Power and Cooling: Investment Notes for the AI Factory Era

A Research Report on Rubin 800V, Liquid Cooling, Modularization, and AI Factory Power Architecture

Analysis Date: 2026-06-24

Opening Note

Let me put the conclusion up front: power and cooling systems are in the early-to-middle stage of a structural upgrade cycle. But the real change is not as simple as “liquid cooling replacing air cooling,” nor as simple as “800V replacing legacy power supplies.” Its essence is this: after NVIDIA Rubin, AI data centers are beginning to require power, thermal management, energy storage, control, installation, and warranty to be designed as one integrated system.

Push the logic one layer deeper: the companies with the strongest value capture in this cycle will not necessarily be the ones shipping the largest volume of any single component. They will be the companies that can participate in reference architectures from the source, bundle “electricity” and “heat” together, prefabricate systems in the factory, and take responsibility for actual on-site operation. In other words, the industry’s pricing logic is shifting from “selling equipment” to “selling control rights”: whoever can control how power is scheduled, how electricity is buffered, and how heat is removed will occupy the scarcest position in the next generation of AI factories.

This note is organized around that main thread, laying out the timing, essence, key positions, future trends, and specific companies in one coherent line. One clarification is necessary: every company mentioned below is only a personal research and tracking subject for me and does not constitute any buy or sell recommendation. Many of these views depend on my own inferences about capacity expansion, hiring, and order cadence, and they may be wrong. Please apply your own framework and make an independent judgment.

Chapter 1: Timing: Early Stage of the Industrial Upgrade, Not Yet Fully Reflected

Start with the demand base, which is verifiable and checkable. In its baseline case, the International Energy Agency (IEA) expects global data-center electricity consumption to rise to about 945 terawatt-hours (TWh) by 2030, roughly double the 2024 level; Goldman Sachs expects U.S. data-center power demand to increase from 31 gigawatts (GW) in 2025 to 41 GW in 2026 and 66 GW in 2027, more than doubling in two years. In other words, AI workloads, power density, and data-center electricity consumption together form a strong long-term demand root. But several things that are often conflated need to be separated: workload demand, planned megawatts, megawatts that can actually be delivered, and the revenue that ultimately shows up in company financial statements. A strong demand root does not mean every announced capacity plan can become equipment revenue on schedule. Projects still have to pass, in sequence, through customer return-on-capital hurdles, financing, grid-interconnection approvals, architecture freeze, and finally on-site acceptance. So the real expectation gap is not whether demand exists, but who can absorb that demand and turn it into profit and cash.

It is important to note, however, that today's main market thread is still the previous-generation and transitional-generation architecture: conventional AC power distribution, an in-rack power path around 54V, and high-density liquid-cooling deployments based on the Blackwell generation (GB200/GB300). Nvidia's Rubin generation (VR200) will begin entering a clearer design, certification, and delivery cadence in the second half of 2026. But pushing 800V DC into the mainstream at large scale is more likely to happen after Rubin Ultra in 2027, corresponding to the Kyber rack. Vertiv's plan to launch its 800V DC product line in the second half of 2026 to support Rubin Ultra in 2027 confirms this cadence.

The next few years, therefore, will most likely be a transition period in which multiple voltages and multiple cooling paths coexist, and the timeline needs to be separated clearly. The mainstream architecture in 2026 will be closer to a conventional AC facility base plus in-rack 50/54V, overlaid with an 800V high-voltage DC "bridge" on the rack or Pod side, while air cooling and liquid cooling coexist. In 2027, native 800V and higher-density racks begin commercialization. Further out, after 2028, rectification, energy storage, and control may move further upstream to the Pod, data hall, or even facility boundary. Viewed across these topologies, the durable value is not in any single sidecar, CDU, or chiller, but in the interfaces, safety certifications, factory testing, service, and operating data that can span those topologies. This is why "position matters more than orders." For investors, value will be highly concentrated in a few links, but that is not simple "monopoly." 800V, OCP, racks, and liquid cooling are simultaneously pushing open standards and multi-sourcing. The more accurate formulation is: positions with high certification barriers, high responsibility boundaries, high switching costs, and the ability to collect structural rent. This point will recur throughout the discussion.

