Who Is Building the Power for AI, and Which of Them Have a Moat?
A layer-by-layer map of the companies supplying the AI power buildout, what their order books really say, and how to tell a franchise from a cycle.
The first part of this series, the AI energy bottleneck, argued that the AI power story is not about a shortage of electricity. It is about a shortage of the equipment that converts and moves electricity, and of the people who install it.
This part answers the obvious follow-on question. If the scarce thing is turbines, transformers, switchgear and skilled crews, who sells them, what do their order books actually say, and which of these businesses are still good businesses if the AI capital expenditure cycle turns?
That last clause is the one that matters. Almost every company in this article is currently reporting record orders. That tells you the cycle is on. It tells you nothing about whether the company has a durable advantage. Separating those two things is the entire job.
What You'll Learn
- How to read a backlog properly, and the three questions that reveal whether it is worth anything
- The turbine oligopoly, with real numbers from GE Vernova and Siemens Energy filings
- Why the grid equipment layer may be the better business than the turbine layer
- The speed-to-power hierarchy, from 90-day fuel cells to 2029 turbine slots
- How long-term contracts quietly re-rated the merchant nuclear industry
- Why the installation contractors earn one third the margin of the manufacturers, and what that tells you
- Six ways the hyperscalers are routing around the bottleneck entirely
- The bear case, drawn from what happened to this exact industry between 2008 and 2015
The Seven Chokepoints, in 400 Words
If you have not read Part 1, here is what you need to carry into this one.
Electricity reaches a GPU through nine layers. Fuel is abundant. Rack-level power distribution is competitive and unconstrained. The chip itself is constrained, but by memory and advanced packaging rather than anything electrical, which is a separate story covered in the real AI chip shortage.
Everything in between is tight, and four layers are genuinely binding.
Generation equipment. Three companies build utility-scale gas turbines. GE Vernova disclosed 116 GW of gas equipment under contract at the end of its second quarter of 2026 against roughly 20 GW of annual production capacity.
Interconnection. Permission to connect a large load to the grid routinely takes four to seven years, because the grid operator must study the effect on everyone else's reliability. PJM's capacity auction for the 2026/2027 delivery year cleared at $329.17 per MW-day, exactly its regulatory price cap, which is the market's way of saying it would have paid more for capacity that does not exist.
Transmission. New high-voltage lines take five to ten years to permit and build, and are contested at every mile.
Transformers. Quoted lead times of roughly 128 weeks for power transformers and 144 weeks for generator step-up units, with prices up 77% and 45% respectively since 2019. The causes are specialised grain-oriented electrical steel, multi-year supplier qualification cycles, and a small ageing workforce of skilled winders.
Two further layers are tight without being binding: switchgear and UPS systems at around 44 weeks, and cooling, where the technology works but vendor qualification takes two to three years.
And one constraint responds to no amount of money at all: skilled labour. High-voltage electricians, commissioning engineers and substation technicians cannot be manufactured, imported or accelerated. Apprenticeships take years.
The geography of the bottleneck determines the geography of the profit. Because the constraints sit in the middle of the chain, the pricing power sits there too. Fuel producers compete on a commodity. Rack power distributors compete on price. The companies in between, selling custom electrical equipment with multi-year lead times to customers who cannot wait, are the ones currently able to raise prices and improve margins simultaneously. That is what the rest of this article maps.
Part 1: How to Read a Backlog Before You Read Any Company
Every company below will tell you its backlog is at a record. In a boom, that statement is nearly meaningless on its own. Three definitions and three questions make it useful.
The Three Terms
Orders are the value of new work won in a period. BacklogBacklogThe value of orders a company has won but not yet delivered. It offers visibility into future revenue, but not all backlog is equal: what matters is how firm the commitments are, whether the customer can cancel without penalty, and whether prices can be revised if the company's own costs rise.See all terms in the glossary is the cumulative value of work won but not yet delivered. Book-to-billBook-to-Bill RatioNew orders received in a period divided by revenue billed in the same period. Above 1.0 means the backlog is growing because the company is winning work faster than it can deliver it. Below 1.0 means the backlog is shrinking. It is one of the earliest signals that a capital goods cycle is turning.See all terms in the glossary is orders divided by revenue in the same period. Above 1.0, the backlog is growing. Below 1.0, it is being consumed faster than it is replaced.
