InitCompany — coming soon
The Vanguard Manifesto — show / hide

Germany is losing an industrial economy and has not decided what to build with the pieces.

Between now and 2035, more than 110,000 job cuts have already been announced across carmakers, suppliers, steel, and chemicals, on published schedules, before Volkswagen's Future Plan 2030 of September 2026 added a further 50,000 group-wide. Count the supplier chains, logistics, and regional service economies that depend on those plants and the realistic exposure is 200,000 to 300,000 jobs. A large share are not line workers. They are development engineers, R&D staff, software teams, and internal IT, which means the country is releasing engineering capacity at scale at the exact moment it says it wants to build sovereign AI infrastructure.

Everything that build requires already exists. The capital is committed: €500 billion over twelve years through the infrastructure fund, with a national strategy targeting quadrupled AI capacity by 2030. The engineers exist and are being let go. The robotics industry exists, ranks among the most capable in the world, and is watching its largest customer contract. And the closing plants hold the one thing money cannot buy quickly: live grid interconnection, in a country where connection queues run for years and the main hub has no capacity left at all.

This is what a digital revolution looks like when a country builds it out of what it is losing rather than out of something new. Nothing has to be invented. The parts have to be pointed at each other.

Sovereignty is the part worth getting right, and it is winnable. Europe is already funding an open processor architecture through RISC-V, and the engineers who can build a trust chain on it, people who have spent careers on embedded systems, functional safety, and hardware that must not fail, are the same ones now leaving German industry. Open silicon, auditable software, and keys held where no foreign order can reach them are all buildable today. The window is open now, and it is the good kind of problem: not whether Germany has what this takes, but whether it decides to use it.

Germany's digital revolution, built out of what industry is losing

I. Three things are true at once

Germany is in its third consecutive year of recession, which is without precedent in the postwar record. EY's analysis put industrial job losses at roughly 124,000 in 2025 alone, about 50,000 of them in automotive, and the fourth quarter of 2025 marked the tenth consecutive quarter of shrinking industrial revenue. Since 2019 the sector has shed on the order of 266,000 positions, with automotive down about 111,000, a 13 percent contraction. A survey by the Institute of the German Economy found four in ten industrial companies planning further layoffs in 2026.

At the same time, Germany has committed €500 billion over twelve years through the Special Fund for Infrastructure and Climate Neutrality, established by a two-thirds constitutional amendment in March 2025 and exempt from the debt brake. Digitalisation is one of its seven named target areas, and about €58 billion flows through the fund in 2026 alone.

And on 18 March 2026 the federal cabinet adopted a National Data Centre Strategy targeting a doubling of overall data centre capacity and a quadrupling of AI and HPC capacity by 2030, support for at least one AI gigafactory, and expansion across the stack from microelectronics to software.

Money, mandate, and released skilled labour, all inside the same eighteen months. Nothing about that combination is permanent.

Germany's digital revolution, if it happens, will not look like the ones it is usually compared to. There will be no founding garage and no breakthrough product. Nothing in it needs to be invented: the capital exists, the grid connections exist, the engineers exist, the robots and the people who build them exist, the legal framework exists. What is missing is the decision to point them at each other. A digital revolution accomplished by reallocation rather than by creation is less romantic than the American version and considerably better suited to what Germany actually has.

II. The ledger: what 2026 and 2027 actually release

The abstract figures obscure what is really happening, which is a set of dated, negotiated, already-signed programmes with known end points. Reading them as a schedule rather than as a tragedy changes what can be planned around them.

BMW confirmed on 29 July 2026 that it will cut roughly 8,000 positions globally, about 5 percent of its workforce, through the largest voluntary redundancy programme in the company's history. The programme runs from October 2026 to the end of 2027 and is expected to save around €1 billion annually from 2028. The detail that matters most is the scope: it targets administrative, development, and corporate functions and explicitly excludes production. Around 40,000 of BMW's roughly 85,000 German permanent employees will be eligible for redundancy offers from October, concentrated in Munich, Regensburg, Dingolfing, and Leipzig. BMW's Munich research centre, the FIZ, employs roughly 25,000 engineers, developers, designers, and specialists, and administrative and development roles there are expected to absorb the bulk of the reduction.