Chapter 2: The Essence: Power and Heat Are Becoming One System

Structurally, the opportunity in this cycle should not be split into two isolated tracks: "power-supply stocks" and "liquid-cooling stocks."

The reason lies in the physics itself. AI training and intensive inference workloads are not smooth; they create synchronized peak power demand. Once rack density rises, the same force pulls on a long chain of components, from the current, copper, and busbars inside the rack, to cold plates, CDUs (explained later), and final heat rejection. Therefore, the companies that will be truly competitive in the future will integrate power and thermal management into unified systems and move toward becoming "full-stack solution providers" - Vertiv, Schneider Electric, and Eaton (ETN) are examples. Traditional power-equipment giants such as Siemens have also begun working with companies such as nVent (NVT), which provide thermal-management and liquid-cooling interfaces, to participate in NVIDIA's designs. NVIDIA itself has long ceased to be merely a GPU vendor; it is delivering a complete reference architecture for an "AI token factory."

Several underlying changes are taking place in this process, and each is worth explaining clearly.

2.1 Why High-Voltage Direct Current Is Necessary (What 800V Solves)

Historically, data centers supplied power by taking medium-voltage alternating current (AC) from the grid, stepping it down in stages, rectifying it, stepping it down again, and finally delivering it to the GPU. The future direction for AI factories is to move part of the conversion process "upstream, centralized, and into DC": use solid-state transformers at the facility entrance to convert medium-voltage AC directly into 800V direct current, deliver power close to the racks through high-voltage DC busbars, and then perform the final step-down stages at the rack, shelf, and board levels. In short, the old model was "many small power supplies performing distributed conversion independently"; the future model is "centralized rectification + high-voltage DC distribution + high-efficiency DC-DC conversion close to the chip."

There is a critical distinction here that is often overlooked: today's 48V/54V DC still performs AC-to-DC conversion inside the rack; it simply changes power distribution inside the rack from AC to DC. By contrast, 800V DC pushes the conversion step directly to the facility layer, leaving only one efficient DC step-down stage inside the rack. This is an architectural leap from "rack-level DC" to "facility-level DC"; it is not the same thing.

The reason for this change is that AI rack power density has already moved far beyond the carrying capacity of traditional low-voltage AC distribution and server-level power supplies. There are two mechanisms behind this.

First, conversion losses compound stage by stage. In a traditional path, each conversion stage may look highly efficient on its own, such as 96%, 97%, or 98%; but after being multiplied across stages, the losses become substantial. If each stage retains 97%, after five conversions only about 86% remains (0.97 to the fifth power) - meaning an additional loss of roughly 14% of the electricity. All of that electricity turns into heat, which then adds load to the cooling system. (The five stages and 97% here are only examples used to illustrate compounded losses. In real systems, the number of conversion stages and efficiency curves vary by topology and load factor, and should not be read as measured values for any specific platform.)

Second, the real problem in low-voltage systems is excessive current. Power equals voltage multiplied by current. When power rises while voltage stays unchanged, current rises sharply; once current becomes large, resistive losses in copper wiring, connectors, heat generation, and safety redundancy all deteriorate quickly. The essence of high-voltage DC is to use higher voltage to push current down. NVIDIA's data shows that, with the same amount of copper, 800V can deliver more than 150% more power, and a single rack can save roughly 200 kilograms of copper busbars. Current affects more than copper wiring; it also pulls along a chain of supporting equipment, including UPS systems (uninterruptible power supplies, the batteries that take over during an outage), PDU/RPP equipment (power-distribution equipment that distributes electricity to racks), switchgear, and related infrastructure. This redundancy crowds out space for GPUs, HBM, NVLink, network-switching chips, and liquid cooling, while directly increasing construction difficulty. The new architecture compresses multi-stage low-voltage AC distribution and distributed rectification into fewer layers of "centralized rectification + high-voltage DC distribution."

