The AI Power Bottleneck: Siemens Energy, Schneider Electric and ABB
For most of the artificial intelligence boom, the scarce thing was silicon. Investors learned to track chip allocations, foundry capacity and packaging constraints the way an earlier generation tracked oil rigs. Every conversation about who would win the AI build-out began, and usually ended, with the question of who could get their hands on accelerators.
That conversation has quietly changed. Ask anyone actually building a large computing campus today what keeps them awake, and the answer is rarely chips. It is power. Not power in the abstract, philosophical sense of energy transition slogans, but power in the brutally literal sense of copper, steel, switchgear, turbines, transformers and a grid connection date that a utility is willing to put in writing.
This is the AI power bottleneck. It is the point at which a story about software collides with the physical world and discovers that the physical world moves on its own schedule. And it is arguably the most investable consequence of the AI era for European equity investors, because the companies that sit at that collision point are not in California. They are in Munich, in Rueil-Malmaison and in Zurich.
This article looks at three of them: Siemens Energy, Schneider Electric and ABB. Each one touches a different part of the electron’s journey from the generator to the processor. Each one has re-rated hard as the market has caught on. And each one carries a distinct set of risks that a thoughtful investor should understand before deciding that the bottleneck automatically translates into shareholder returns.
The bottleneck is not one bottleneck
The phrase “power constraint” gets used as though it describes a single problem. It does not. It describes a chain of separate constraints, each with its own suppliers, its own lead times and its own economics. Understanding where each company sits on that chain is the whole game, because a company exposed to the tightest link captures the most pricing power, while a company exposed to a link that loosens first gets squeezed the earliest.
The first constraint is generation
Somebody has to make the electricity. Large computing campuses do not want intermittent supply and they do not want to wait a decade for a nuclear plant. In practice, that has pushed an enormous amount of near-term demand toward gas turbines, with nuclear and renewables filling in over longer horizons. The market for large gas turbines had been written off as structurally declining for years. Manufacturers had shrunk capacity accordingly. When demand returned, it returned into a supply base that had been deliberately made smaller.
Turbine slots are now booked years ahead. Buyers negotiate on delivery date rather than efficiency and pay premiums to move up the queue. The industry has flipped from a buyer’s market to a seller’s market, and manufacturers are careful not to over-expand into what may prove a spike rather than a plateau.
The second constraint is transmission
Electricity that exists somewhere is useless if it cannot reach the place where the racks are. This is where the situation becomes genuinely awkward. Grids across Europe and North America were designed for a world where demand grew slowly and predictably. They were not designed for a single customer to arrive and request the load of a mid-sized city, on a site chosen for fibre and land prices rather than for proximity to existing substations.
The hardware needed to fix this is unglamorous and desperately scarce. Large power transformers now carry lead times measured in years rather than months. High-voltage direct current links, which are the only practical way to move very large amounts of power over long distances, are effectively sold out, with utilities signing reservation agreements simply to hold a place in a factory queue that does not yet have a product design attached to it.
Underneath that sits an even narrower chokepoint: grain-oriented electrical steel, the specialised material at the heart of a transformer core. There are very few producers globally. Expanding that capacity is capital intensive, slow and unattractive if you suspect the demand surge is temporary. So the shortage persists even as everyone agrees it should not.
The third constraint is the building’s own power train
Once electricity arrives at the fence line, it has to be stepped down, distributed, protected, backed up and delivered to racks that draw more power in a single cabinet than an entire room used to draw. This is the world of medium-voltage switchgear, busway, uninterruptible power supplies, power distribution units and the increasingly exotic direct-current architectures that high-density AI racks are pushing toward.
The fourth constraint is heat
Every watt that goes into a processor comes out as heat, and modern accelerators produce heat at densities that air simply cannot carry away. Liquid cooling has moved from exotic to mandatory. That has created a new equipment category almost overnight, with cooling distribution units, cold plates, manifolds and closed-loop systems becoming as strategically important as the electrical gear they sit beside.
It has also created a political constraint that is only beginning to bite. Cooling consumes water, and communities that were relaxed about a data centre arriving are considerably less relaxed once they understand its water draw. Regulators in several jurisdictions have begun writing rules around it, which pushes the industry toward closed-loop designs that use less water but more electricity, tightening the first constraint further.
The single most useful mental model for this theme: AI capital expenditure is not a technology budget any more. It is an infrastructure budget wearing a technology badge. And infrastructure budgets are spent with industrial companies.