Volkswagen agreed in December 2024 to cut at least 35,000 jobs at its core brand by 2030 and had eliminated about 15,000 of those by the end of 2025. On 3 September 2026 its Supervisory Board unanimously approved the Future Plan 2030, which goes further: beyond the existing programmes, a group-wide workforce adjustment of approximately 50,000 positions including management roles; European capacity that exceeds demand by more than 500,000 units; no competitive production allocation currently securable for the Emden, Zwickau, Hanover and Neckarsulm plants from 2031 to 2034, with alternative uses for them under assessment; and a model portfolio streamlined by around half by 2035. The 50,000 is a group-wide, global figure; the German share is not stated. Audi announced 7,500 German job cuts by 2029, mainly in administration and development. Porsche has raised its total announced reduction to roughly 9,400 positions, around a fifth of its workforce, by 2035. Mercedes-Benz is running its own voluntary redundancy programme.

On the supplier side the numbers are larger relative to headcount. Bosch announced 13,000 cuts at its mobility division in September 2025 to close a €2.5 billion cost gap, alongside earlier programmes: around 600 of 750 Hildesheim positions go by the end of 2026, and up to 1,300 at Schwäbisch Gmünd between 2027 and 2030. ZF plans up to 14,000 German job cuts by the end of 2028, including 7,600 in its electrified drivetrain division, while targeting more than €500 million in savings by 2027. Continental is cutting 3,000 automotive research and development jobs by the end of 2026. Schaeffler is removing 4,700 positions across 2025 to 2027, about 2,800 of them in Germany across ten sites. Ford plans 2,900 German cuts by 2027, roughly one in four jobs at its Cologne plant.

Outside automotive, thyssenkrupp Steel is restructuring 11,000 of its 27,000 steel positions, with 5,000 eliminated and 6,000 outsourced, and has pulled its Bochum closure forward to 2027. Evonik will cut 3,200 jobs between 2027 and 2029, 2,150 of them in Germany, on top of 2,800 already scheduled through the end of 2026, and will close its Witten site in 2027. BASF raised its annual cost savings target to €2.3 billion and is cutting its internal IT division while establishing a new hub in India.

That last item deserves a pause. At the exact moment the federal government is funding digital sovereignty, one of Germany's largest industrial employers is moving its internal IT capability out of the country. Sovereignty policy and corporate cost discipline are currently pulling in opposite directions, and nobody is reconciling them.

Two observations about the ledger as a whole. First, most of these are voluntary, negotiated, and phased over two to four years, which means the release is predictable rather than sudden. That is the difference between a labour shock and a planning input. Second, and contrary to the usual assumption, a large share of the 2026 and 2027 wave is white-collar: development, administration, R&D, software, IT. BMW, Audi, Continental, Cariad, Evonik, and BASF are all cutting engineers and specialists, not line workers.

III. The bottleneck is the interconnection queue

The instinct is to read a closing plant as a real estate opportunity: cheap land, a large shell, industrial zoning already in place. That undervalues it badly. The scarce asset on an old industrial site is the grid connection.

The IEA reports that grid connection wait times inside the EU run from two to ten years depending on the country, and that in the established FLAP-D hubs the queue averages seven to ten years. ACER put direct grid congestion costs at €4.3 billion in 2024, excluding the knock-on cost of delayed projects. Frankfurt is the sharpest case: data centres there account for as much as 40 percent of municipal electricity demand, available capacity for new projects inside the city is effectively zero, high-voltage connection lead times run 24 to 36 months, and grid operators have signalled no new utility-scale substations before the second quarter of 2027. Operators are moving out into Hanau, Hattersheim, and the wider Rhine-Main region because there is no delivery path left inside the city.