Book-to-bill is the most useful of the three because it turns first. Revenue is the last thing to fall in a downturn, because it is delivering work sold years ago. Orders are the first. A company can post record revenue and record backlog in the same quarter that its book-to-bill falls through 1.0, and that quarter is the signal.
The Three Questions
Watch disclosure quality as a signal in its own right. Vertiv's second quarter 2026 earnings release, published on 29 July 2026, contained no backlog figure and no book-to-bill ratio, having previously been one of the more forthcoming disclosers in the sector. That may be entirely reasonable commercially. It also removes the single most useful forward indicator investors had for that business. When a company stops publishing a number during a boom, note it.
The guide on how to read annual reports covers where these disclosures live in a filing and how to compare them across periods.
Part 2: Turbines, the Tightest Oligopoly in Industrials
Three companies. No credible fourth. Order books stretching past the end of the decade.
GE Vernova
Spun out of General Electric in 2024, GE Vernova is the purest large-cap exposure to this theme, and its second quarter 2026 filing is worth reading directly.
The company raised full year 2026 guidance to revenue of $45.5 billion to $46.5 billion, and free cash flow of $11.5 billion to $12.5 billion, up sharply from a prior range of $6.5 billion to $7.5 billion. A free cash flow guidance raise of that magnitude mid-year is unusual and reflects large customer deposits arriving ahead of delivery.
Note the Wind line in that grid, because it is the honest counterweight. GE Vernova is not a pure AI power play. Its Wind segment posted orders down 40% organically and a segment EBITDA loss of $275 million in the quarter. The company contains a booming gas and grid business and a struggling wind business, and reported results blend them.
The most under-discussed number in GE Vernova's disclosure is the split between firm backlog and slot reservation agreements. At the end of Q2 2026, 53 GW was firm and 63 GW was reservations, and 10 GW converted from one to the other during the quarter. That conversion rate is the cleanest available real-time test of whether announced AI power demand is turning into committed spending. It deserves more attention than the headline 116 GW, which is routinely quoted as though it were all firm.
Siemens Energy
The European counterpart, and arguably the more balanced business. Its third quarter fiscal 2026 results reported group orders of €17.9 billion, a book-to-bill ratio of 1.57, and an all-time high order backlog of €162 billion. Gas Services took €10 billion of orders, up 62%, a record intake. Profit before special items reached €1,623 million against €497 million a year earlier.
Two details are worth pulling out. First, Grid Technologies raised its full year profit margin guidance to a range of 18% to 20%, which for a grid equipment business is a striking level of profitability. Second, Siemens Gamesa, the wind division that has been a persistent drag, reported its first positive quarterly result since 2022.
Mitsubishi Heavy Industries
The third supplier, and the least accessible to most investors given its Tokyo listing and highly diversified structure spanning aerospace, defence and shipbuilding alongside energy. Industry reporting indicates its turbine order book is sold out into 2028 and that it is working to roughly double manufacturing capacity, though these figures come from trade press rather than company disclosure and should be treated as directional.
The Investment Question
A three-firm oligopoly selling a product with a five-year wait to customers who cannot substitute is about as strong a near-term position as industrials offers. The question is what it looks like in 2032.
Gas turbines are, ultimately, capital equipment sold into a cyclical end market. The service revenue that follows an installed turbine for thirty years is the genuinely durable part of the business, and it is high margin and recurring. The equipment sales are the cyclical part. A useful way to think about these companies is that the current boom is building the installed base that generates the durable earnings a decade from now. That framing, distinguishing a temporary price advantage from a structural one, is the subject of understanding economic moats.
Part 3: The Grid Layer, Which May Be the Better Business
Turbines get the headlines. Grid equipment may be the more attractive economics, for three reasons: the demand is less dependent on AI specifically, the replacement cycle is independent and enormous, and the margins have expanded faster.
