| Direction of travel | |||
| net +9 | |||
|
|
|||
|
▲
15
Towards autonomy
|
▼
6
Towards dependence
|
◆
3
Both directions
|
■
6
No movement
|
| 30 developments | |||
What Europe did this month to depend less on everyone else, and where it went the other way.
Why This Newsletter Exists
European technology sovereignty is argued about constantly and recorded almost nowhere in one place.
This publication is the aggregation. Once a month: what moved across European technology, industry, infrastructure and defence, where the money went, and which way each thing pushed Europe's dependence on suppliers, systems and law it does not control. Every claim is traced to the document it came from and marked with how firmly it stands up. Where something could not be established, that is stated too, because in this subject the gaps are frequently the finding.
Summary for July and August
Both halves of that are literal. Roughly €13bn of European public and quasi-public money moved or opened for bids in this window, and exactly one procurement landed on a European vendor. EuroHPC opened the largest pool of public compute money in European history, up to €10bn of public backstop against more than €30bn total, and attached no European-processor condition to it. Eight days earlier it had signed an €80m contract for 1,008 Nvidia GPUs. Meanwhile the space and defence side went the other way hard: IRIS², Europe's answer to Starlink, was expanded to 348 satellites and pulled forward to 2029, and SAFE defence loan money kept moving, with Greece drawing its first tranche. The pattern across the month: where a European supplier exists, European money reaches it. Where one does not, the money buys the capability from whoever has it, and no condition is attached requiring anything else.
1. Policy and Law
▲ 2 ▼ 3 ◆ 0 ■ 5 net −1
Summary. The two big sovereignty laws proposed in June are moving slowly and nothing binding exists yet. The one law that did land went the other way: the AI Act's hardest rules were postponed by up to two years. A European preference for public procurement, the single most direct tool available, exists only as a leaked draft.
▼ The AI Act's high-risk rules were pushed back, and it is now law. Regulation (EU) 2026/1744 of 8 July moves standalone high-risk AI obligations from 2 August 2026 to 2 December 2027, and product-embedded high-risk AI to 2 August 2028. In force since 27 July. Why it matters: the rulebook that was meant to set the global standard just gave European industry between 16 months and two years of breathing room instead.
SOURCE VERIFIED · EUR-Lex
▲ The transparency rules did not get delayed and started applying on 2 August. Chatbots must disclose they are machines, AI-generated content must be machine-readably marked, deepfakes must be labelled. The Commission published its guidelines on 20 July. Why it matters: this is the part of the AI Act that now actually binds anyone deploying AI in Europe.
SOURCE VERIFIED · Commission guidelines
■ The June Tech Sovereignty Package has four parts, and two get forgotten. Chips Act 2.0 and the Cloud and AI Development Act take all the attention. The package also contains an Open Source Strategy and a Strategic Roadmap for Digitalisation and AI in the Energy Sector, both of which carry procurement consequences and neither of which is being tracked anywhere prominent.
SOURCE VERIFIED · Commission, 3 June 2026
■ Chips Act 2.0 got as far as officials reading it line by line. The Council's industry working party finished Pillar I on 8 July and continued Pillar II on 14 July, aiming to draft a first compromise text in the autumn. Parliament has not appointed a rapporteur. Why it matters: nothing binding on European chip capacity will exist before late 2027 at the earliest.
SOURCE VERIFIED · Council WK 10387/2026
SOURCE VERIFIED · Procedure file
■ The Cloud and AI Development Act has a rapporteur, according to trade press. Reinier van Lanschot (Greens/EFA, Netherlands) is reported to have taken the file on 7 July. ⚠️ Trade press only. Parliament's procedure file still shows the referral pending before a joint committee, and its Key players section records no rapporteur, group or appointment date. The file's own transparency section, however, has listed van Lanschot as IMCO rapporteur on dated meetings since 12 May, eight weeks before the reported 7 July appointment and three weeks before the proposal itself was published. Why it matters: CADA would create four legal sovereignty levels for cloud services and force public buyers to require them, so whoever holds the pen decides how hard Europe pushes non-EU providers out of government IT.
