Mapping the dependencies

So far this site has worked with concepts. This chapter goes to numbers. The aim is not to raise alarm but to give the European CIO, procurement director, or legislator the tables they can calibrate their decisions against. The exposure is structured into three tiers: hardware, cloud and software, and critical inputs (raw materials, energy, intellectual property).

Hardware: where Asia holds the European industry by the wrist

The European chip discussion too often reduces to manufacturing capacity. The actual dependency is more layered. Producing a modern chip requires six separate competencies — design, IP, wafer fabrication, lithography equipment, memory, advanced packaging, and substrates — and Europe holds a different position in each.

Global concentration in the semiconductor supply chain and EU exposure (2026)

Layer	Global leader	Market share	EU position	Type of exposure
Advanced logic, ≤3 nm	TSMC (Taiwan)	~90 % of world production	0 %	Geopolitical (Taiwan Strait)
Advanced logic, 7–5 nm	TSMC + Samsung	~85 %	0 %	Geopolitical
Mid processes, 8–27 nm	TSMC, Samsung, GlobalFoundries, UMC	~75 %	<5 % production in EU	Structural
Mature processes (≥28 nm), capacity	China (SMIC, HuaHong)	~25–31 % of world capacity	~10 % of world capacity	Adversarial (kill-switch risk for automotive, IoT)
Automotive and industrial chips (share of global market by revenue)	Infineon, NXP, STMicro (EU)	—	~30 % of global market	EU-competitive
EUV lithography	ASML (NL)	100 %	100 %	EU strategic asset
Advanced packaging (CoWoS)	TSMC	~95 %	0 %	Geopolitical
HBM3e/HBM4 memory	SK hynix + Samsung + Micron	100 % (3 firms)	0 %	Sold out through 2026
Standard DRAM	Samsung + SK hynix + Micron	~95 %	0 %	Pricing pressure (+45–50 % Q3 2025)
ABF substrates	Japan + Taiwan	~85 %	AT&S (AT)	Structural bottleneck
EDA (design software)	Synopsys, Cadence, Siemens EDA	~70 % (+ U.S. export controls)	Siemens EDA (DE)	Mixed
AI accelerators	Nvidia	~80–85 %	<5 % (SiPearl, Graphcore)	Structural

Methodology. Foundry shares from Counterpoint Research (Q3 2025); leading-edge concentration cross-referenced with TSMC's Q4 2025 SEC 6-K filing. HBM shares from Counterpoint Research (Q3 2025); DRAM shares and the +45–50 % Q3 2025 contract-price figure are TrendForce's published QoQ projection. Mature-node figures span TrendForce's end-2025 estimate (Chinese share >25 % of top-10 capacity) and Digitimes/Mordor projections of ~31 % by 2027. Automotive shares are TechInsights 2024 vendor data — Infineon 13 % + NXP 10 % + STMicro 9 % ≈ 32 % combined. ASML's 100 % EUV share is uncontested. TSMC's CoWoS dominance reflects Morgan Stanley capacity data — Nvidia alone is booked for ~60 % of global CoWoS wafer demand through 2026. ABF film material is a Japanese near-monopoly (Ajinomoto holds >95 %); ABF substrates are geographically concentrated in Japan and Taiwan (Unimicron, Ibiden, Nan Ya PCB, Shinko), with Europe present in this layer through AT&S (Austria). EDA shares from TrendForce 2024 — Synopsys 31 % + Cadence 30 % + Siemens EDA 13 % = 74 %. Nvidia AI accelerator share spans 80–87 % across Bloomberg Intelligence, IDC, and Silicon Analysts FY2026 estimates. Where industry sources disagree, the table presents the consensus midpoint.

Four important patterns emerge from this table.

First: Europe is not "weak in semiconductors" as a whole — Europe is strong in some layers and weak in others. ASML controls 100 % of the world market for EUV lithography. Without ASML no modern chip would be manufactured anywhere in the world, including TSMC, Samsung, and Intel. Infineon, NXP, and STMicroelectronics together hold around a third of the global market for automotive and industrial chips. Siemens EDA (after acquiring Mentor Graphics) is one of the three big players in chip design software. This "invisible advantage" in some layers is leverage that Europe systematically underplays in geoeconomic negotiations.