Against that, an industrial site carrying live high-voltage interconnection, water rights for process cooling, rail and road access, and an existing industrial designation is not a discounted warehouse. It is a position in a queue that money alone cannot jump.

The clearest recent demonstration is American. TeraWulf bought a former aluminium plant in Hawesville, Kentucky for $200 million in February 2026 and leased it to Anthropic on a twenty-year agreement, and the reporting is explicit that the purchase was for the grid connection on the property rather than for the buildings.

Two qualifications belong here, because this is where the argument usually gets oversold. Brownfield reuse compresses the interconnection wait, not the construction schedule: that same Hawesville site is not expected to bring initial capacity online until the second half of 2027, with full buildout in early 2028. And historic load rights do not automatically convert into data centre capacity. Utilities can require new studies, protection equipment, upgrades, or a fresh queue position, and a retired industrial site is only a shortcut once the grid path is confirmed. Contamination, demolition cost, unknown subsurface conditions, and legacy easements are all real risks. The German strategy's proposed move to a "first ready, first reserved" allocation model matters precisely because it changes who wins that race.

The narrow claim survives all of that: reusing industrial interconnection is the fastest available path to sovereign compute capacity in a country where the grid is the binding constraint.

IV. Where these people actually go

The German retraining debate keeps offering a false choice between preserving the old jobs and moving workers into services they have no relationship to. The ledger in section II points at something more specific, and it splits cleanly into two tracks that require completely different handling.

Track one: critical facilities

This is the track most people already imagine, and the labour data supports it. Randstad's analysis of 50 million job postings between 2022 and 2026 found demand for cooling and HVAC system engineers up 67 percent, industrial automation technicians up 51 percent, robotics technicians up 107 percent, and electricians up 27 percent. The Uptime Institute has found 53 percent of data centre operators reporting difficulty staffing roles, most acutely at the experienced electrician level where industrial electrical work meets critical facilities work. In Germany the electrician trade has topped shortage occupation rankings for years, and BDA and IW forecasts project up to 768,000 missing workers by 2028.

The people this fits are the ZF, Schaeffler, Ford, and thyssenkrupp cohort: industrial electricians who have worked three-phase distribution, mechatronics and maintenance technicians from automated production lines, plant cooling and process engineers, high-voltage specialists. The gap between what they know and what a data centre needs is UPS systems, building management and power monitoring systems, rack-density thermal behaviour, concurrent maintainability, and change control under a 24/7 uptime regime. That is a bridging problem measured in months of structured training, not a career reinvention. Germany already has the institutional machinery for exactly this: the Handwerkskammer and IHK certification system, works council agreements, and Transfergesellschaft structures that can be pointed at a defined qualification rather than at generic job search support.

Track two: assurance and platform engineering

This track is the one almost nobody is discussing, and it is where the 2026 and 2027 wave actually lands hardest.

A BMW FIZ development engineer, an Audi administration and development specialist, a Continental automotive R&D engineer, a Cariad software developer, a BASF internal IT specialist: none of these people is going to accept a facilities technician role, and nobody should be designing a programme that assumes they would. Their salary expectations, seniority, and skill profile all point somewhere else.

But consider what the German automotive engineering corps is actually trained in. Functional safety under ISO 26262, with hazard analysis, safety cases, and argued assurance rather than assumed assurance. Requirements traceability from specification through implementation to verification. Validation and test engineering under regulatory scrutiny. Embedded systems, firmware, deterministic real-time behaviour, and hardware-adjacent debugging. Supplier qualification and component provenance across deep multi-tier supply chains. Documentation discipline sufficient to survive a type approval audit.

That is an almost exact description of what regulated digital infrastructure now requires and cannot find. The EU Cyber Resilience Act, the AI Act's conformity assessment regime, NIS2, and the assurance layer of any sovereign infrastructure programme all demand argued, evidenced, auditable safety and security cases produced by people who think in hazard analysis and traceability. The software industry has historically been bad at this. German automotive engineering has been doing it under regulatory obligation for two decades.