Why this is a European story
There is an understandable instinct to treat AI as an American phenomenon with a handful of Asian manufacturing dependencies. For the compute layer, that is broadly right. For the power layer, it is not.
The heavy electrical industry consolidated over decades into a small group of firms with genuinely global scale, deep installed bases and the engineering depth to certify equipment for utility-grade reliability. A disproportionate number of them are European. This is partly historical accident and partly the legacy of Europe’s own electrification and grid-interconnection programmes, which forced these companies to build capabilities in high-voltage transmission and grid stability long before anyone was thinking about training clusters.
The consequence is that European investors have something rare: direct, high-quality exposure to the AI build-out that does not require owning an expensive American semiconductor stock. It is exposure with different risk characteristics, different cyclicality and, historically, different valuation. That last point has been eroding fast, which is precisely why the analysis matters more now than it did when the theme was young.
Siemens Energy: the company that sells the electrons and the wires
Background
Siemens Energy exists because Siemens decided it did not want to own the energy business any more. The spin-off arrived at what looked like a spectacularly bad moment: a portfolio anchored in gas turbines, in a world convinced that gas turbines were a stranded asset, bolted to a wind turbine business that was about to discover serious quality problems in its onshore fleet and serious cost problems in its offshore ramp-up.
The years that followed were genuinely difficult. Losses in the wind unit were large enough to threaten the group. The company ended up seeking state-backed guarantees to keep issuing performance bonds to customers, which is about as close to an emergency as a large industrial company gets without an actual restructuring. The share price behaved accordingly.
What happened next is one of the more remarkable reversals in recent European industrial history, and it happened for reasons that had almost nothing to do with anything the company chose. The world simply decided it needed enormous quantities of firm generation and grid equipment, immediately, and Siemens Energy happened to own two of the best franchises for supplying both.
Business model
The group is best understood as four quite different businesses sharing a balance sheet.
- Gas Services builds and, more importantly, services large gas turbines. The service annuity here is the part investors often underrate: turbines run for decades and require parts, upgrades and overhauls throughout, generating recurring, high-margin, low-cyclicality revenue that dwarfs the original equipment margin.
- Grid Technologies makes the transmission equipment that everyone now wants and nobody can get quickly: high-voltage switchgear, transformers, direct-current converter stations and the associated engineering. This has become the growth engine.
- Transformation of Industry covers compressors, electrolysers and industrial decarbonisation equipment. It is smaller, more heterogeneous, and its fortunes depend heavily on the pace of hydrogen and industrial capital spending, which has been slower than the enthusiasts hoped.
- Siemens Gamesa is the wind business, still majority-owned, historically the source of the group’s pain and now the subject of a long argument among investors about whether it should be fixed, shrunk or separated.
Market position
In large gas turbines, the competitive set is small enough to name in a breath: essentially two American players and one Japanese group alongside Siemens Energy. That is an oligopoly with high barriers, decades-long product qualification cycles and installed bases that lock customers in through the service relationship. When demand exceeds supply in a structure like that, pricing does not improve gradually. It improves sharply.
In transmission, the picture is similar. Building a high-voltage transformer factory is not a matter of renting a shed. It requires specialised presses, testing facilities, skilled winders and a certification history that customers trust. New entrants do not appear overnight. The incumbents therefore capture most of the surge, and because customers now care more about delivery certainty than about squeezing the last percentage point out of a price negotiation, the commercial terms have shifted meaningfully in the supplier’s favour.
Why investors care right now
Siemens Energy is the purest listed expression of the idea that AI demand shows up first as an electricity problem. Its backlog stretches far enough into the future to give unusual revenue visibility for an industrial company. Management has been explicit that the market is supply-constrained rather than demand-constrained, and that hyperscaler customers are willing to pay for speed.
Simultaneously, the historic drag is easing. The wind business has been working through its problem fleet and improving its commercial discipline, which removes the tail risk that dominated the equity story for years. A company that goes from “may need rescuing” to “cannot make product fast enough” in a relatively short span will re-rate violently, and it did.
The bull case
The optimistic view rests on duration rather than intensity. Grid investment is not a one-year event. Ageing infrastructure in developed markets needs replacing regardless of AI. Electrification of heating, transport and industry adds load independently of data centres. Renewable connections require the same equipment. Layer AI demand on top of all of that, into a supply base that spent a decade shrinking, and you have a genuinely multi-year up-cycle in which the constrained supplier sets the terms.