SOURCE VERIFIED · Procedure file
SOURCE VERIFIED · Full text
■ SEAL is a scoring tool, not a law, and it is widely described wrongly. SEAL stands for Sovereignty Effectiveness Assurance Level, not "Sovereign European Assurance Level" as commonly repeated. There are five levels, SEAL-0 to SEAL-4, and a service is scored separately against each of eight sovereignty objectives, SOV-1 to SOV-8. The framework awards no single SEAL grade, though it does compute a composite Sovereignty Score alongside the objective scores. It was published in October 2025 as procurement guidance. Why it matters: buyers are already being scored against it, and a supplier claiming to "be SEAL-3" is describing something the framework does not issue.
SOURCE VERIFIED · Cloud Sovereignty Framework v1.2.1
▼ The Quantum Act has slipped and no new date exists. It was scheduled for Q2 2026 in the Commission Work Programme and reconfirmed for Q2 2026 in the April 2026 roadmap. It had not been tabled as of 10 August. No Commission, Council or Parliament document assigns it a replacement date, and the "Q3 2027" figure in circulation is the target for co-legislator agreement, not for the proposal. Why it matters: quantum procurement decisions are being made now against a rulebook that does not exist.
SOURCE VERIFIED · Parliament Legislative Train
▲ €659m of German chip subsidies cleared state aid on 14 July. Four first-of-a-kind plants: Element 3-5 (€353m, silicon carbide epi-wafers), Vishay Siliconix Itzehoe (€214m, power transistors), KLA-Tencor (€74.4m, metrology), KETEK (€17.9m, detectors). Why it matters: this is where chip strategy stops being a document and becomes a factory, and notably it funds niche links rather than another headline mega-fab.
SOURCE VERIFIED · Commission IP/26/1585
■ A European preference for public procurement is drafted but not tabled. Reporting on 9 July described a Commission regulation expected in September 2026 allowing buyers to exclude foreign suppliers on security grounds and apply an optional European preference across energy, water, transport and postal procurement. ⚠️ Trade press only. No such document appears on EUR-Lex or the Council register.
SOURCE REPORTED · Euronews
▼ The data half of the Digital Omnibus stalled before summer. No Council position, no finished Parliament report. Why it matters: the Data Act's cloud switching rules, including the removal of switching fees due by 12 January 2027, are tangled up in it, and those are the main legal lever for actually leaving a US hyperscaler. ⚠️ Reported blocking minority could not be confirmed at a Council source.
SOURCE REPORTED · EDRi
2. Compute and Infrastructure
▲ 2 ▼ 2 ◆ 2 ■ 0 net 0
Summary. The biggest month for European compute money on record, and the terms attached to it decide almost nothing about who supplies the hardware. Four separate EuroHPC procurements opened or signed in this window. One of them specifies a European supplier. It is the quantum one.
▼ EuroHPC opened the AI Gigafactories call on 30 July with no European silicon requirement. Up to seven sites across at least seven member states, bids due 12 November 2026, selection early 2027. Up to €10bn public, more than €20bn expected private, more than €30bn total. Why it matters: the largest pool of European public compute money ever opened, and on current terms most of it clears through Nvidia and AMD.
SOURCE VERIFIED · EuroHPC
SOURCE VERIFIED · Commission, for the funding figures
▼ EuroHPC and Luxembourg bought 1,008 Nvidia GB200 GPUs for €80m on 22 July. MeluXina-AI, built by Italian integrator E4 with Dell as subcontractor, 252 nodes carrying 1,008 GPUs in total, over 20 AI exaflops. EuroHPC funds half the cost through the Digital Europe Programme. Why it matters: the clearest case in the window of European public money buying non-European hardware. Europe gets the integration and the building; the chip, the interconnect and the software stack are American.
SOURCE VERIFIED · EuroHPC
▲ A European-silicon gigafactory bid appeared at Strasbourg on 8 July. The ÆTHER consortium (SiPearl, 2CRSi, Axelera AI, Dassault Systèmes/OUTSCALE, Socomec, Viridien, Equans, ÉS Group and others) announced two sites targeting 42 MW rising towards 400+ MW, built on SiPearl high-performance processors and Axelera accelerators. The Grand Est region and the Eurométropole de Strasbourg back the bid as local authorities rather than as consortium members. Why it matters: the only gigafactory bid putting European compute silicon at the centre. Everything depends on whether it is selected in early 2027 and whether SiPearl's Rhea1, due at the end of 2026, ships on time. ⚠️ No investment figure disclosed. The consortium document names no processor generation; the Rhea1 date comes from SiPearl.