The same applies at the hardware layer. Measured boot, remote attestation, secure element integration, and firmware trust chains are embedded systems work. An engineer who has debugged a CAN bus timing fault or qualified an automotive-grade microcontroller is closer to that work than a typical web developer will ever be. If Europe is serious about a RISC-V trust chain, the people who can actually build the attestation path are currently being made redundant in Munich and Ingolstadt.

The bridging gap here is real but narrow: Linux and container fundamentals, cryptographic key management, cloud-native operational patterns, and the specific regulatory vocabularies. Twelve to eighteen months for a senior engineer, and considerably less for the strong ones.

What this does not solve

Three honest limits, because the version of this argument without them is propaganda.

Data centres are not labour-intensive at steady state. A large facility employs a few hundred people in operations against a capital base comparable to a plant employing thousands. The construction and commissioning phase absorbs far more labour than the operating phase, and that phase ends. Nothing in this proposal absorbs 14,000 ZF positions and 13,000 Bosch positions on a one-for-one basis, and claiming otherwise would be dishonest. The realistic framing is that digital infrastructure absorbs a meaningful minority of the released workforce into higher-value work, and that this is worth doing on its own terms rather than as a complete answer to industrial unemployment.

Voluntary redundancy self-selects against retraining. BMW's programme is voluntary, and the people most likely to accept a severance package are those near retirement or those who already have somewhere to go. The cohort genuinely available for redirection is smaller and younger than the headline numbers suggest.

And this reallocates a shortage rather than resolving one. Data centres compete for exactly the same electricians and HVAC technicians that factories, grid operators, and the heating transition are already fighting over. What still makes the case is that a laid-off automotive electrician or a redundant FIZ engineer is currently being lost to the labour market entirely, and the destinations described here are closer to their existing skills than anything else on offer.

V. The robotics asset is being stranded with the cars

There is a second stranded capability alongside the workforce, and it gets almost no attention in this debate because it looks like a manufacturing story rather than a digital one.

Germany is the most automated economy in Europe and among the most automated in the world. The IFR's World Robotics 2025 report puts German robot density at 449 units per 10,000 manufacturing employees, third worldwide behind South Korea and Singapore, growing about 5 percent annually since 2019. Nearly half of Europe's 821,384 operational robots sit in Germany, and Germany accounts for roughly 32 percent of annual European installation volume.

What makes this different from a country that simply buys a lot of robots is that Germany also builds them. KUKA, Siemens, Bosch Rexroth, and several hundred specialised automation firms make Germany both a major consumer and a major producer, which is a combination almost nobody else has at this scale.

And that capability is now tied to a shrinking customer. Automotive accounts for roughly 42 percent of German industrial robot installations and remains the leading client sector. Global industrial robot installations have shown no meaningful growth since 2021, China now accounts for more than half of worldwide deployments, and the VDMA Robotics and Automation association warned in February 2026 that Germany is losing market share. The domestic demand base for German robotics is the same automotive sector described in section II, contracting on a published schedule through 2029.

Meanwhile a new customer class is appearing, and it is appearing in Germany. Data centre robotics was estimated at $13.7 billion in 2024 and is projected to reach $44.2 billion by 2030, a compound annual growth rate above 21 percent. Google demonstrated mobile robots for data centre management at the 2024 Open Compute Summit, handling material movement, media management, and rack servicing. Operators are moving toward hybrid and lights-out models where robots run unsupervised for part of the day. The applications are unglamorous and well suited to existing German competence: rack and equipment transport, cabling, thermal and environmental inspection patrols, predictive maintenance, drone-based inspection, security, and automated media handling.

The clearest signal is already on the ground. At Deutsche Telekom's Munich AI factory, the roughly 75 kilometres of fibre optic cable connecting the GPUs and the site were laid partly by robots from Agile Robots, a Munich company that is also part of the sovereign AI ecosystem forming around that facility. That is a German robotics firm building German AI infrastructure, and it is the template rather than the exception.

