By
Javier
Surasky
The
exponential rise of artificial intelligence (AI) has triggered a fast-moving,
highly complex transformation in the global semiconductor value
chains—materials such as silicon that can behave as electrical conductors or
insulators depending on how they are used.
This
duality allows extremely precise control over electrical current, making
semiconductors the physical foundation of all modern chips. These
chips—essentially semiconductors integrated into a single piece of silicon—are
the driving force behind computers, mobile phones, and the processing capacity
required by AI. They have become a strategic infrastructure, with production
concentrated in a small number of countries and firms.
As a
result, the transformation of production chains follows three interconnected
vectors:
- The expansion of the market for
advanced accelerators and memory.
- The geographic concentration of
semiconductor manufacturing capacity; and
- The sustained environmental and
energy impact associated with increasingly powerful hardware.
In our
previous article on U.S.–China interdependence in their race to lead AI, we
examined how the structural power of a digital world is reshaping the global
economy while generating new vulnerabilities. In this second blog, we “zoom in”
on the material, productive, and regulatory foundations that sustain—and
strain—the global expansion of AI, with special attention to the chip market,
memory requirements, cloud-computing growth, and export-control frameworks.
A
distracted reader might assume that this is a text for engineers—something I
could never write, as I am not one—but in reality, it is aimed primarily at
social scientists. Their diverse areas of expertise are also being transformed,
and their work increasingly requires new analytical capacities to understand
the multiple systemic impacts of the digital transition we are living through.
Let us
begin by stating an evident truth: the market for AI accelerators has
maintained an upward trajectory. According to recent estimates, the global
market value reached USD 140.55 billion in 2024 and could exceed USD 440
billion by 2030, implying an annual growth rate of nearly 20% (Mordor
Intelligence, 2024).
The
structural demand behind these figures stems from ever-larger models and
increasingly complex compute workloads, driven by advances in video-generation
AI, multimodal systems, and autonomous agents.
But as this
demand skyrockets, the industry is running into physical limits. A Graphics
Processing Unit (GPU) is a type of chip designed to perform multiple parallel
calculations, giving it massive, simultaneous computational capacity ideal for
highly intensive and repetitive operations such as training and running AI
models. In other words, GPUs are the core of any data center that trains large
language models, image-processing systems, or generative AI models.
Next-generation
GPUs—such as NVIDIA’s H100 and GB100 Blackwell—are approaching the maximum size
allowed by lithography reticles, the process through which circuits are “drawn”
inside a chip. This creates a physical barrier and helps explain why chips cannot
increase capacity simply by “getting bigger.” Instead, new chip architectures
are required, such as multi-die systems in which a chip is not made from a
single silicon piece but from several smaller pieces (dies) interconnected
within the same package, with each die performing a specific function so that,
together, they operate as a single, more powerful chip.
Following a
similar “decoupling and recoupling” strategy, we find High Bandwidth Memory (HBM),
a type of ultra-fast memory that enables the system to operate with large
volumes of data without experiencing delays. Unlike its predecessors, HBM is
stacked in tiers. To understand its architecture, we can think of building a
house of cards, where the goal is to add layer upon layer to reach the highest possible
level without the structure collapsing. Each complete layer of cards acts as an
HBM, saving space and increasing operational speed. Currently, and without
knowing where the upper limit lies, a next-generation accelerator is reaching 1
terabyte (Russell,
2025) of HBM by stacking 16-level tiers, something like taking 16 houses of
cards, each with 16 levels, and connecting them so that, together, they form a
structure that remains stable without crumbling—resulting in the ability to
handle enormous volumes of inputs and outputs per second locally, without
relying on “the cloud,” which tends to slow processes down.
Memory
demand—once driven primarily by the need for greater processing efficiency—is
now driven by the challenge of overcoming the functional limits of today’s AI,
limits imposed by the inadequacy of existing physical infrastructures. A direct
consequence is the continual rise of the hardware’s embodied carbon footprint.
According to the Global AI GPU Manufacturing Carbon Emissions Forecast (Russell,
2025), emissions from accelerator manufacturing rose from 1.8 million
tonnes of CO₂ equivalent in 2024 to an expected
21.6 million tonnes by 2030, nearly 9% of the semiconductor sector’s total emissions for that year.
New
accelerators use large amounts of HBM, which is more complex and
resource-intensive to manufacture. For example, a chip like NVIDIA’s
H100—introduced in 2023 and still in use in many laboratories—generates most of
its emissions through its central processor. But newer models (yes, 2023 is no
longer “new”), such as AMD’s MI300X or NVIDIA’s upcoming Rubin Ultra, shift the
dominant share of emissions to fabrication processes (Chen et al., 2025; TechInsights,
2025a, 2025b;
Rteil,
2025).