The Order Books
| Company | Disclosed grid or electrical metric | Period |
|---|---|---|
| Siemens Energy Grid Technologies | €51bn backlog, orders €5.4bn, up 28%, transformers the largest contributor | Q3 FY2026 |
| GE Vernova Electrification | $40.6bn equipment backlog, up 69% year on year, book-to-bill about 1.7 | Q2 2026 |
| Eaton Electrical Americas | Backlog up 44% year on year, rolling 12-month orders up 42% organically | Q1 2026 |
| Eaton Electrical Global | Backlog up 73% year on year | Q1 2026 |
| Hitachi Energy | Over $1bn of announced North American investment, South Boston plant expected around 2028 | Announced |
Eaton, and an Honest Complication
Eaton's first quarter 2026 results show Electrical Americas revenue of $3.6 billion, up 20%, with operating margins of 25.6%. Rolling twelve-month orders were up 42% organically and backlog was up 44%. The Electrical book-to-bill sat at 1.2.
Those are excellent numbers. And in the same quarter, total segment margins were 22.7%, down 120 basis points year on year.
That deserves an explanation rather than an omission. Eaton is spending heavily to expand capacity, and it has been buying aggressively: $9.55 billion for Boyd Thermal in March 2026, a thermal management business aimed squarely at data center cooling, plus $1.53 billion for Ultra PCS in aerospace, on top of $1.43 billion for Fibrebond in 2025. Acquisition and integration costs, and the mix effects of newly acquired businesses, compress reported margins in the short run.
Eaton's Boyd Thermal purchase is strategically revealing. Eaton was a power company. Cooling was Vertiv's territory. By spending $9.55 billion to buy into thermal management, Eaton is betting that the customer eventually wants a single vendor for the whole electrical and thermal package inside a data center, rather than best-of-breed components from specialists. If that bet is right, it compresses the addressable market for pure-play specialists. If it is wrong, Eaton has paid a great deal for a business outside its core competence at the top of a cycle. Watch which way that goes.
Hitachi Energy
The most direct pure exposure to the transformer shortage, and the hardest for most investors to own cleanly. Hitachi Energy is a subsidiary of Hitachi Ltd, listed in Tokyo, whose group results span IT services, rail and industrial systems. Buying the parent for the transformer business means buying a great deal else alongside it.
It has committed over $1 billion to North American capacity, including a South Boston, Virginia facility expected around 2028. As Part 1 noted, that timing helps 2028 rather than 2026.
The Structural Argument
Roughly 55% of American distribution transformers are more than 33 years old. That replacement wave exists whether or not another AI data center is ever built. It is the strongest argument that the grid equipment cycle outlasts the AI capital expenditure that triggered it.
Grid equipment demand has at least four independent drivers: AI data centers, general electrification including electric vehicles and heat pumps, renewable integration requiring new connection points, and pure age-related replacement. Turbine demand has fewer. A portfolio of demand drivers is worth more than a single one, even when the single one is growing faster today.
Part 4: Cable and Interconnect
A thinner section, because the disclosure here is less useful and the constraint is less acute.
Prysmian and Nexans dominate high-voltage cable, particularly the subsea and HVDC segments where technical barriers are highest. HVDC, which moves large quantities of power over long distances with lower losses than conventional alternating current, is structurally advantaged as grids attempt to connect distant generation to distant load.
Copper is an input cost and an occasional constraint, and it is covered properly in the copper and uranium trade, which examines the mining and refining side of that market including why miners and smelters capture very different economics.
The honest assessment: cable is tight, but not in the way transformers are tight. Manufacturing capacity is more expandable and the products are more standardised.
Part 5: On-Site Power, Where Speed Is the Product
This is the most interesting competitive dynamic in the whole story, because the customer is not buying electricity. The customer is buying time.
If a grid connection is four to seven years away and a competitor can energise a campus in eighteen months, the eighteen-month option is worth an enormous premium regardless of its cost per unit.
Approximate deployment times from commitment to running power. Shorter is worth a large premium in the current market.
Bloom Energy
Bloom makes solid oxide fuel cells, which convert natural gas directly into electricity through a chemical reaction rather than by burning it to spin a turbine. The output is modular, the installation is fast, and the emissions profile is better than combustion.
The company reports having deployed 1.5 GW across more than 1,200 installations globally. Its 2026 Data Center Power Report, a survey of 152 decision-makers across hyperscalers, colocation developers, utilities and GPU service providers conducted in November 2025, found that roughly one third of data centers in 2030 are expected to run entirely on on-site power, and that 45% of respondents expect to adopt direct-current distribution architectures in new facilities by 2028.