The demand-side argument is therefore symmetrical with the labour argument in section IV. German robotics integrators are losing their largest customer exactly as a new one emerges domestically at industrial scale, and the engineering skills involved, precision motion control, safety-rated automation, industrial networking, and machine vision, transfer almost directly.

One contradiction has to be stated rather than glossed. Robotics reduces data centre operational headcount, which cuts directly against the employment argument. A lights-out facility employs very few people, and any honest version of this manifesto has to admit that automating data centre operations makes the jobs case weaker, not stronger. The resolution is not to pretend the tension away. It is that the durable German position is in building, integrating, and maintaining the robots rather than in supplying the labour they displace. That is the higher-value side of the transaction, and it is the side Germany is already equipped for. A country with KUKA and Bosch Rexroth should be selling the automation to the world's data centres, not competing to staff its own.

VI. Jurisdiction is the first sovereignty axis

Here is where the opportunity gets squandered if it is handled carelessly. Building the shells with German public money and filling them with foreign-controlled stacks under foreign jurisdiction would solve part of the employment problem while giving away the sovereignty problem in the same transaction.

The legal exposure is on the record under oath. In sworn testimony to a French Senate inquiry into public procurement and digital sovereignty in mid-2025, Anton Carniaux of Microsoft France was asked whether he could guarantee that data belonging to French citizens held under public contracts would never be transferred to US authorities without French authorisation. He answered that he could not guarantee it. He described contractual commitments to resist unfounded requests and noted the situation had not arisen, but under the CLOUD Act a US-headquartered company can be compelled regardless of where the data physically sits. Separately released UK government material records Microsoft advising Scottish police authorities that it could not guarantee data sovereignty for M365.

Note carefully what that establishes. It does not show that transfers happen routinely. It shows that the guarantee cannot be given, which means the protection rests on a company's continued willingness to resist rather than on anything structural. A control that depends on someone else's restraint is not a control.

The alternative is a specific set of decisions rather than a slogan: open-source stacks that can be audited instead of trusted, providers incorporated and operated inside EU jurisdiction, and key custody arranged so that compelled disclosure produces ciphertext. Sovereignty is a question of where keys live, who can be served an order, and what an adversary actually obtains when they win.

VII. Silicon is the second axis, and it is not closed

Jurisdictional sovereignty does not settle the trust question underneath it. Every x86 server runs on silicon designed by two American companies, with an out-of-band management layer, Intel ME or AMD PSP, that is closed, always powered, and unauditable by the operator running workloads on top of it. A flawless legal posture can still sit on hardware nobody in Europe can inspect.

Germany's flagship sovereign AI build illustrates the tension rather than resolving it. Deutsche Telekom and NVIDIA's Industrial AI Cloud in Munich represents more than €1 billion of private investment, went live in the first quarter of 2026, and raises German AI compute by roughly 50 percent using up to 10,000 NVIDIA Blackwell GPUs. Telekom's commitment that data stays in Germany and that only certified German and European personnel have access is a real jurisdictional guarantee. It is also, at the silicon layer, an American accelerator monoculture. Both things are true at once, and pretending otherwise is how sovereignty rhetoric loses its credibility.

RISC-V is the right direction on this second axis. The instruction set is open and owned by no single company or state, and Europe is funding it seriously. The DARE project, Digital Autonomy with RISC-V in Europe, launched in March 2025 under the EuroHPC JU, led by the Barcelona Supercomputing Center with 38 partners under a framework agreement running to 2030. Its first specific grant agreement carries roughly €240 million, combining up to €120 million of EU funding with matched contributions from participating states, covering a general-purpose HPC processor, a vector accelerator, and an AI processing unit built as European-designed chiplets.