There is another detail directly tied to the investment opportunity: the current problem becomes most acute at the "last centimeter" near the chip. A GPU core operates at around 1V yet needs to consume thousands of amperes of current, while copper's resistive loss is proportional to the square of current (I^2R). At thousands of amperes, even a few milliohms of resistance on the circuit board can burn a meaningful amount of power and create localized hot spots. There are two responses. One is the upstream approach described above: raise voltage and reduce current. The other is to move the voltage regulator module (VRM) from beside the chip (lateral power delivery) to directly underneath the chip (vertical power delivery, VPD), shortening the high-current path to the millimeter scale and directly cutting resistive losses across the plane of the circuit board. This trend of "moving the power supply underneath the chip" is becoming an independent battleground. The beneficiaries are companies making intermediate bus converters (IBC), VRMs, and board-level power modules, as well as the packaging and substrate segments. It is also precisely where a company such as Flex can enter by participating earlier in chip-level power-supply design.

The investment implication is direct: value will migrate from ordinary UPS systems, low-voltage distribution, and traditional data-center power supplies toward new links such as high-voltage DC protection, solid-state circuit breakers, rectifier cabinets, busway systems, and board-level DC-DC modules.

2.2 The Three "Imbalances" Magnified by the High-Voltage Era, and the Three Types of New Hardware They Create

As systems move toward 800V and even higher voltages, three types of "imbalance" will be magnified, and each type of imbalance corresponds to a new category of hardware demand. This is the main framework for assessing investment opportunities in the power chain.

The first is the imbalance in voltage stability, or the problem of "multi-timescale energy stability." GPUs require extremely stable low-voltage power. If voltage becomes unstable, a GPU can raise current sharply in an extremely short time and drag the voltage lower; in severe cases, this may damage core components. The solution relies on high-end capacitors (high-capacitance MLCCs and other ceramic capacitors). These need to stabilize voltage at the nanosecond level and are typically mounted near the GPU's voltage regulator module (VRM, the regulator circuit that steps voltage down to the level usable by the chip), CPU power modules, and DC-DC outputs. The larger the voltage span and the higher the density, the greater the demand for these high-end capacitors.

The second is the imbalance in compute demand, commonly called a power spike. The mechanism is worth spelling out: a training job synchronizes thousands or tens of thousands of GPUs, so load can step up by tens of megawatts within seconds, or suddenly drop during all-reduce pauses or checkpoint saving. This kind of large and fast step change is exactly what the grid dislikes; it can shock frequency and voltage and trigger protection systems. As a result, more utilities are beginning to impose ramp-rate limits on data centers. In the past, UPS systems only needed to handle very short outages. Now the problem is larger, longer-duration power buffering and ramp smoothing. Therefore, the attachment probability of energy-storage systems - BESS (Battery Energy Storage System) - will rise significantly in projects with strict grid-interconnection requirements, volatile AI loads, clear peak-valley price spreads, or a need for behind-the-meter power supply. But whether BESS becomes a universal standard still depends on UPS topology, interruptible loads, gas-fired generation configuration, fire-safety codes, and project economics; this should not be treated as inevitable. The distinction must be clear: the high-end capacitors discussed above and BESS are completely different pieces of hardware. One handles nanosecond-level voltage stability, while the other handles power buffering from seconds to hours. But both are being forced into existence by the same thing: compute pulses.

Pushing this one step further, there is a structural trend: the role of energy storage in AI data centers is evolving from "backup power" into a "real-time power buffer." It must perform three tasks at once: smooth short-cycle power fluctuations, buffer load when the grid is congested, and arbitrage costs across peak and off-peak electricity prices. In other words, energy storage will increasingly resemble a "market maker" or "liquidity provider" in the power market: absorbing electricity when prices are low, discharging when prices are high, and stepping in when the grid is tight. This is also the core logic behind Fluence (FLNC), discussed separately later. At the same time, UPS systems, supercapacitors, BESS, and solid-state protection will be integrated into a "multi-timescale buffering system": supercapacitors for the millisecond level, batteries for seconds to minutes, and campus-level storage or on-site generation for the hourly level.