SOURCE VERIFIED · 2CRSi
SOURCE VERIFIED · SiPearl
◆ Microsoft made a multibillion dollar commitment to Mistral's European GPU capacity on 21 July. Mistral is adding capacity "drawing on thousands of the latest NVIDIA Vera Rubin GPUs" and Microsoft will use part of it to increase its own European AI capacity; Mistral models ship through Microsoft Foundry and Copilot Studio. The release quantifies the commitment no further than "multibillion dollar". Why it matters: a European lab is the landlord and an American hyperscaler the tenant, which inverts the usual arrangement and gives Mistral demand certainty without dilution. European ownership covers the site and the operating company. It does not reach the accelerators, and Mistral's distribution now runs through Azure.
SOURCE VERIFIED · Microsoft
▲ Orange and Morrison announced a €3bn French datacentre joint venture on 27 July. Five facilities across four campuses, targeting 400 MW, roughly 10x Orange's current capacity, explicitly framed as sovereign. Why it matters: European-operated capacity on a low-carbon grid. Neither the tenants nor the hardware inside are specified anywhere in the announcement.
SOURCE VERIFIED · Orange
◆ Two sovereign clouds launched five days apart, meaning two different things. Atos launched Atos Sovereign Cloud on 23 July, designed and engineered in the EU. Zscaler and Schwarz Digits announced a sovereign security platform on 28 July, hosted in German datacentres and operated by STACKIT. The release carries no ownership or intellectual-property language, and it does not need to: the platform is Zscaler's Zero Trust Exchange, and Zscaler is American. Why it matters: both use the word sovereign. In one case it describes where the engineering was done, in the other where the servers sit while American software runs on them. European policy has not settled which of these it means, and this is what that unsettled question looks like in practice.
SOURCE VERIFIED · Atos
SOURCE VERIFIED · Zscaler and Schwarz Digits
3. Defence and Space
▲ 6 ▼ 1 ◆ 0 ■ 0 net +5
Summary. The strongest month of the year for European strategic autonomy, and it happened in orbit and in defence industry rather than in computing. IRIS² became a contracted build. SAFE money actually moved. The counterweight is that when a member state needed interceptors this month, it bought American.
▲ IRIS² expanded to 348 satellites and pulled forward to 2029. The Commission and the SpaceRISE consortium (SES, Eutelsat, Hispasat) signed the implementation agreement on 7 August, adding 66 satellites as a dedicated defence, security and emergency layer, raising secure governmental capacity inside the EU by 60%. Why it matters: this is the programme meant to end European dependence on Starlink for secure government connectivity, and this month it moved from agreement to signed implementation. Two caveats. 2029 is when launches begin, not when the system is complete; services deploy progressively from that point. And the cost rose from €10.55bn to a reported €15.6bn. Whatever the finish date, the dependency on commercial providers runs at least three more years. ⚠️ The €15.6bn total is trade press; official releases do not state it. The two official sources also disagree on when negotiations closed: ESA records 6 August, the Commission 7 August.
SOURCE VERIFIED · Commission
SOURCE VERIFIED · ESA
▲ Poland's IRIS² commitment reached €656m, up from the up-to-€500m it pledged in September 2025, joining Hungary (€500m) and Spain (€1.6bn to €2bn) as national co-investors. Why it matters: Poland is the most Starlink-exposed large member state. A frontline government putting its own satcom budget into the EU system rather than buying more commercial capacity is the clearest available signal of whether IRIS2 will actually displace the incumbent.
SOURCE VERIFIED · ESA
▲ The Commission proposed the first five European Defence Projects of Common Interest on 3 July. Drones and counter-drone, maritime and seabed, space, air and missile defence, eastern flank security. An average of 18 member states each, €325m allocated under the €1.5bn European Defence Industry Programme, stated combined ambition near €190bn by 2036. Not yet law: the Council must formally establish them.
SOURCE VERIFIED · Commission
▲ Greece received its first SAFE payment, €118.2m, on 23 July. 15% of its €787.7m allocation under the €150bn loan instrument. It is Greece's first, not the instrument's: Poland drew the first SAFE payment, €6.6bn, on 29 May. Why it matters: SAFE carries a European content requirement, capping components originating outside the Union, the EEA EFTA States and Ukraine at 35% of the estimated cost of the components of the end product, so disbursed money is procurement steered away from third-country primes.