What RISC-V does not yet deliver is a complete production trust chain. No shipping RVA23-class part currently pairs full vector and crypto extensions with a discrete TPM or hardware security module. No cloud provider rents that class of compute to an individual for validation work. The European Processor Initiative's Rhea prototype and EPAC accelerator reached validation boards rather than volume production, which is why European supercomputing still runs on off-the-shelf NVIDIA and AMD parts for large-scale training.

Fabrication remains foreign regardless. The EU Chips Act set a target of 20 percent of global manufacturing capacity by 2030; the Commission's own scorecard put the bloc at 9 percent in June 2026, and the Chips Act 2.0 draft of 3 June 2026 abandons the 20 percent goal in favour of design capacity, specialty materials, equipment, and advanced packaging. Intel's €30 billion Magdeburg fab was cancelled, leaving TSMC's ESMC joint venture in Dresden as the sole advanced-node anchor, itself 70 percent Taiwanese-owned and backed by €5 billion in German state aid. Advanced packaging is the sharper dependency: CoWoS, the standard for HPC and AI packaging, is executed almost exclusively by TSMC in Taiwan, so a wafer fabricated in Dresden still does not make a finished European part.

An open instruction set architecture is not fabrication sovereignty, and conflating the two is exactly the overclaiming this argument cannot afford. The defensible position is to build the measured-boot and remote-attestation architecture now, on hardware that exists today, so it is ready when RISC-V silicon closes the trust-chain gap. That is also, as section IV argued, work that displaced German embedded systems engineers are unusually well placed to do.

VIII. Prove it small before arguing it big

No government commits to an industrial-to-infrastructure transition on the strength of an argument. It needs a working reference implementation, small enough for one architect to build and complete enough that the objections have already been answered somewhere concrete.

That is what InitCompany is for. It is a full-stack, EU-sovereign, entirely open-source infrastructure platform, specified across a documented architecture corpus and built to run a real organisation end to end: identity, secrets and key custody, networking, compute, CI/CD, monitoring, and forensics, on sovereign cloud, with no dependency on a jurisdiction that can be compelled against its will. The SME scope of up to a hundred users is deliberate. Every governance and technical decision a national build would face, including key custody and quorum, attestation, classification propagation, licence provenance, and auditable rather than trusted controls, gets made, justified, and written down at a scale where one person can defend every line of it.

The claim is not that the vision is compelling. It is that the reference architecture exists, the decisions are documented with their reasoning, and the arguments a scale-up would need to make have already been made once against real constraints.

IX. The mandate

Germany is unlikely to get another convergence like this one. The €500 billion will be spent regardless, and the only live question is on what. Between October 2026 and the end of 2029, tens of thousands of engineers, technicians, and specialists will leave German industry on a published schedule, and the only live question is whether anything is built to receive them. The industrial sites and their grid connections will be repurposed or demolished, and the only live question is by whom, under whose jurisdiction, and on whose silicon.

Take the interconnection. Redirect the electricians to critical facilities and the safety engineers to assurance. Point the robotics industry at the customer that is arriving rather than the one that is leaving. Hold both sovereignty axes, and stay honest about which one is not yet closed.

Sources

Facts, figures, and testimony come from the following. Interpretation and argument are mine.

Industrial contraction: aggregate figures

Company-specific restructuring programmes

Public finance and national strategy

Grid constraints and brownfield conversion

Labour demand and skills

Robotics and automation

Jurisdiction and the CLOUD Act

Note on dating: reporting gives the Senate hearing date variously as 10 June, 18 June, and 10 July 2025. This document says "mid-2025" rather than pick one. The substance of the testimony is consistent across all accounts.

Silicon, RISC-V, and fabrication

Sovereign AI infrastructure in Germany

Revision of 11 September 2026: section II's Volkswagen paragraph and its source added from Volkswagen Group Media Information No. 74/2026 (3 September 2026); the text of 24 August 2026 is kept unchanged as Vanguard Manifesto.md. Everything else as written on 24 August.