Treat vendor-sponsored survey findings with the appropriate scepticism. Bloom has an obvious interest in the finding that data centers are leaving the grid. The survey is still useful, particularly its finding that utilities project delivery timelines roughly 1.5 to 2 years longer than their customers expect, which is a disagreement between two parties who both have reason to be optimistic.
Caterpillar, Cummins and Generac
Reciprocating engine generator sets sit between fuel cells and turbines on both speed and scale. Historically these were sold as backup power, running a few hours a year. The change in this cycle is that they are increasingly being sold as primary power in microgrid configurations, which is a different and much larger revenue opportunity, and a far more demanding duty cycle.
Cummins announced in June 2026 that it would supply natural gas generator sets and integrated microgrid controls for a high-performance computing data center in West Texas, with deliveries running from 2026 through 2030. Caterpillar has been expanding engine capacity, including a substantial investment in its Lafayette engine facility.
There is a durability argument specific to engines that does not apply to turbines. An engine sold as backup power runs a handful of hours a year and needs little service. An engine sold as primary power runs continuously and consumes parts and service revenue for decades. If the shift from backup to primary duty holds, it changes the aftermarket economics of these businesses permanently, independent of how many further data centers get built.
Part 6: Storage, Which Sells Speed Rather Than Energy
Grid-scale batteries do not generate electricity. They move it across hours. Their role in this story is almost entirely about interconnection.
A data center with battery storageBattery Energy StorageGrid-scale batteries that absorb electricity when it is plentiful and release it when it is scarce. Storage does not create energy, it moves energy across hours. Its role in the AI buildout is mainly speed: batteries can be installed in months, so they let a site smooth its demand and connect sooner.See all terms in the glossary on site can smooth its demand profile, shave its peak draw, and in some jurisdictions qualify for flexible connection arrangements that let it energise years earlier than a firm connection would allow. Utilities including Dominion and PG&E have introduced programmes along these lines. The battery is not an energy source. It is a queue-jumping device.
Tesla Energy and Fluence are the principal suppliers at scale. Both have been shipping increasingly dense products, with Fluence's Smartstack rated at 7.5 MWh per unit against Tesla's Megapack at up to 3.9 MWh, density mattering because constrained sites have limited acreage.
The caution here is that storage is the layer of this entire chain with the most competition, the most Chinese manufacturing capacity, and the fastest declining costs. Falling costs are excellent for the buildout and difficult for the suppliers' margins.
Part 7: The Electrons Themselves, and How Nuclear Got Re-Rated
Someone still has to generate the power, and the most interesting corporate story in this whole complex is what happened to merchant nuclear.
The Change Was Contractual, Not Physical
A merchantMerchant PowerA power plant that sells its output into the open wholesale market at whatever price prevails, rather than earning a regulated return on its assets. Merchant generators carry price risk and price upside, which is why a long-term contract with a hyperscaler changes their economics so sharply.See all terms in the glossary nuclear plant sells its output into the wholesale market at whatever price prevails. For two decades that was a difficult business. Reactors have enormous fixed costs, cannot ramp up and down economically, and were competing against cheap gas. Several closed.
Then hyperscalers arrived wanting baseloadBaseload PowerGeneration that runs continuously at a steady output rather than ramping up and down with demand. Nuclear and coal are traditional baseload sources. AI data centers want baseload because they run near flat out around the clock, which is why nuclear output has become unusually valuable to them.See all terms in the glossary power, carbon-free, at a known price, for twenty years. That is precisely what a reactor produces and precisely what the market had been undervaluing.
No new reactor was built to cause this re-rating. The same physical assets, generating the same electrons, became substantially more valuable because a new class of buyer appeared who wanted exactly the product they make and would sign a twenty-year contract for it. A long-term power purchase agreementPPA (Power Purchase Agreement)A long-term contract in which a buyer agrees to purchase electricity from a specific generator at an agreed price, often for 10 to 20 years. Hyperscalers use PPAs to lock in nuclear or renewable output years before they need it, which gives the generator revenue certainty and the buyer supply certainty.See all terms in the glossary converts a commodity price taker into a contracted infrastructure business, and the market values those two things very differently. This is a useful general lesson: sometimes an industry re-rates because its customers change, not because it does.