The third is the imbalance in voltage span. Future power delivery must step down from 800V to 48V and then to the roughly 1V required by the chip, a much larger span than in the past. Wherever there is a "voltage change," power semiconductors are required to perform the conversion. The larger the span and the higher the density, the more demanding each conversion stage becomes. Current no longer flows steadily; it jumps in high-frequency transients, requiring microsecond-level response and high-frequency control. Traditional silicon MOSFETs are not fast enough or efficient enough, so two new-material power semiconductors have been pushed to the front of the stage.

SiC (silicon carbide) withstands high voltage (600V to above 3300V), tolerates high temperature, has good stability, and is suited to high power. Its weaknesses are lower switching frequency than GaN, higher cost, and larger size. GaN (gallium nitride), by contrast, operates at extremely high frequency (megahertz level), responds very quickly, and can be made very small. Its weaknesses are lower voltage-withstand capability (not suitable for the 800V input stage) and slightly lower maturity. It is mainly used in board-level DC-DC, VRM, and GPU power delivery - the "last centimeter." There is a layer of logic here that is often overlooked: GaN's value is not only efficiency, but density. In switching converters, passive components such as inductors and capacitors are roughly inversely proportional in size to switching frequency. GaN switching at the megahertz level means magnetic components and capacitors can be made smaller, lifting the power density of the entire converter sharply and occupying less board area. In a 120 to 600 kilowatt rack, space itself is the scarcest resource, so GaN's pull is strongest at the board level and closest to the chip.

The division of labor can be summarized this way: SiC is the "backbone material" of high-voltage power, while GaN is the "nervous system" of high-speed power. The two are complementary. The high-voltage trunk (800V) uses SiC; intermediate conversion (48V) uses SiC plus traditional MOSFETs; low-voltage, high-speed GPU power delivery uses GaN. Demand for the SiC technology path is likely to be significantly amplified in the future, because the high-voltage trunk of AI factories is precisely its home market.

In U.S. equities, the purest SiC name is Wolfspeed (WOLF). It is vertically integrated across the full chain from SiC substrates and epitaxy to power devices, with no traditional business to provide a buffer. But this company is extremely high risk and must be assessed separately. It has just completed a Chapter 11 debt restructuring in 2025 (filed in late June and completed in late September), reducing debt by about 70%. The cost was that original shareholders were almost wiped out, and the company was reborn after debt-to-equity conversion. Its technology path is highly concentrated on one direction: "high voltage for AI data centers," with no other product buffer. The market has repriced it from an "EV-cycle SiC company" into a "core SiC supplier for AI power infrastructure," but it is also carrying an obvious contradiction: the AI narrative is strengthening, while oversupply, equity pressure, and still-weak profitability remain unresolved. Overall, it is a high-volatility name with strong option-like characteristics, not a core holding.

One additional metric-level trend is worth noting: in the future, the core metric for judging data centers will gradually shift from PUE (power usage effectiveness) toward output-based measures such as tokens/watt, tokens/MW, and revenue/MW. A data center with an attractive PUE but low GPU utilization and serious queueing is not a truly efficient factory. This will in turn affect the valuation logic for the entire power and cooling chain.

Chapter 3: Position Matters More Than Orders: Participating in Design at the Source

A key change in this cycle is that companies such as Vertiv, Schneider Electric, Siemens, nVent, and Fluence are no longer merely waiting for customers to issue tenders before submitting quotes. Instead, they are entering Nvidia's AI factory reference architecture ahead of time. This is equivalent to seeing the bid documents in advance, and in some cases even helping define them.