SOURCE VERIFIED · Commission, Greece
SOURCE VERIFIED · Commission, Poland
SOURCE VERIFIED · SAFE Regulation, Article 16(10)
▼ NATO's procurement agency awarded Raytheon a Patriot interceptor contract for Poland on 7 July. Why it matters: the clearest decision in the window deepening dependence on a US prime for a critical munition, and it sits directly against the SAFE and EDIP direction. Neither NSPA nor Raytheon disclosed a value. Poland's own agreement with NSPA for the same missiles was reported at $988m for several hundred interceptors, with deliveries running to 2031, which is a different leg of the same procurement. ⚠️ The $988m figure is trade press. In the same week the Netherlands announced over €3bn of packages built substantially on US-origin systems. ⚠️ Netherlands figure is trade press. ⚠️ The NSPA release blocks automated retrieval; this item was checked against a verbatim mirror and two trade reports rather than read at the source.
SOURCE REPORTED · NSPA
▲ AGILE, a €115m rapid defence innovation programme, reached political agreement on 15 July. Expected to be operational from 1 January 2027, delivering to member states and associated countries only, which means EEA/EFTA states and Ukraine, and aimed at startups and SMEs rather than incumbents. The budget is small. The delivery restriction is the significant part, because it shapes the market a programme aimed at firms that have never sold to defence is building for.
SOURCE VERIFIED · Commission
▲ The EU-Ukraine Drone Alliance launched in Kyiv on 17 July with 18 founding members, nine EU companies and nine Ukrainian firms. Why it matters: drones are where European primes are furthest behind and Ukrainian firms are furthest ahead. No funding figure attached yet.
SOURCE VERIFIED · Commission
4. Industry, Investment and R&D
▲ 5 ▼ 0 ◆ 1 ■ 1 net +5
Summary. European hardware had a genuinely strong month, in lithography, in power electronics and, in Barcelona, in first silicon. The first two sit upstream and downstream of the AI accelerator, and neither replaces it. European defence tech raised record money, mostly from American investors.
▲ ASML posted €9.3bn quarterly sales on 15 July and is adding 30% to its low NA EUV capacity for 2027. Net income €2.9bn, full-year guidance €43bn to €45bn, against a 2026 low NA capacity of around 65 systems, with a further 30% under investigation for 2028 and a parallel 30% on DUV immersion. Why it matters: ASML is Europe's only genuine chokepoint in the semiconductor chain, and expanding capacity expands that leverage. Worth holding alongside it: almost none of the chips produced on ASML machines are consumed by European compute vendors.
SOURCE VERIFIED · ASML
▲ Infineon posted record revenue on 5 August, driven by AI datacentre power silicon. €4.172bn for the quarter, 19.1% segment margin, guidance now around €16.3bn, where the previous outlook was directional rather than numeric, €2.7bn capex including a fourth Dresden module. Why it matters: European silicon winning real revenue inside the AI buildout, manufactured on European fabs. The product is power delivery for datacentres. It sits alongside the accelerator rather than competing with it, so it grows with Nvidia's shipments instead of displacing them.
SOURCE VERIFIED · Infineon
▲ Openchip showed first silicon of BER10 on 30 July, the same day the AI Gigafactories call opened. A Linux-capable 64-bit RISC-V processor with four out-of-order cores and a wide vector unit, on a sub-2nm Gate-All-Around process, taped out in 2025 and delivered 2.5 years after the company started from scratch. Why it matters: a European team now designs application-grade silicon for the most advanced process node in production, and no European foundry offers that node. The design capability moved; the fab dependence did not. The release names no foundry, volume-production date, customer or independent benchmark, and its own wording is "designed to be fully functional". Worth recording: the Spanish version of the same release drops that hedge and states «totalmente funcional»; the English text is the one quoted here.
SOURCE VERIFIED · Openchip
◆ Helsing raised $1.8bn at an $18bn valuation on 13 July, and Quantum Systems raised $1.2bn at roughly $8bn on 2 July. Why it matters: European defence AI now has capital at competitive scale. Both rounds lean on non-European institutional money, with Blackstone, Advent, Fidelity and Wellington among the Quantum Systems backers and CPP Investments among Helsing's, alongside European names including Airbus, Balderton and HV Capital. Helsing's own statement that it remains predominantly European-owned is a claim about voting control rather than about where the capital came from. ⚠️ The description of these as Europe's largest defence technology rounds is trade press; neither release makes the claim.