As
expected, the limited availability of such a strategic resource generates
international tensions, amplified by the increasing concentration of
capabilities. To understand this, it is key to recall that semiconductor
“generations” are measured in nanometers (nm). When we talk about “5 nm,” “3
nm,” or “2 nm” chips, we are not describing a transistor’s physical size but
the technological generation of the manufacturing process. Fewer nanometers
means a more advanced process.
With that
in mind, the Semiconductor
Industry Association (2024) notes that over 70% of global fabrication
capacity below 10 nm is located in South Korea and Taiwan, with TSMC
(2024) producing about 90% of leading-edge nodes below 5 nm.
This
results in a distributed production chain beginning with natural-resource
extraction—metallurgical silicon (mainly from China, followed by the U.S. and
Germany), copper (Chile, Peru, China), cobalt (highly concentrated in the
Democratic Republic of the Congo, with smaller contributions from Indonesia),
gallium (almost exclusively from China), rare earths (China, then the U.S. and
Australia), nickel (Indonesia, the Philippines, Russia), and lithium
(Australia, Chile, China), according to the U.S. Geological
Survey (2024) and the International
Energy Agency (2024).
The United
States leads the intellectual and design segments, controlling the software and
intellectual inputs needed to design and manufacture advanced chips, while
fabrication is concentrated in specific Asian countries (OECD,
2024a). Extreme-ultraviolet (EUV) lithography—used to etch microscopic
circuits—lies in the hands of very few actors, notably ASML in the Netherlands
(OECD,
2024b; ASML,
2024).
Assembly
and packaging occur primarily in China, Malaysia, and Vietnam. The main chip
“consumers” are China (35–40%), the U.S. (≈20%), Germany (≈10% EU), Japan
(6–7%), South Korea (5–6%), India (≈3%), and Taiwan (≈3%) (SIA 2024; OECD
2024a).
Data-center
integration brings the chain back to the United States, home to Amazon
(Seattle), Microsoft (Redmond), Google (Mountain View), Meta (Menlo Park), and
Oracle (Austin), along with specialized providers like CoreWeave, Lambda, and
Crusoe. Nebius (Zürich) is the main exception (Synergy
Research Group, 2025).
The Top500
supercomputer ranking adds another layer of insight. As of November 2025, the
top ten systems were Frontier (U.S.), Aurora (U.S.), Eagle (U.S.), Fugaku
(Japan), Leonardo (Italy/EU), Summit (U.S.), Sierra (U.S.), MareNostrum
(Spain), Lumi (Finland/EU), and Polaris (U.S.) (Top500 Project, 2025).
This list
reveals:
·
Heavy
U.S. leadership (six of the ten systems).
·
No
Chinese systems—not because they lack capacity, but because China stopped
reporting results. Dongarra (2022) and Shilov
(2023) estimate that Sunway Oceanlite and Tianhe-3 would rank among the top
five.
·
Two
systems at exascale performance; estimates suggest China may also have two.
·
Nine
of the ten systems are in public consortia or national labs; only Eagle
(Microsoft) is privately owned.
All this
underscores the importance of export-control policies. The Netherlands, for
example, has imposed restrictions on the export of advanced lithography
equipment to China due to U.S. pressure.
Physical
infrastructure also faces severe energy bottlenecks. Studies (Patterson et al., 2021; Thompson & Spanuth, 2021; Chamness,
2025; Russell,
2025) show that compute-demand growth is outpacing improvements in energy
efficiency. Major AI companies are increasingly competing for low-carbon
electricity sources, raising questions about the long-term sustainability of
the current AI expansion model.
In this
scenario, the social sciences play an indispensable role in complementing
techno-economic analyses. Decisions on export controls, fiscal incentives,
datacenter siting, manufacturing subsidies, and environmental requirements
emerge not only from physical and economic constraints but also from
institutional frameworks, political priorities, democratic strength, and
state–business relationships.
At the same
time, the pace of technological adoption depends on regulatory coordination,
international trade relations, strategic alliances, standard-setting, and
multilateral negotiations within the complex AI ecosystem—placing digital
justice at the center of emerging debates.
As we have
repeatedly argued in our blog, the accelerated expansion of AI is not a purely
technological phenomenon but a reconfiguration of the global political order.
Social sciences must claim their seat in debates on distributive impacts,
political and social analysis, industrial policy design, geopolitical scenario
forecasting, environmental protection tools, and risk assessment, among many
other areas.
Without a
multifaceted understanding of the social dimensions of AI development, we
cannot grasp the whole meaning of the digital infrastructure being built, who
will benefit from it, and its consequences for people’s lives.
