Every conversation about artificial intelligence eventually collides with a physical constraint: electricity. Models are trained in buildings, cooled by water and glycol, and fed by transformers, switchgear and uninterruptible power supplies. Somewhere in that chain — in roughly four out of every ten data centres on earth — sits equipment designed by a French company that most candidates never think of as an AI company at all. Schneider Electric closed 2024 with revenue of €38.2 billion, up 8.4% organically, and its Energy Management business, which sells the electrical backbone of data centres, grew faster than the group. The company's own investor communication now describes data centres and networks as its single most dynamic end market.
This is the story junior candidates in France and Europe consistently misread. They apply to the model labs — the twenty-person research teams — and ignore the industrial layer that determines whether those models can run at all. The hiring volume, the salary premiums and the durable career paths are disproportionately in the second group.
The demand shock: from 460 TWh to something no grid planned for
The International Energy Agency put global data centre electricity consumption at approximately 460 TWh in 2022 — around 2% of global demand — and projected a range of 620 to 1,050 TWh by 2026, with the central case landing above 800 TWh. To put that in European terms, the upper bound is roughly the annual electricity consumption of Germany and France combined. Nothing in the grid planning cycle of 2015 anticipated a load of that shape: dense, continuous, and geographically clustered around fibre routes and cheap power rather than around population.
The consequence is that AI capacity has stopped being a software procurement problem and become an electrical engineering problem. A traditional enterprise server rack drew 5 to 10 kW. A rack of accelerators for training workloads draws 40 to 130 kW, and the reference architectures now being deployed assume liquid cooling because air physically cannot remove that much heat from that volume. Every one of those numbers is an engineering brief: higher-density busways, medium-voltage distribution pushed closer to the rack, coolant distribution units, and power monitoring at a granularity that did not exist five years ago.
Why this lands on Schneider specifically
Schneider Electric sits at an unusual intersection. It sells the low- and medium-voltage electrical distribution equipment (Schneider Electric brand, plus Square D in North America), the uninterruptible power supplies and rack infrastructure inherited from its 2007 acquisition of American Power Conversion, and the software layer — EcoStruxure — that monitors and optimises the whole assembly. In October 2024 it added the missing piece by agreeing to acquire a 75% controlling stake in Motivair, a US specialist in liquid cooling and coolant distribution units, for approximately $850 million, with the remaining 25% scheduled for 2028. Read as a strategy document, that transaction says the company expects thermal management to be as commercially important as power distribution.
In parallel, Schneider has been publishing joint reference designs with NVIDIA for high-density AI clusters — pre-engineered blueprints covering power, cooling and controls for racks in the 100 kW class. Reference designs matter more than press releases: they are how an equipment vendor becomes the default assumption inside a hyperscaler's build standard, and they are also the artefact junior engineers are most likely to work inside during their first two years.
The European policy layer nobody should ignore
The demand is not only commercial. In February 2025 the European Commission announced InvestAI, a mobilisation package of up to €200 billion for AI in Europe, including €20 billion earmarked for a network of AI gigafactories — facilities built around very large accelerator clusters. In the same month, at the Paris AI Action Summit, the French government announced roughly €109 billion of private investment commitments in French AI infrastructure, with data centre construction as the dominant line item.
For a candidate, those two announcements should be read as a hiring forecast rather than as politics. Gigafactory-scale compute in Europe implies grid connections, substation work, permitting, cooling water strategy, heat reuse into district networks, and a multi-year maintenance workforce. That work cannot be offshored: it is physically located in Mulhouse, Marseille, Dunkirk, Grenoble, and the industrial belts around Paris and Lyon. It also cannot be automated away in the 2020s, because the constraint is skilled hands and certified engineers, not code.
Where the jobs actually are: five families, ranked by durability
1. Electrical distribution and power systems engineering
The core discipline. Sizing medium-voltage transformers, designing selectivity and protection schemes, modelling short-circuit behaviour, coordinating with the grid operator on connection studies. Entry routes come through French engineering schools with electrical specialisations (Supélec/CentraleSupélec, INSA, Arts et Métiers, Polytech network) and increasingly through BUT Génie Électrique graduates who move into design offices. This family is the least exposed to automation and the most exposed to demand: a grid connection study cannot be produced by a language model, and the volume of studies is rising with every announced facility.
2. Thermal and liquid cooling engineering
The fastest-growing and least crowded family. Direct-to-chip cooling, rear-door heat exchangers, coolant chemistry, leak detection, and the hydraulic design of coolant distribution loops. Mechanical engineers who can speak fluently about both fluid dynamics and rack-level electronics are, at present, scarce relative to demand across Europe. Candidates from mechanical or process engineering backgrounds routinely underestimate how transferable their training is to this market.
3. Controls, monitoring and industrial software
The EcoStruxure layer: building management systems, power monitoring, digital twins of the electrical topology, and increasingly predictive maintenance models trained on telemetry from installed equipment. This is where software candidates belong if they want AI-adjacent work with industrial durability. The stack is unglamorous — Modbus, BACnet, OPC UA, time-series databases, dashboards that operators actually trust — and precisely because it is unglamorous, competition per opening is a fraction of what it is for a machine-learning role.
4. Commissioning, field service and operations
Every installed system must be tested, energised, and kept alive for fifteen years. Commissioning engineers and field service technicians are the roles hyperscalers and colocation operators struggle hardest to fill, and they are the roles with the clearest apprenticeship path for candidates without a five-year engineering degree. They also travel: a commissioning engineer who has energised facilities in three countries becomes structurally difficult to replace.