Second: where Europe is weak, it is deeply and structurally weak. Advanced logic below 7 nm — where every GPU, AI accelerator, and top-tier server processor is produced — has zero presence in Europe. ESMC in Dresden, the flagship of the Chips Act, will fabricate at 28/22 nm and 16/12 nm from late 2027 — three to four generations behind the leading edge. This is not a problem for the automotive sector (which these processes serve fine), but it is a problem for European AI and high-performance computing, which must purchase 100 % of their chips from Asia or the U.S.

Third: the Chips Act target — to double EU world-production share from 10 % to 20 % by 2030 — will be missed with high confidence. In July 2025 Intel formally cancelled its planned mega-fab in Magdeburg (€30 bn investment, ~€10 bn in pre-committed German subsidies) along with the packaging plant in Wrocław. STMicroelectronics and GlobalFoundries cancelled their joint Crolles fab. The European Court of Auditors found in 2025 that despite €86 bn of funding, the EU share remains at 10 %.

Fourth — and most important: HBM memory is a hard supply bottleneck in 2026. Three manufacturers (SK hynix ~57 %, Samsung ~22 %, Micron ~21 %) have capacity sold out through end-2026, and contracts with Nvidia, AMD, and the hyperscalers consume nearly all HBM4 production. Standard DRAM prices rose 45–50 % quarter-on-quarter in Q3 2025. For the European CIO this means: if you intend to deploy AI infrastructure on-premises in 2026, expect extended lead times and significantly higher server costs. Plan procurement 6–12 months ahead.

Cloud and software: structural U.S. dominance

If the EU's hardware exposure to Asia is physical, its cloud exposure to the U.S. is legal and contractual. U.S. providers function technically without issue and typically operate in Europe through Irish or Luxembourg subsidiaries that fall under European jurisdiction for GDPR purposes. The problem is that this structure does not isolate the customer from U.S. legal obligations: the CLOUD Act of 2018 explicitly extends to data held by U.S. firms regardless of where the data sits or which subsidiary technically holds it.

Cloud and enterprise software concentration in the EU (2025)

Segment	Global leader	EU share	European alternative	Realistic EU position
Hyperscale IaaS/PaaS	AWS, Azure, Google Cloud	~70 % of €75 bn market	OVHcloud, IONOS, T-Systems, STACKIT, Aruba	~15 % ("minnows on home turf")
Sovereign cloud (de jure)	—	growing	Bleu, S3NS, STACKIT, OVHcloud	Majority built on licensed U.S. technology
Enterprise productivity	Microsoft 365, Google Workspace	>80 %	OnlyOffice, Nextcloud, Collabora	<5 % in enterprises
CRM	Salesforce	dominant	SAP CX, Sage	Growing but partial
ERP (large enterprises)	SAP (DE)	Leader in EU large-enterprise ERP; ~6 % global ERP applications	—	EU-competitive
Workflow / ITSM	ServiceNow	dominant	—	None
Foundation AI models	OpenAI, Anthropic, Google, Meta	dominant	Mistral (FR), Aleph Alpha (DE)	<10 % global, growing in EU
Cybersecurity (EDR/XDR)	CrowdStrike, Microsoft Defender, Palo Alto	dominant	ESET (SK), WithSecure (FI), Tehtris (FR)	Partial

Methodology. Hyperscaler share and the €75 bn EU IaaS/PaaS figure are Synergy Research Group's H1-2025 update (full-year 2025 projection at +24 % YoY). The ">80 % productivity" line combines Gartner's enterprise-segment figure of 77 % for Microsoft 365 alone with the residual Google Workspace share, and is independently confirmed by the Asterès/Cigref 2025 study, which finds 83 % of EU corporate cloud-and-software spending captured by US vendors (€264 bn/year, ≈ 1.5 % of EU GDP). Salesforce's CRM dominance is IDC 2024 data — 20.7 % global share, #1 in Western Europe specifically, ahead of Microsoft, Oracle, SAP and Adobe combined. SAP's ERP position uses Apps Run The World 2024 vendor data — Oracle and SAP each hold ≈6.5 % of the global $135.9 bn ERP applications market — but no public tracker pins SAP's EU large-enterprise share to a precise number, so the table reports the qualitative position rather than a percentage. Foundation AI shares are Menlo Ventures' end-2025 enterprise survey: Anthropic 40 % / OpenAI 27 % / Google 21 % (≈ 88 % combined), with European providers (Mistral, Aleph Alpha) within the residual 12 % by enterprise spend. The qualitative "<5 % in enterprises" position for Nextcloud, OnlyOffice and Collabora reflects the absence of any tracker that places these suites within the top vendors by enterprise seat-count, and matches Forrester and Gartner analysts' framing of EU public-sector pilots as early-stage rather than at scale.