The Deals
Constellation Energy signed a twenty-year agreement with Microsoft in September 2024 to restart Three Mile Island Unit 1, an 835 MW reactor, now renamed the Christopher M. Crane Clean Energy Center. Constellation has accelerated the restart to 2027, roughly a year ahead of the original schedule, and a FERC waiver on grid connection rights granted in June 2026 removed a significant remaining obstacle. Project startup costs are estimated at around $1.6 billion, supported by roughly $1 billion of federal loan financing.
Talen Energy expanded its arrangement with Amazon to 1,920 MW through 2042, supplying data centers from the Susquehanna station in Pennsylvania.
Vistra signed a twenty-year agreement with Meta covering 2,609 MW of nuclear output.
Meta has contracted for up to 6.6 GW across agreements with TerraPower, Oklo, Vistra and Constellation, with the newer capacity targeted for delivery in the 2032 to 2035 window.
Google signed a master agreement with Kairos Power for an initial 50 MW demonstration reactor, scaling toward roughly 500 MW across multiple sites by 2030.
The Small Modular Reactor Reality Check
Small modular reactorsSMR (Small Modular Reactor)A compact, factory-built nuclear reactor. Several data center operators have signed agreements for future SMR capacity, though the economics are unproven at scale and meaningful deployment remains years away.See all terms in the glossary attract enormous attention and very little near-term electricity. Oklo, NuScale, X-energy, TerraPower and Kairos are all developing designs, and the first hyperscaler-procured units are expected to deliver first power around 2030 at the earliest.
Small modular reactors solve the 2032 problem, not the 2027 problem. Any investment case in this area is a bet on execution against regulatory approval, first-of-a-kind construction cost, and a supply chain that does not yet exist at volume, with essentially no revenue in the interim. That is a venture-shaped risk profile wearing an infrastructure label. Position sizing should reflect which of those two things it actually is.
Part 8: Cooling
Every watt consumed becomes a watt of heat. As rack densities have climbed, direct-to-chip liquid cooling has become standard rather than exotic.
Vertiv is the largest specialist, and the only sizeable company covering both power and cooling. Its second quarter 2026 results reported net sales of $3,274 million, up 24%, with 18% organic growth, adjusted operating margin of 22.6%, up 410 basis points, and adjusted free cash flow of $925 million. Full year 2026 guidance implies about $14 billion of revenue and 31% organic growth.
As noted earlier, that release disclosed no backlog figure or book-to-bill ratio.
Munters, Modine, Johnson Controls, Trane and nVent all compete across thermal management and precision cooling. The competitive moat in this layer is the two to three year qualification cycle that hyperscalers impose before designing in a new cooling vendor, which protects incumbents and slows the arrival of new capacity. Vertiv's position in this layer was covered in more detail in the hidden AI stack.
Part 9: The Layer With the Least Competition and the Lowest Margins
Somebody has to install all of this, and the contractors who do it are in a genuinely unusual position: demand vastly exceeds their capacity, and yet their margins are a fraction of the equipment makers'.
Quanta Services is the largest. It reported record backlog of $48.5 billion in the first quarter of 2026, comprising $28.2 billion of twelve-month backlog and $26.2 billion of remaining performance obligations, with the Electric Power Infrastructure Services segment accounting for $40.1 billion. That segment generated $6.47 billion of revenue at an operating margin of 8.7%.
Compare that with Siemens Energy Grid Technologies guiding to 18% to 20%, GE Vernova Electrification at 18.4%, or Eaton Electrical Americas at 25.6%.
Margins of roughly 18% to 26%.
Constrained by factory capacity, specialised materials and supplier qualification.
Scale, intellectual property and decades-long qualification create genuine barriers. A new entrant cannot appear.
Capacity is expanded by building factories, which is capital-intensive but plannable and dated.
Margins of roughly 8% to 10%.
Constrained by the availability of trained crews.
Barriers are real but different: safety records, utility relationships, union agreements and the scarcity of qualified people.
Capacity is expanded by training people, which is slower, less certain, and cannot be bought.
Margins of roughly 18% to 26%.
Constrained by factory capacity, specialised materials and supplier qualification.
Scale, intellectual property and decades-long qualification create genuine barriers. A new entrant cannot appear.
Capacity is expanded by building factories, which is capital-intensive but plannable and dated.
Margins of roughly 8% to 10%.
Constrained by the availability of trained crews.