SOURCE VERIFIED · Helsing
SOURCE VERIFIED · Quantum Systems
▲ Airbus anchored a €500m European defence tech fund on 21 July. E2D was launched in June 2026 by AVP and Earlybird and targets growth-stage companies across both defence-native and dual-use applications, around twenty of them at roughly €25m each. Airbus Defence and Space came in as anchor investor on 21 July. Why it matters: it attacks exactly the gap above, the absence of European growth-stage defence capital. It is smaller than a single Helsing round, and it is European institutional money.
SOURCE VERIFIED · Airbus
▲ Finland's LUMI AI Factory chose IQM, a Finnish supplier, for its quantum system on 8 July. LUMI-IQ, an IQM Halocene H4 with 150 qubits in the first 2027 system, funded by EuroHPC with Finland, Czechia, Norway and Poland. Why it matters: the one procurement in this window where European public money bought European hardware. Quantum is also the only compute domain where European suppliers currently compete at the frontier, which is likely the reason. Contract value not published as an absolute figure. IQM's prospectus of 1 July states it is approximately equal to the company's total revenue for the year ended 31 December 2025.
SOURCE VERIFIED · IQM
■ EuroHPC opened two more tenders: a semiconductor spin-qubit quantum computer in Luxembourg with an acquisition budget of up to €11.95m (bids 1 September) and an industrial supercomputer at CINECA Bologna with a total project budget of €11.6m, of which the EU contributes €4.07m (bids 8 September). The two figures are not like for like: one is an acquisition budget, the other includes industry money. Neither specifies a European supplier requirement.
SOURCE VERIFIED · MeluXina-Q
SOURCE VERIFIED · INNOVATE Bologna
⚠️ Not verified: a reported €3bn Mistral Series D at a €20bn valuation, said to have closed 21 July led by EQT's Scaleup Europe Fund. Mistral's own newsroom shows no funding announcement in the window; its last posted round is the €1.7bn Series C of September 2025. If real it would be one of the largest items of the month, which is exactly why it is not being reported here as fact.
5. The Month in Numbers
Across all four sections: ▲ 15, ▼ 6, ◆ 3, ■ 6. Net +9 across 30 tracked developments.
The count is unweighted, so a €10bn call and a single press release score the same. That is why net +9 sits oddly beside a month defined by €13bn of public money reaching one European supplier.
The weighted figure would mean more, and this month it cannot be computed precisely: IQM's contract value was published only by reference to the company's 2025 revenue rather than as a figure, so the European share of European public compute spending can be approximated but not stated. From issue.01 these tallies become a series.
6. Tools and Resources
Everything cited here is a public document, and the registers those documents live in do not change from month to month. They sit on one page now rather than being reprinted in every issue: sources and databases, which carries the legislative trackers, the procurement and funding databases, the compute and defence registers, and the two datasets this publication maintains itself, the European Tech Sovereignty Stack map and the EU Sovereignty Ledger.
One place also means one correction. A register that moves, or a description that turns out to overstate what the register actually covers, is fixed once instead of in every back number. Five descriptions were narrowed that way before this issue went out, and a description that overstates its register is the kind of error a reader acts on and then cannot reproduce.
Definitions used here
Sovereignty is not one property, which is why two products launched in the same week can both be called sovereign and mean opposite things. This newsletter separates four questions: who owns it, who operates it, whose law governs it, and who made the parts. Most European instruments test the first three and stop before the fourth.
Method
Every claim is checked at a primary document where one exists. Where only trade press exists, the item says so. Figures are not repeated from press summaries when the underlying document is available. Corrections run in the next issue with the original claim quoted, never edited silently.
This newsletter covers instruments, programmes and market structure. It does not report private financings in motion, including ones already named in press.
Corrections and tips: reply to this email.
This is issue.zero, the pilot. The format is not settled and it will change at Issue 01. If something here is wrong, or something obvious is missing, that is what this issue is for.
|
Next issue at the start of October. Corrections run at the top of it. Reidar Balstad |