. For example, “IBM Research EU, Rüschlikon, Switzerland” has Switzerland as the “country on paper” but the US as the “headquarters.”

Which RTOs are listed under “other EU RTOs”?

Total Papers 


Fraunhofer Society  62
Centre Suisse d’Electronique et de Microtechnique (CSEM)  29
Netherlands Organisation for Applied Scientific Research (TNO)  16
Tyndall National Institute  6
VTT Technical Research Center Finland  4
Silicon Austria Labs   3
Helmholtz Association  2
Austrian Institute of Technology  1
The Foundation for Scientific and Industrial Research (SINTEF)  1

Data Processing and Analysis  

How did we identify institutions?

We started with 58,376 affiliations (33,617 from IEEE and 24,759 from OpenAlex) on 17,783 papers. Many of these affiliations were redundant; they described the same institution but referred to different departments, spellings, or abbreviations.

There were 19,406 unique affiliation strings among the 58,376. Our goal was to identify different affiliation strings belonging to the same entity, such as a specific company or university: “Samsung,” for example, has 749 different affiliation variations in our dataset, like “Samsung Electronics, Hwaseong, Korea,” “Assignees at IMEC from Samsung, Kapeldreef 75, Leuven, Belgium,” or “Process Development Team 2, Semiconductor R & D Center, Samsung Electronics Co., Ltd., South Korea.”

To identify all affiliations referring to the same entities, we first automatically removed unnecessary information, such as geolocation, department, or contact information, from the raw affiliation strings. “Department of Electrical Engineering, University of Notre Dame, Notre Dame, IN, USA,” for example, was transformed to “University of Notre Dame, Notre Dame.”

To obtain the resolved institution names, we manually assigned an entity name to each previously cleaned affiliation string. To make this labeling process more efficient, we embedded all cleaned affiliations using OpenAI’s text embedding model text-embedding-ada-002. This provided us with a vector representation of all affiliations, which encodes semantic relationships to a certain extent. During labeling, we randomly sampled an affiliation and showed it to a human labeler, together with other affiliations with similar representation vectors. Hence, we were able to label many affiliations that belonged to the same entity at once.

Using vector representations helped us find different spellings, misspellings, or abbreviations belonging to the same entities. For example, the “SeoulTech” representation vector is close to that of the “Seoul National University of Science and Technology.” Comparing representation vectors also helped us find incomplete affiliation strings (resulting from parsing errors in IEEE Xplore or OpenAlex). For example, “Technology CAD, Hilsboro” and “Technology CAD, Intel, Hilsboro” have similar representation vectors, although the former is missing the name of the parent company “Intel.” By checking the original paper’s PDF in relevant cases, we verified that representation vectors reliably cluster variations of affiliation strings belonging to the same entity, even in the case of incomplete affiliations.

Using the above procedure, we resolved the names for more than 90% of the affiliation strings in our data. The remaining affiliation strings could not be resolved due to either being incomplete, such as “Department of Physics, USA,” or referring to smaller institutions with few occurrences in the data that were not economical to label. These unresolved affiliations were treated differently in terms of how we determined headquarters and type, as described below.

How did we determine countries?

Affiliations in IEEE and OpenAlex typically ended with a country name representing the location of the research institution. We assembled a dictionary with country names in different languages and US states to match these countries. We were able to identify a country for 89% of all raw affiliations. We found a country for 96% of IEEE affiliations, while for OpenAlex, 28% had no country. However, many affiliations were included in both sources, effectively reducing the number of missing countries. In total, 99% of papers had at least one country assigned. Affiliations without a country were removed from the dataset and were not part of the analysis.

Please note that OpenAlex provides country information, which was obtained by matching the affiliation with an entry from the RoR database. However, we did not use this information because we found that the country information provided does not always refer to the location that appears in the original paper. For example, in some cases, the location provided referred to the wrong subsidiaries in another country, such as “Philips Research Laboratories, Eindhoven (Netherlands)” being matched to the RoR entry “Philips (Finland).”

How did we determine headquarters?

Headquarters were determined manually for all resolved institution names (see “How do we identify institutions?”) that occurred more than once in the dataset. For the remaining 10% of affiliation strings, which occurred in the dataset only once (mostly the names of small companies or universities, or incomplete affiliations, which could not be further resolved), we used OpenAI's GPT-3.5 chat completion API to identify the headquarters location from the raw affiliation string. To determine headquarters with GPT-3.5, we use the following prompt: “Please provide the headquarters of the following academic institutions and companies: {institutions}. For each institution or company, in the same order, in a new line, repeat the EXACT input string followed by a colon and the ISO 3166-1 alpha-2 country code of their headquarters.”

Manually verifying 250 randomly sampled headquarters determined in this fashion resulted in an accuracy of 95%. For affiliations of institutions that we categorized as academia, accuracy was 99%; for affiliations categorized as industry, it was 89%. Although these errors might bias the results of our analysis in unexpected ways, we believe this is unlikely, since only 10% of headquarters were identified via the language model.