Key numbers: Synergy Research calculated the European cloud market (IaaS/PaaS/hosted private) at €75 bn annually in mid-2025, of which approximately 70 % goes to three U.S. hyperscalers. European providers — despite tripling their revenue since 2017 — hold 15 %.

Sovereign cloud — what it actually offers. The promises of "sovereign" partnerships (Bleu = Capgemini/Orange + Microsoft, S3NS = Thales + Google, STACKIT = Schwarz Group, T-Systems Sovereign Cloud) are a step in the right direction, but with a fundamental caveat: the majority of these offerings still build on licensed U.S. technology. Data residency in the EU and encryption under European control genuinely limit a U.S. authority's direct access to data, but they do not eliminate dependency on a software stack whose lifecycle, security patches, and long-term maintenance remain in the licensor's hands. Sovereign cloud in 2026 addresses the bit and interpretation layers, but instrumentation dependency remains.

Foundation AI: Mistral as the European pole. Here the picture is changing rapidly. Mistral AI closed a Series C round in September 2025 at €1.7 bn, led by ASML (€1.3 bn) at an €11.7 bn valuation. The company reached approximately $400M ARR in January 2026 (from $20M a year earlier — 20× growth in 12 months). Sixty per cent of revenue comes from Europe. Mistral Compute is currently building 18,000 Nvidia Grace Blackwell GPUs powered by nuclear energy — the largest AI infrastructure in Europe independent of U.S. cloud providers.

That said, the relationship remains radically unequal. OpenAI is valued at ~$300 bn, Anthropic at ~$170 bn — Mistral's €14 bn valuation is 20–30× lower. Mistral is strategically essential and operationally relevant; but a European CIO who would rely exclusively on European top-tier models cannot today match the capability parity of Anthropic Claude Opus 4.5 or OpenAI GPT-5.2 for the most complex agentic tasks. The realistic strategy is hybrid: Mistral for sovereign workloads, regulated sectors, and multilingual European workloads; U.S. models where absolute frontier performance is needed, with appropriate legal and contractual isolation.

Critical inputs: raw materials, energy, IP

The discussion of digital sovereignty too often stops at silicon and code. The actual depth of dependency goes lower.

• Rare earths and critical minerals. China controls about 70 % of world mining and over 90 % of rare-earth processing. Between 2023 and 2025 it imposed export controls on gallium, germanium, graphite, and rare-earth magnets. The EU Critical Raw Materials Act (May 2024) set targets that by 2030 at least 10 % of strategic raw materials should come from EU mining, 40 % from EU processing, and no single third country should hold a share above 65 %. The starting position is far from that.
• Energy for AI. AI workloads are energy-intensive at a scale for which the European energy economy is not dimensioned. Mistral Compute has contracted nuclear energy for its Swedish cluster. The shift in European stance on nuclear power has been consolidating since 2024. The EU SMR Industrial Alliance (October 2024) selected nine Small Modular Reactor projects for European deployment. Czechia signed a contract with KHNP for new units at Dukovany in 2025 — one example of Korea-European industrial partnership. For European data centres, realistic near-term options include lifetime extensions of existing reactors, long-term PPAs between data centres and nuclear operators, and pilot SMR deployments.
• Intellectual property and licensing models. The Instruction Set Architectures (ISA) x86 belong to Intel and AMD; ARM Holdings (owned by SoftBank, registered in the UK) licenses the ARM ISA. RISC-V as an open ISA solves this problem — but switching to a new instruction-set architecture is a generational project, not a quarterly decision. The processors chapter develops this argument.
• Human capital. According to the Atomico State of European Tech 2025, Europe remains a net recipient of tech talent (about +26,000 people in 2024), but this number has roughly halved compared with 2022 (~52,000). Migration between Europe and the U.S. has returned to parity since 2022. The structural pay gap (U.S. AI engineers earn on average 30–70 % more than their European peers) persists, however, and at the senior engineer and technical leadership level, Europe continues to lose net.

This is the terrain map. But terrain alone does not dictate strategy. Europe has two paths: continue to respond mainly through regulation, or deliberately structure partnerships in which its own strengths enable a more balanced negotiating position vis-à-vis partner technologies. The latter is the subject of the partnerships section and the matrix of regional self-sufficiency.