Barriers are real but different: safety records, utility relationships, union agreements and the scarcity of qualified people.
Capacity is expanded by training people, which is slower, less certain, and cannot be bought.
Quanta is responding to this by moving upstream, committing $500 million to $700 million to roughly double its own transformer manufacturing capacity and expanding its off-site manufacturing and fabrication footprint toward around 6.7 million square feet. That is a contractor attempting to capture some of the manufacturer's margin, and whether it works is one of the more interesting strategic questions in the sector.
MYR Group, EMCOR, Comfort Systems USA and Sterling Infrastructure occupy adjacent positions across electrical construction, mechanical systems and site work.
There is a paradox worth sitting with. The contractors are the most severely capacity-constrained layer in the entire chain, and they earn the lowest margins in it. Scarcity does not automatically confer pricing power. It confers pricing power only when the scarce resource is something you own. The equipment makers own factories and intellectual property. The contractors mostly rent labour that could, in principle, go and work for a competitor tomorrow. Owning the constraint is what matters, not being constrained.
Part 10: How the Hyperscalers Are Routing Around All of This
The buyers are not passive. Faced with a four to seven year queue, they have developed six distinct strategies, and each one redirects money to a different set of suppliers.
Strategy six deserves more attention than it gets. Every other approach competes for scarce physical assets. Efficiency creates capacity out of engineering. If the industry moves from a PUE of 1.3 to 1.1, it has effectively freed up a meaningful share of its power procurement without building anything. That is also why the electrical architecture transition matters commercially: an 800 volt direct current rack standard reshuffles which component vendors are designed in, and being inside that transition is worth a great deal to the companies concerned.
Part 11: The Bear Case
This industry has done this before, and it ended badly.
The Historical Rhyme
Between roughly 2008 and 2015, the gas turbine market went from boom to severe oversupply. Post-crisis electricity demand growth disappointed, renewables took share faster than forecast, and manufacturers who had expanded into the boom carried expensive idle capacity for years. The damage to General Electric's power business was one of the largest industrial value destructions of the decade, and it is a direct ancestor of today's GE Vernova.
The people running these companies remember it, which is why they are expanding capacity so cautiously against contracted rather than forecast volume. That discipline genuinely reduces the risk of a repeat. It does not eliminate it.
The Specific Risks
Multiple independent demand drivers. Grid equipment demand comes from AI, general electrification, renewable integration and age-related replacement. Losing one does not end the cycle.
Ageing infrastructure is not optional. Over half of American distribution transformers are past 33 years old.
Supplier discipline. Capacity is being added against contracted volume, not forecasts, which limits overshoot.
Service annuities. Installed turbines and engines generate high-margin service revenue for decades after the equipment sale.
Long backlogs delay damage. Even a sharp order slowdown takes years to reach revenue.
Concentrated buyers. A handful of technology companies drive the marginal order. Their capital expenditure decisions are correlated and can change fast.
Off-balance-sheet commitments. Reporting through 2026 has flagged very large AI-related obligations sitting outside the major technology companies' balance sheets. In August 2026, shares across the behind-the-meter energy complex fell sharply in a single session on such a report, which shows how tightly these are now coupled to AI capital expenditure sentiment.
Backlog is not a guarantee. Slot reservations can lapse. Contracts can be renegotiated. Firmness is what matters, and much of the headline volume is not firm.
Valuations embed durability. This complex has re-rated substantially. Prices already assume the buildout continues, which means being right about the shortage is not sufficient to make money.
Capacity arrives together. Every expansion announced in 2025 and 2026 lands around 2028. If demand moderates before then, relief becomes glut quickly.
Multiple independent demand drivers. Grid equipment demand comes from AI, general electrification, renewable integration and age-related replacement. Losing one does not end the cycle.
Ageing infrastructure is not optional. Over half of American distribution transformers are past 33 years old.
Supplier discipline. Capacity is being added against contracted volume, not forecasts, which limits overshoot.
Service annuities. Installed turbines and engines generate high-margin service revenue for decades after the equipment sale.
Long backlogs delay damage. Even a sharp order slowdown takes years to reach revenue.
Concentrated buyers. A handful of technology companies drive the marginal order. Their capital expenditure decisions are correlated and can change fast.