5. Supply chain, quality and industrial project management
Transformer and switchgear lead times became a global bottleneck after 2022. That turned procurement, supplier qualification and project scheduling from back-office functions into strategic ones. Candidates with industrial engineering or supply chain training who can read a technical specification — not merely a purchase order — are being promoted unusually fast.
What the market pays, and why the premium is structural
Compensation in this segment behaves differently from generalist engineering. The premium is not paid for the diploma; it is paid for three specific, verifiable things: certification (electrical safety and equipment-specific qualifications), demonstrated experience on high-density or mission-critical sites, and bilingual capability, because European projects are executed in English while French regulatory and site work happens in French. A graduate who arrives with two of the three negotiates from a materially different position than one who arrives with a strong average and no site exposure.
The durability argument matters more than the starting number. Data centre electrical infrastructure carries a service life measured in decades, and the maintenance obligation is contractual. That produces a demand curve that does not collapse when a funding cycle turns — the opposite of the pattern candidates experienced in consumer software hiring between 2022 and 2024.
How to become the obvious candidate in six months
The single most common failure in applications to this sector is a CV that describes coursework instead of evidence. Recruiters in industrial infrastructure are reading for proof that the candidate has touched a real system. Four concrete moves close that gap:
- Build one artefact, not five projects. A single-line diagram for a 2 MW hall with protection coordination, or a thermal model of a 60 kW rack with a defended cooling choice, communicates more than any list of technologies.
- Get the safety and equipment certifications early. French electrical clearance levels (habilitation électrique) and vendor-specific equipment training are cheap, fast, and function as a hard filter in screening.
- Learn the vocabulary of the buyer. PUE, WUE, N+1 and 2N redundancy, Tier classification, commissioning levels L1 to L5. Using these terms correctly signals a candidate who has read the industry rather than the recruitment page.
- Target the layer below the headline. Applications concentrate on the operators whose names appear in the press. The engineering and integration firms, the panel builders, and the service partners hire more juniors, interview faster, and give broader exposure in the first two years.
The honest risks
Three caveats belong in any serious brief. First, projections are projections: the IEA range for 2026 spans 620 to 1,050 TWh precisely because efficiency gains and model architecture changes could bend the curve. Second, local resistance to data centre construction — over land, water and grid capacity — is rising in several European regions and can delay projects by years. Third, the concentration of demand in a handful of hyperscale customers means suppliers carry customer-concentration risk, which propagates into hiring plans.
None of these invalidate the thesis. They argue for building transferable capability — power systems, thermal engineering, industrial controls — rather than a career bet on one facility or one customer. Those skills are equally in demand in grid and energy transition work, in semiconductor fabs, and in electrified manufacturing generally.
The value chain, decoded: who buys, who builds, who operates
Candidates lose months by applying into the wrong layer of a four-layer market. Understanding the chain is the single highest-return hour of research in this sector.
At the top sit the demand owners — hyperscalers and AI labs that commission capacity. They hire few juniors and almost never for design work. Below them are the developers and colocation operators, who own the buildings and the long-term service obligation; they hire operations, commissioning and energy managers. Below those are the engineering, procurement and construction firms and design offices that actually specify the electrical and mechanical systems — in France this includes the large infrastructure contractors and independent design offices, and this is where the majority of graduate design roles live. Finally there are the equipment manufacturers and their integration partners: Schneider Electric and its competitors, the panel builders who assemble switchgear to specification, and the service networks that maintain it.
The practical implication is precise: a candidate who wants to design should target layer three and layer four, not layer one. A candidate who wants operational responsibility fastest should target layer two, where a 25-year-old can hold accountability for a live facility within three years. The compensation gap between these layers is narrower than the visibility gap, which is exactly why the less visible layers are easier to enter and often better training grounds.
The grid queue: the constraint that will define the next five years
The binding limit on European AI capacity is increasingly not chip supply but grid connection. Across several European markets, requests for large industrial connections now face multi-year queues, because a new load of tens or hundreds of megawatts may require reinforcement of the transmission network, not merely a new substation. In the Netherlands and parts of Ireland, connection constraints have already forced operators to pause or relocate projects; in France, the relative headroom created by the nuclear fleet has become an explicit commercial argument for locating compute domestically.
This creates two job families that barely existed as distinct careers a decade ago. The first is grid interconnection engineering: professionals who model load profiles, negotiate connection agreements, and design on-site generation or storage to smooth demand. The second is energy strategy and flexibility: structuring power purchase agreements, designing demand-response participation, and quantifying carbon intensity per unit of compute for corporate reporting. Both sit at the intersection of engineering and commercial negotiation, and both are unusually well paid relative to years of experience, because the population of people who can credibly do them is small.
For a graduate, the actionable read is that fluency in the grid side of the equation — connection processes, tariff structures, capacity mechanisms — is a differentiator that costs weeks to acquire and years for competitors to replicate. It is also portable across every electrified industry, which is the definition of a durable skill.
What to do this week
Choose one of the five families above and spend the week producing a single defensible artefact inside it. Then map three employers per family: the equipment vendor, the engineering firm that integrates it, and the operator that runs it. Apply to all three layers, not just the one with the recognisable logo. The AI build-out in Europe is, for the next decade, a hiring event in electrical and mechanical engineering wearing a software company's branding — and the candidates who understand that distinction are the ones who get the offers.