This position matters because the more complex next-generation architectures become, the less willing customers are to assemble and debug a pile of equipment on site that has not been jointly validated. Whoever can enter the reference design early is more likely to have its power cabinets, protection systems, cooling loops, control software, and services designed into the customer's "default solution."

But discipline is still required: a design-in is not the same as an order, and a reference architecture is not the same as backlog. More precisely, a "true design-in" is not a company saying in a press release that it "supports 800V." It means the company can enter the Rubin/Rubin Ultra reference design, participate in defining the boundaries for power and thermal management, complete joint validation before the platform is formally launched, and continue to be responsible for commissioning, operation, and warranty during actual customer deployment. A position that meets this standard is far more valuable than ordinary supply. What it truly improves is the probability of certification, customer trust, and the speed of subsequent orders; but whether it ultimately generates profit still has to be tested against orders, gross margin, cash flow, and warranty obligations.

Chapter 4: Full-Stack Integration: Power and Cooling Are Closing Each Other's Gaps

The industry's moves are now very clear: companies that used to focus on power are buying cooling assets, companies that used to focus on cooling are moving toward power and controls, and companies that used to make cabinets and connectors are moving toward liquid-cooling interfaces and modular systems. This is an unmistakable consolidation trend.

Eaton completed its acquisition of Boyd Thermal for approximately $9.5 billion in March 2026 and folded it into its electrical segment. After this transaction, Eaton's logic is no longer just electrical equipment, but a combined power-and-cooling portfolio from "grid-to-chip" supply to cooling. After acquiring Motivair, Schneider is no longer merely selling power distribution and software; it is filling in its capabilities in thermal management "near the chip." Siemens and nVent are collaborating on a reference architecture for NVIDIA AI data centers, and they have brought in Fluence's energy storage as well. This shows that energy storage, power distribution, protection, and liquid cooling are being drawn into the same design blueprint. Vertiv already had broad boundaries across power, thermal management, cabinets, prefabrication, and services, so in this cycle it looks most like a "system platform."

These companies can be viewed in several layers, rather than simply comparing who has more "AI exposure." At the top are full-stack platforms that can participate in source-level design and package power, cooling, and controls into a complete deliverable system with clear accountability, such as Vertiv, Schneider, Eaton, and the "Siemens + nVent + Fluence" combination. Their risks are high valuations, long project cycles, M&A integration, and heavy engineering responsibility. Below them is the interface and protection layer, including busbars, cabinets, connections, protection, and prefabricated electrical houses, such as nVent, Hubbell, Powell, and Eaton. The risk is that if they are only supplying to drawings, margins can easily fall back toward ordinary industrial products. Further down is facility-level thermal management, including chillers, fan walls, CDUs, and final heat rejection, such as Modine, AAON, Trane, JCI, and Carrier. The risks are insufficient absorption of capacity expansion, warranties, receivables and inventory, and customer concentration. Then there is the power-device layer, including onsemi, Infineon, MPS, Vicor, AOSL, Navitas, and Wolfspeed. After qualification, their revenue can ramp faster and their torque can be greater, but volatility is also higher. Finally, there is the on-site power and energy-storage layer, including Fluence, Quanta, EMCOR, Caterpillar, Cummins, and GE Vernova, which is affected by regulation, project cycles, gas-turbine slot availability, and capital intensity.

Chapter 5: Modularization and Industrialization

AI data centers will become increasingly complex. If all equipment is piled onto the site for assembly, engineering risk rises materially: power, cooling, controls, energy storage, and networking all have to be commissioned at the same time, and a problem with any single interface can delay the launch of the entire data center. For customers, idle GPUs that cannot be put into production carry an extremely high opportunity cost.