The service annuity strengthens the argument. Every turbine and every substation installed today creates decades of aftermarket revenue tomorrow. The equipment cycle may be cyclical; the installed base it leaves behind is not. Investors who focus only on order intake miss the compounding quality of what those orders leave in their wake.
The bear case
The first risk is the oldest one in capital goods: execution. Large energy projects are fixed-price, multi-year, technically complex and full of opportunities to lose money. This company has demonstrated, painfully and recently, that it can misprice risk at scale. A booming order book taken at aggressive terms is a wonderful thing until it is delivered badly.
The second is the wind business. It has improved, but improvement from a very poor starting point is not the same as structural health. Offshore wind economics remain fragile across the industry, project cancellations have been common, and the unit still consumes management attention that could be spent on the parts of the group that are actually winning.
The third is the shape of the demand curve. Gas turbine demand collapsed once before, precisely because the industry over-extrapolated a boom. Management appears to have learned that lesson and is expanding capacity cautiously, which is prudent but also caps how much of the surge can be converted into revenue. If capacity does eventually catch up with demand, the pricing premium that is currently flattering margins disappears.
The fourth is valuation. The stock is no longer the misunderstood recovery story it was. Expectations embedded in the price now assume the up-cycle persists. That is a very different risk profile from the one early buyers took on.
Schneider Electric: the company that owns the last few metres
Background
Schneider Electric began life in heavy industry, in steel and armaments, and spent the modern era systematically reinventing itself into something closer to an energy management and automation platform. The transformation was executed through a long series of acquisitions that most investors have forgotten individually but that collectively assembled an unusually complete position in electrical distribution and, crucially, in the data centre.
Business model
Schneider organises itself around two broad activities. Energy Management is the larger by far and covers everything from medium-voltage switchgear down through building distribution to the rack-level power equipment inside a data hall, plus the software that monitors and optimises all of it. Industrial Automation covers control systems, drives and industrial software.
What makes the model distinctive is the layering. Schneider sells hardware, but it also sells the design software engineers use to specify electrical systems, the simulation tools used to model them, the monitoring platforms used to run them and the services used to maintain them. When a customer designs a facility in your software, specifies your equipment inside it and then manages that equipment with your monitoring layer, you are not really competing on the price of a circuit breaker.
The company has also pushed hard into prefabrication. Rather than building electrical rooms on site, it delivers factory-assembled modules and complete power pods that arrive ready to connect. For customers whose binding constraint is time and skilled-labour availability rather than capital, this is enormously valuable, and it converts a commoditised installation business into an engineered product business.
Market position
In data centre power, Schneider is the closest thing the industry has to a default choice. Its position in uninterruptible power supplies, rack distribution and reference architectures gives it a presence in a very large share of facilities globally, and its relationships extend from the hyperscalers down through colocation operators to the enterprise rooms that still exist in the background.
It has extended that position into cooling, which was the obvious defensive and offensive move once liquid became mandatory. Owning both the power train and the heat rejection system lets it sell an integrated design rather than a component, and integrated designs are much harder for a competitor to unpick.
The strategic prize it is chasing is to be specified before the tender exists. Working directly with chip designers on reference architectures for high-density racks means that by the time a customer is choosing equipment, the reference design already assumes Schneider’s approach. That is the most valuable position in any industrial market: not winning the bid, but writing the specification the bid is based on.
Why investors care right now
Schneider offers the cleanest read-through to data centre capital expenditure of the three. Its data centre and networks end market has been growing at rates that would embarrass most software companies, and unlike a turbine order, this revenue converts relatively quickly. Investors get a faster feedback loop between the announcement of a hyperscaler capital plan and the appearance of revenue in the accounts.
The business also carries an attractive financial profile: high returns on capital, a large services component, structural exposure to electrification well beyond data centres, and a software layer that supports the multiple. It is, in short, the quality name in the group, and it is priced as such.
The bull case
The optimist argues that Schneider has been systematically converting a hardware business into a platform business, and that the market still values it partly as the former. Every additional layer of software and services makes the customer relationship stickier and the earnings stream more resilient. The AI build-out accelerates that process by pushing customers toward integrated, prefabricated, designed-in solutions rather than component purchasing.
There is also a breadth argument. Data centres are the exciting part, but the same electrical content is being pushed into buildings, industry, transport and grids by electrification generally. If AI capital spending pauses, the underlying electrification demand does not vanish. That gives the business a second engine that a pure data centre supplier lacks.