Off-balance-sheet commitments. Reporting through 2026 has flagged very large AI-related obligations sitting outside the major technology companies' balance sheets. In August 2026, shares across the behind-the-meter energy complex fell sharply in a single session on such a report, which shows how tightly these are now coupled to AI capital expenditure sentiment.
Backlog is not a guarantee. Slot reservations can lapse. Contracts can be renegotiated. Firmness is what matters, and much of the headline volume is not firm.
Valuations embed durability. This complex has re-rated substantially. Prices already assume the buildout continues, which means being right about the shortage is not sufficient to make money.
Capacity arrives together. Every expansion announced in 2025 and 2026 lands around 2028. If demand moderates before then, relief becomes glut quickly.
The Correlation Trap
A specific portfolio risk worth naming. It is easy to construct what feels like a diversified basket here: a turbine maker, a grid equipment company, a cooling specialist, a contractor, an independent power producer. Five companies, five industries, five business models.
They share a single demand driver. If hyperscaler capital expenditure decelerates, all five reprice together and in the same direction. That is not diversification, it is one position expressed five ways, and the guide to building a non-correlated portfolio covers why this pattern is one of the more common ways investors misjudge their actual exposure.
The same caution applies to owning these alongside the semiconductor and hyperscaler names in AI data center economics. They are all downstream of the same capital expenditure decision.
Part 12: What to Watch, Quarter by Quarter
A monitoring checklist that would give you early warning, roughly in order of how quickly each turns.
| Indicator | Where to find it | What a turn looks like |
|---|---|---|
| Book-to-bill ratios | Quarterly releases from Siemens Energy, GE Vernova, Eaton | Falling through 1.0 while revenue still looks strong |
| Slot reservation conversion | GE Vernova quarterly gas disclosure | Reservations stop converting into firm orders |
| Capacity auction clearing prices | PJM and other grid operator auction results | Clearing well below the price cap |
| Interconnection queue withdrawals | Grid operator queue reports | Rising withdrawal rates, revealing how much demand was speculative |
| Announced versus energised gigawatts | Industry trackers and utility filings | The gap widening rather than closing |
| Replacement versus new-build mix | Company commentary on order composition | A shift toward speculative new build, which is more fragile |
| Lead times | Trade press and manufacturer commentary | Falling faster than expected, signalling demand rather than supply relief |
| Hyperscaler capital expenditure guidance | Quarterly results from the major technology companies | Any moderation in the growth rate, which is the upstream driver of all of it |
Valuation discipline matters as much as any of these. Being correct that a shortage exists is not the same as making money from it, because the shortage is widely known and substantially priced. The framework in valuation 101 applies with particular force to cyclical industrials at the top of an order cycle, where earnings are peak and multiples often look deceptively reasonable precisely when they are most dangerous.
A Note for Indian Readers
This article deliberately covers the United States and global suppliers, because that is where the bottleneck is most acute and the disclosure is most complete.
The Indian angle is genuinely substantial and is covered separately. Indian electrical equipment manufacturers are exporting into this global shortage, and India is running its own data center buildout with its own power constraints. The hidden AI stack maps the Indian listed companies across cables, electronics manufacturing and power utilities, and AI data center economics covers India's data center capacity trajectory and the listed exposures to it. The commodity layer beneath both, copper and uranium, is covered in the AI commodity trade.
Sources
- GE Vernova, second quarter 2026 Form 8-K, filed 22 July 2026
- Siemens Energy, Earnings Release Q3 FY 2026
- Eaton, first quarter 2026 results
- Vertiv, second quarter 2026 results, published 29 July 2026
- PJM, base residual auction results for the 2026/2027 delivery year
- POWER Magazine, Transformers in 2026: Shortage, Scramble, or Self-Inflicted Crisis?
- Bloom Energy, 2026 Data Center Power Report
- Cummins, natural gas generators for West Texas data centers, June 2026
Frequently Asked Questions
Which companies actually benefit from the AI power buildout?
The direct beneficiaries sit in the middle of the chain. Gas turbines go to GE Vernova, Siemens Energy and Mitsubishi Heavy Industries. Grid and transformer equipment goes to Hitachi Energy, Siemens Energy Grid Technologies, GE Vernova Electrification, Eaton, Schneider Electric and ABB. On-site power goes to Caterpillar, Cummins, Generac and Bloom Energy. Nuclear output goes to Constellation, Vistra and Talen. Installation goes to Quanta Services, MYR Group, EMCOR and Comfort Systems.