That is why the future direction will inevitably involve more factory prefabrication, more factory acceptance testing (FAT, meaning power, cooling, controls, redundancy, and failover are validated before equipment leaves the factory), and more standardized modules. This path will create several specific trends: the delivery unit will shift from "project engineering" to "industrial product"; standardized AI Pods (delivery units with fixed GPU counts, fixed power, fixed cooling, and fixed networking) will become the new unit; power, cooling, pump stations, and energy storage will be packaged as skid-mounted modules (skids, meaning equipment packages prefabricated in the factory and transported as complete units to the site for assembly); onsite construction time will decline, but system complexity will not go away. It will simply move from the site to design, simulation, interfaces, controls, certification, and supply chain synchronization. Over the longer term, standardized modules can also be procured, financed, leased, and even securitized in advance. Financialized arrangements such as "power module financing," "energy storage capacity contracts," and "cooling as a service" are very likely to emerge.

But modularization is not a one-sided bullish story. It has two sides. If a company can turn modularization into a repeatable product, both gross margin and turnover can be very attractive. But if it is merely moving onsite engineering risk onto its own balance sheet, then fixed-price contracts, interface errors, onsite rework, and accounts receivable collection will gradually erode profits. So when assessing a modularization company, the key is not how many orders it has announced, but three things: first, whether its modules are a repeatable platform; second, whether factory acceptance can truly reduce onsite commissioning; and third, whether project revenue can turn into cash rather than inventory, receivables, and warranty obligations.

The most direct stock tied to this theme is Flex (FLEX). It was originally one of the world's largest electronics manufacturing services (EMS) providers, and over the past two years it has packaged data center power, cooling, and compute integration into a high-growth business. Its scope runs from facility-level power distribution, in-rack power, and chip-level power modules, to direct-to-chip liquid cooling (from its acquisition of JetCool), and then to assembling compute, storage, networking, power, and cooling into complete systems that are factory tested before shipment. If the companies discussed earlier are selling "system design," Flex is selling the ability to "build the system, assemble it, test it, and ship it to the site." It is the closest fit for the positioning of an "industrialized assembly plant for AI data centers."

But one easily exaggerated claim needs to be corrected first: Flex is not an "AI factory operating system," nor is it "eating" the profits of Eaton, Vertiv, and Modine. At its core, it is an integrator and assembler that buys components from others and puts them together. Its moat lies in engineering, global capacity, and full-system testing, not in IP-heavy areas such as busbars, switches, or SiC. The real structural question for an integrator is whether it can capture rent while sitting between "component manufacturers with IP" and "hyperscale cloud customers with scale and bargaining power." I will return to this when discussing the specific stock later, because the answer is hidden precisely in the gross margin.

Chapter 6: A Misunderstanding That Needs Clarification: Liquid Cooling Will Not Replace Air Cooling, and 800V Will Not Eliminate Heat Rejection

A common misunderstanding needs to be clarified here: liquid cooling will not turn the data center into a place with no air cooling. It is true that more of the heat from GPUs and CPUs will enter the liquid loop, but networking, storage, power, the data hall environment, and residual thermal loads still need to be handled on the air side. Schneider Electric's reference design for Rubin is also not "liquid cooling only"; it is liquid cooling and the air side coexisting.

What really changes is the location of the profit pool, and NVIDIA Rubin's push to bring 45°C warm-water liquid cooling to the foreground is the key to understanding that shift. First, answer a question many people will ask: why does the new generation use "warm water" at 45°C instead of lower-temperature water? The key is that chips do not actually need cold water. The upper limit for GPU junction temperature is roughly above 85°C, and modern cold plates already have extremely low thermal resistance (on the order of about 0.03°C per watt), so even with 45°C inlet water and chip power above one kilowatt, junction temperature still has a large margin below the limit. In other words, the constraint is not "whether the chip is cool enough," but "where the heat goes and how expensive it is to reject it."