Finally, the density trend works in Schneider’s favour. As racks draw more power, the electrical and cooling content per unit of compute rises. Even a flat number of facilities would imply growing content per facility. That is the quiet compounding underneath the headline growth.
The bear case
The valuation is the loudest concern. Schneider trades at a premium that reflects quality, growth and thematic appeal simultaneously. Premiums of that kind are unforgiving. A single disappointing quarter in the data centre end market, or any evidence that hyperscaler spending is being digested rather than accelerated, has an outsized effect on a stock priced for continuation.
Customer concentration is a subtler risk. A small number of enormously powerful buyers now account for a growing share of the industry’s demand. Those buyers have engineering departments, procurement sophistication and every incentive to standardise, dual-source and design out cost. The relationship is warm today because supply is tight. It will be tested when it is not.
There is also the question of whether the hyperscalers eventually internalise more of the electrical design. Several already design their own racks, their own power shelves and increasingly their own facility architectures. Every step in that direction moves value away from the equipment vendor and toward the buyer.
Industrial Automation, meanwhile, has been the weaker half of the business, exposed to a manufacturing capital spending cycle that has been soft in Europe and inconsistent elsewhere. It dilutes the growth story and occasionally supplies the negative surprise.
ABB: the focused electrifier
Background
ABB is the product of a merger between a Swedish and a Swiss engineering group, and for much of its recent history it suffered from the classic conglomerate problem: excellent individual businesses, obscured by a structure that made them hard to value and harder to manage. Successive management teams attempted to fix this. The current one has been considerably more decisive about it.
The defining recent decision was to part with the robotics division. Robotics was the most visible ABB business to the outside world and the one that generated the most magazine covers, but it was also cyclical, exposed to automotive capital spending and only loosely connected to the rest of the group. Selling it to a buyer with a strategic interest in intelligent machines simplified ABB into what it had arguably always been at its core: an electrification and automation company.
Business model
Post-simplification, ABB is organised around electrification, motion and process automation. Electrification is the crown jewel and the piece that matters most for this theme. It spans low-voltage products, medium-voltage switchgear, protection equipment, uninterruptible power systems and the emerging direct-current architectures that very high-density racks are pulling the industry toward.
Motion covers motors and drives, which is a quietly excellent business with strong efficiency-driven demand and a large service tail. Process Automation serves heavy industries and has now absorbed the machine automation activities, tightening the structure further.
Market position
ABB holds strong positions across the electrical value chain, particularly in medium-voltage and in the protection and distribution equipment that sits between the grid connection and the white space. In data centres it competes directly with Schneider and with a set of American specialists, and it has been aggressive about repositioning itself as an AI infrastructure supplier rather than a generalist.
The most interesting strategic move has been embedding its equipment models into the digital design environments used to plan data centre campuses. If an engineer designing a facility in a simulation environment can drop in a validated digital twin of ABB switchgear and see it work, the specification battle is substantially won before procurement begins. It is the same logic Schneider applies through its own software, approached from a different angle.
Why investors care right now
ABB offers something the other two do not: a portfolio reset happening simultaneously with a demand boom. Simplification stories and cyclical upswings rarely coincide. When they do, the combination of margin improvement from focus and revenue growth from the cycle can produce earnings momentum that surprises even attentive investors.
The bull case
The core argument is quality plus focus. ABB now has fewer businesses to explain, all of them oriented toward electrification and automation, all of them exposed to secular demand drivers that extend well beyond data centres. Margins have improved substantially over recent years through operational discipline rather than cyclical luck, which suggests the improvement is durable.
Electrification specifically benefits from a broad set of drivers at once: data centres, grid upgrades, building electrification, industrial efficiency and the electrification of transport. That breadth means the business does not depend on any single theme staying hot.
There is also a re-rating argument. Conglomerate structures attract discounts. Focused businesses attract premiums. If the market comes to see ABB as a pure electrification play rather than a diversified engineering group, the multiple can expand independently of earnings.
The bear case
Short-cycle exposure is the principal vulnerability. Much of the electrification business flows through distributors, which means it moves with construction and industrial activity and can turn quickly. Distributor inventory cycles amplify this: channels build stock when lead times stretch and then stop ordering entirely when they normalise, producing an air pocket in reported orders that has nothing to do with end demand.