What is the difference between backlog and orders, and why does it matter?
Orders are what a company won in a period. Backlog is the total won but not yet delivered. Book-to-bill is orders divided by revenue, so above 1.0 means backlog is growing. What matters is not the size of the backlog but its firmness, whether the customer can cancel without penalty, and whether prices can be revised if input costs rise. GE Vernova's gas figures illustrate this: of 116 GW under contract, only 53 GW was firm backlog and 63 GW was slot reservation agreements.
How are hyperscalers getting around the power bottleneck?
Six strategies. Building generation behind the meter to skip the queue. Signing long-term nuclear power purchase agreements. Agreeing to curtail consumption during peak hours in exchange for faster connection. Siting campuses where power already exists. Bridging with fuel cells and engines that install in months. And improving efficiency, because a watt saved is a watt that does not need to be procured.
Is nuclear power going to solve the AI energy problem?
Restarted and uprated conventional reactors help before 2030. Small modular reactors mostly do not, because the first hyperscaler-procured units are expected to deliver first power around 2030 or later. The nearer-term nuclear story is contractual rather than physical: twenty-year power purchase agreements converted merchant nuclear generators from commodity price takers into contracted infrastructure businesses without a single new reactor being built.
What would go wrong for these companies if AI spending slowed?
A long backlog delays damage rather than preventing it. Orders fall first, then book-to-bill drops below 1.0, then backlog erodes, and revenue declines several quarters later. The precedent is the gas turbine market of roughly 2008 to 2015, when post-crisis demand collapsed and manufacturers carried idle capacity for years. Current valuations across this complex embed an assumption that the buildout is durable.
What should an investor actually monitor each quarter?
Book-to-bill ratios, because they turn before revenue. The conversion rate of slot reservations into firm orders. Interconnection queue withdrawal rates. Capacity auction clearing prices. Announced versus energised gigawatts. The replacement versus new-build mix in order books. And hyperscaler capital expenditure guidance, which sits upstream of all of it.
Key Takeaways
-
The profit sits where the bottleneck sits. Fuel producers compete on a commodity and rack power distributors compete on price. The companies in the constrained middle, selling custom electrical equipment with multi-year lead times to customers who cannot wait, are the ones raising prices and expanding margins at the same time.
-
Read backlogs with three questions, not one. How firm is the commitment, can the price be revised, and what is it made of. GE Vernova's disclosure that 63 GW of its 116 GW is slot reservations rather than firm orders is the clearest illustration of why the headline number misleads.
-
The grid equipment layer may be the better business than turbines. It has four independent demand drivers rather than one, and the replacement of America's ageing transformer fleet continues whether or not another data center is built.
-
Scarcity does not equal pricing power. The installation contractors are the most capacity-constrained layer in the chain and earn roughly 8% to 10% margins against the manufacturers' 18% to 26%. What confers pricing power is owning the constraint, not suffering from it.
-
Nuclear re-rated without building anything. Twenty-year contracts with hyperscalers converted merchant reactors from commodity price takers into contracted infrastructure. Small modular reactors, by contrast, address the 2032 problem and carry venture-shaped risk under an infrastructure label.
-
Efficiency is the workaround nobody discusses enough. Every other strategy competes for scarce physical assets. Better cooling and higher-voltage rack architectures create capacity out of engineering, and the transition toward 800 volt direct current will reshuffle which component vendors are designed in.
-
A basket of these names is one position, not five. A turbine maker, a grid supplier, a cooling specialist, a contractor and a power producer all share a single demand driver. If hyperscaler capital expenditure decelerates, they reprice together. Being right that the shortage exists is not the same as making money from it, because it is already substantially priced.
Disclaimer
Nothing on this site is investment advice. All content is for educational and informational purposes only. Do your own research and consult a registered financial adviser before making any investment decisions.
Finished reading? Mark this article to track your learning progress.
Software Engineer, Self-Taught Investor
Software engineer who started learning about money in 2016 after a layoff coincided with a new home loan. Went from bank deposits to mutual funds to picking stocks in India and the US, learning through YouTube, screener.in, TradingView, and the hard way. Still learning. This site is her notes made public — for education and sharing only, not financial advice.