Too much is
at stake to leave these matters solely in the hands of a small group of
technical experts. It is essential for social scientists to “get their hands
dirty” with AI—before others make decisions in silence, under the astonished
gaze of academia, when it may already be too late.
References
ASML.
(2024). Export controls and lithography systems. Government of the
Netherlands. https://www.government.nl/topics/export-controls-of-strategic-goods
Chamness,
L. (2025, 3 de octubre) The Hidden Environmental Cost of Advanced AI Chips.
TechInsights. https://library.techinsights.com/hg-content/35313f98-2500-44f3-8a28-6dd7d8ceccf2
Chen, X., Han, L., Bhagavathula, A., & Gupta, U. (2025).
CarbonClarity: Understanding
and addressing uncertainty in embodied carbon for sustainable computing.
arXiv. https://doi.org/10.48550/arXiv.2507.01145
Dongarra,
J. (2022). A New Era in Computing: Exascale Systems and Their Implications.
Communications of the ACM, 65(6), 34-36
International
Energy Agency. (2024). Global Critical Minerals Outlook 2024. https://www.iea.org/reports/global-critical-minerals-outlook-2024
Mordor
Intelligence. (2024). AI
Accelerators Market. Growth, trends, forecasts (2025-2030). https://www.mordorintelligence.com/industry-reports/ai-accelerators-market
OECD. (2024a).
Mapping the global semiconductor value chain. OECD Publishing. https://www.oecd.org/content/dam/oecd/en/publications/reports/2025/06/mapping-the-semiconductor-value-chain_5ba52971/4154cdbf-en.pdf
OECD.
(2024b). The chip landscape: Economic, geopolitical and environmental
dimensions of semiconductors. OECD Publishing. https://www.oecd.org/content/dam/oecd/en/publications/reports/2025/12/the-chip-landscape_27ef5d87/02dbd028-en.pdf
Patterson,
D.; Gonzalez, J.; Le, Q.; Liang, C.; Munguia, L-M.; Rothchild, D.; So, D.;
Texier, M. y Dean, J. (2021) Carbon Emissions and Large Neural Network
Training. https://arxiv.org/abs/2104.10350
Rteil, N. (2025, 18 de septiembre). Understanding
GPU’s energy and environmental impact – Part I. Interact. https://interactdc.com/posts/understanding-gpus-energy-and-environmental-impact-part-i/
Rusell, S.
(2025) Global AI GPU manufacturing carbon emissions forecast 2026-2030.
TechInsights. https://library.techinsights.com/hg-content/d77d2810-28cb-4f7b-bcad-17c937a10b41#moduleName=My+Library&reportCode=SME-2510-805&subscriptionId=null&channelId=null&reportName=Global+AI+GPU+Manufacture+Carbon+Emissions+Forecast%252C+2026-2030
Semiconductor
Industry Association (2024). State of the U.S. Semiconductor Industry. SIA. https://www.semiconductors.org/wp-content/uploads/2024/10/SIA_2024_State-of-Industry-Report.pdf
Shilov, A.
(2023, 10 de Agosto). China
builds exascale supercomputer with 19.2 million cores. Tom’s Hardware. https://www.tomshardware.com/news/china-builds-exascale-supercomputer-with-192-million-cores
Synergy
Research Group. (2025, 19 de noviembre). Cloud Market Share Trends. https://www.srgresearch.com/articles/cloud-market-share-trends-big-three-together-hold-63-while-oracle-and-the-neoclouds-inch-higher
TechInsights
(2025a, 4 de diciembre). Majority of emissions from semiconductor manufacturing
come from the fabrication of die. https://www.techinsights.com/blog/majority-emissions-semiconductor-manufacturing-come-fabrication-die
TechInsights.
(2025b, 4 de diciembre). AI GPU growth directly impacts carbon
emission growth through 2030. https://www.techinsights.com/blog/ai-gpu-growth-directly-impacts-carbon-emission-growth-through-2030
Thompson,
N., y Spanuth, S. (2021). The Decline of Computers as a General Purpose
Technology. Communications of the ACM, 64(3), 64-72. https://doi.org/10.1145/3430936
Top500
Project. (2025). TOP500 Supercomputer Sites. https://www.top500.org/lists/top500/2025/11
TSMC.
(2024). Annual Report 2024. https://investor.tsmc.com/sites/ir/annual-report/2024/2024%20Annual%20Report-E.pdf
United
States Geological Survey. (2024). Mineral Commodity Summaries 2024. U.S.
Department of the Interior. https://pubs.usgs.gov/periodicals/mcs2024/mcs2024.pdf