Competition in low-voltage products is genuinely intense, and Chinese manufacturers have been improving quality while competing hard on price. In the parts of the portfolio closest to commodity electrical products, pricing power is far weaker than the transmission-equipment narrative suggests.
Simplification also cuts both ways. Divesting robotics removed a cyclical business, but it also removed diversification and a genuine technology asset at a moment when the boundary between physical automation and artificial intelligence is becoming interesting. Reasonable people can disagree about whether that was a good trade.
And ABB, like the others, now trades at a level that assumes the good times continue. The margin of safety that existed when the market treated it as a sleepy Swiss engineer has largely gone.
Three companies, one theme, three different bets
It is tempting to treat these as interchangeable ways to buy the same idea. They are not. They sit at different points on the chain, with different cycle lengths, different customer sets and different failure modes.
Siemens Energy is the
upstream bet. Its revenue arrives through very large, very long projects with utilities and independent power producers. Visibility is exceptional because the backlog is enormous, but conversion is slow and execution risk is high. It is the most leveraged to the idea that the constraint is generation and transmission rather than anything inside the building. It is also the most operationally complex.
Schneider Electric is the
downstream bet. Its revenue arrives faster, from a mix of project and product business, with the data hall itself as the centre of gravity. It has the highest quality financial profile and the most software content, and it is the most directly exposed to hyperscaler capital expenditure decisions, for better and for worse.
ABB is the
midstream bet with the widest industrial base. It touches the data centre meaningfully but is not defined by it, and its fortunes are more tied to general electrification and industrial activity. It offers the self-help story that the other two largely lack.
A portfolio holding all three is not diversified across the theme so much as concentrated in it three times. If AI capital spending disappoints, all three will fall together, whatever their individual merits. Understanding that correlation is more important than fine-tuning the relative weightings.
What the market has already priced in
This is where discipline matters. Each of these companies has enjoyed a substantial re-rating. The narrative has moved from “cheap industrial with a hidden asset” to “essential infrastructure supplier to the most important technology build-out in a generation.” That is a far more crowded and far more expensive place to be standing.
The market has already priced in continued growth in hyperscaler capital expenditure, the persistence of supply tightness and therefore of unusual pricing power, successful execution of enormous backlogs, and — in Siemens Energy’s case — the wind business ceasing to be a problem. None of these assumptions is unreasonable. All of them are assumptions. The useful question is not “is this a good business?” — all three plainly are — but “what has to go right for the current price to make sense, and how confident am I in each of those things?”
What could break the thesis
Digestion, not collapse
The most likely negative scenario is not a crash in AI investment but a pause. Capital spending programmes of this scale are not smooth. Buyers commit, build, absorb and then reassess. A period in which hyperscalers slow new commitments while integrating what they have already ordered would be entirely normal industrial behaviour and would still be devastating for stocks priced for uninterrupted acceleration.
Supply catching up
Every participant in the electrical supply chain is currently expanding capacity. So are their competitors. So are new entrants tempted by the economics. Capacity added into a shortage arrives, as a rule, slightly after the shortage ends. When it does, the premium pricing that has been flattering margins evaporates quickly, and the industry rediscovers that heavy electrical equipment is a competitive business.
This is the single most under-appreciated risk in the theme. The current margin structure is not a permanent feature of these businesses. It is what happens when demand temporarily exceeds supply in an oligopoly.
Efficiency improving faster than expected
The entire theme rests on the assumption that compute demand grows faster than compute efficiency improves. History suggests caution here. Every previous computing era saw efficiency gains that repeatedly confounded power forecasts. If model architectures, chip designs or software approaches deliver a step change in energy per unit of useful work, the load growth curves that underpin current valuations get revised downward.
Political and social resistance
Large computing facilities consume power and water in communities that receive relatively few jobs in return. Electricity prices for households have become a live political issue in several markets, and data centres are an easy target. Moratoria, tariff structures that load costs onto large consumers, and planning refusals are all plausible. This does not stop the build-out, but it can slow it, relocate it and make it more expensive.
Customer power
The buyers here are among the most sophisticated procurement organisations in the world, with balance sheets far larger than their suppliers’. They are currently accepting supplier-friendly terms because they have no choice. They are simultaneously investing in alternatives: designing their own power architectures, signing direct generation agreements, exploring on-site generation and qualifying second and third sources. Every one of those efforts is aimed at reducing the leverage their suppliers currently enjoy.