How did we determine affiliation type?

For all resolved institution names (see “How did we identify institutions?”), the affiliation type was determined manually. For the remaining 10% of affiliation strings, we used a zero-shot classification approach: each institution name was embedded into a vector space using OpenAI’s text embedding model text-embedding-ada-002. As class labels, we chose “university” for academic institutions and “company” for industry. We evaluated different class labels by comparing the type assigned via this method to the types we manually assigned to the resolved institution names. The accuracy of the class labels “university” and “company” was 95%. We also randomly sampled 250 affiliations from the entire dataset and verified the affiliation types. The obtained accuracy was 99%.

How did we deal with missing or incomplete data?

The metadata about conference papers provided by IEEE contained errors stemming from the automatic parsing of PDF files. One of the most common types of errors we observed in the IEEE metadata was missing or incomplete affiliations: if multiple departments/locations of one company worked together on a paper, the company name was often included only in the first or last affiliation, leaving incomplete affiliations, such as “Corporate Research Center, Germany,” which belong to “Infineon” in this specific case.

Using full credit (see “How do we count contributions?”) to count citations mitigated errors stemming from missing data. The count of a country, company, or type was increased by one if at least one entity of that category was on the paper. Hence, it did not matter if we missed Germany once if another German affiliation was on the paper.

Combining OpenAlex with IEEE made it even less likely to completely miss an entity. This effectively increased the reported accuracies for determining attributes, such as countries or types. OpenAlex, in rare cases, provided incorrect affiliations that were not on the actual paper. We mitigated this problem by only adding OpenAlex affiliations to the dataset if the resolved entity also occurred somewhere in the IEEE dataset.


What are limitations of studying published papers?

First, the research output of three, albeit major, conferences reflects only part of the semiconductor engineering research field. Over the 28 years covered by our analysis, the research landscape has broadened. The boundaries of these fields are fluid; adjacent fields, such as artificial intelligence, also conduct research on specialized chips.

Second, research output is not a perfect proxy for innovation and technological leadership (see here). Research conducted by for-profit companies has not always been published. For a more complete picture, an investigation of research output via peer-reviewed papers should be supplemented with analyses of the patents, investments, and market shares of relevant players.

As shown in Chart 1, the number of papers per year has remained roughly constant since 2004. It is important to keep in mind that the total number of papers accepted at conferences is usually limited. This means that an increase in papers from academia necessitates a decrease in industry papers. Analyses such as ours, therefore, do not allow for conclusions on the evolution of the entire research area itself.

What are the potential sources of error and bias in our dataset?

Our dataset has different potential sources of error. First, there are errors in the raw data from IEEE and OpenAlex. As described in “How do we deal with missing or incomplete data?” IEEE data are sometimes incomplete or faulty, and OpenAlex sometimes lists affiliations that can’t be found on the original paper. By using both data sources and filtering out unreliable information from OpenAlex, we obtain a redundant and overall more accurate data basis than with any single source. Since we count using “full credit” (see “How do we count contributions?”), we avoid counting duplicates.

Second, there are errors in our processing pipeline. We randomly checked the error rates for different steps and found that the errors were small enough not to bias our conclusions. Determined institution types are very accurate: Manually verifying 250 randomly selected affiliations yielded only two errors for the affiliation type. Similarly, headquarters were manually checked for 90% of entries, while a random check on 250 of the remaining entities yielded an accuracy of 95%.

In rare cases, IEEE contained multiple affiliations in a single string. For assignees from companies at RTOs, for example, “Panasonic assignee at IMEC,” we used the companies as affiliations, as the RTOs usually have separate authors on these papers. In other cases, we had to choose randomly.

Additionally, we filtered out affiliations with missing countries. This is described in more detail in “How did we determine countries?” Although we only lacked countries for 4% of IEEE affiliations, our primary data source, we missed 28% for OpenAlex. This introduced potential bias if there was regularity in the missing values. We observed, however, that due to the redundancy of our sources, fewer affiliations lacked a country in both sources.

Lastly, readers should be aware of what the data show and what can be generalized from it. As explained in “Limitations of studying published papers,” three conferences do not represent the entire field of research, and peer-reviewed research is not sufficient to draw definitive conclusions about the technological impact of a specific country or institution.

How accurate are our results?

We are confident that the conclusions drawn from our analysis are valid because our interpretation relies only on macroscopic trends, not specific numbers. For example, the shift from industry-dominated to academia-dominated research at selected conferences is clear and consistent over time. All our charts and analyses show no change in character when using only 90% of the data, where institutions were manually resolved. For details about accuracies for specific attributes, see the sections about how each attribute was determined.


We would like to thank Pegah Maham for developing ideas, guidance and feedback on our paper, Shannon Reitmeir for the support with the creation and verification of the dataset, Anna Semenova for reviewing our code and analysis, Alina Siebert for the creation of the layout, Luisa Seeling for her support in editing the text, and Sebastian Rieger and Ernesto Oyarbide-Magaña for helping us spread the word about this publication.