Once water temperature is raised to 45°C, the situation changes: in most climates, the facility side can directly use outdoor dry coolers (essentially large outdoor radiator banks) to reject heat, and for most of the year it does not need to run mechanical compressors - this is "free cooling." NVIDIA's AI factory reference design is moving exactly in this direction: closed-loop, dry-cooled, and almost no mechanical refrigeration (in some climates, chillers may be needed only about 1% of the year). This brings two benefits at scale. The first is power savings: cooling has historically accounted for as much as 40% of total data center electricity consumption, and every 1°C increase in cooling-loop temperature can reduce chiller energy consumption by roughly 2%-4%. Turning the chiller off in most cases means facility overhead falls sharply and more electricity is left for GPUs. The second is water savings: traditional cooling towers reject heat through evaporation and consume about 2.6 million gallons of water per megawatt per year, while a 45°C closed-loop dry-cooling circuit is "filled once, closed loop for life," with water use close to zero - directly relieving water as an increasingly hard site-selection constraint. Why not simply go lower or higher? Lower water temperature only creates thermal headroom the chip cannot use, while paying the energy cost of mechanical refrigeration; that is pure waste. Going higher faces two ceilings: as chip power keeps rising, junction-temperature margin narrows, and the higher the water temperature, the faster the coolant's chemistry degrades (following the Arrhenius law). So 45°C is a deliberately set upper limit: while keeping chips safe, tune the system as close as possible to the point where it can use free cooling in almost all climates. In investment terms, value will migrate away from low-temperature mechanical chillers and compressors toward dry coolers, CDUs, low-thermal-resistance cold plates, coolants and water treatment, manifold quick disconnects, and loop controls. The thermal-resistance metric of the cold plate itself also becomes a battleground for design wins. At the same time, higher heat-rejection temperature makes waste-heat recovery more realistic: the discharged heat is closer to the temperature range usable for building heating and low-temperature industrial heat demand. But this depends on geography, climate, distance to heat users, regulation, and economics; it does not work everywhere.

In this new thermal system, the role of the CDU (Coolant Distribution Unit) changes. To understand why it matters, first recognize that there are actually two water loops here: one is the "facility-side loop" (FWS), connected to the building and the outdoors, where water quality and temperature are managed more coarsely and shared; the other is the "chip-side loop" (TCS), which runs directly to the cold plates and must be clean, deionized, dosed with corrosion inhibitors, free of impurities, and separately controlled for pressure and flow. The CDU is the "liquid-to-liquid heat exchanger + pumps + controls" that sits between the two loops and decouples them: the coarse facility-side water does not directly contact the chips, while the clean chip-side loop is independently pressure-stabilized, flow-stabilized, isolated against leakage, and used to collect all the data. This is why it is called a "thermal-flow router" - it is the central node for data collection, control, and fault isolation. So when judging a cooling company in the future, the question cannot only be whether it "sells cooling equipment." The question is whether it controls cold-plate design, manifolds and quick disconnects, CDU controls, liquid compatibility, leak detection, and the ability to remove chip hot spots (high-power chips are a local hot-spot problem, not an average heat-generation problem) in a stable, predictable way.

At the stock level, companies such as Modine (MOD) are typical: the direction is right and demand is there, but the market has long known that data center cooling is growing and also knows about its large order. The real expectations gap has already shifted from "whether demand exists" to production-line yield, gross margin, and true cash flow after deducting customer prepayments - this part will be discussed in detail later when the individual stocks are covered.

Chapter 7: Where the Power Comes From: Core Constraints on the Supply Side

The preceding sections covered how electricity is used after it enters the data center. But from the supply side, the real bottleneck for the entire industry is an even more upstream question: where the power comes from, and how long it takes to arrive. There is no longer much debate on the demand side (the IEA and Goldman Sachs numbers are already on the table), but the supply side is becoming the hardest constraint. As a result, "speed-to-power" is replacing "GPU count" as the first threshold for whether a new data center can actually be built.

The most direct evidence is in gas turbines. GE Vernova's (GEV) gas turbine orders had already accumulated to about 100GW by the first quarter of 2026, and the company expects to reach more than 110GW by the end of 2026 including reserved capacity, effectively booking its production lines out to 2029 and 2030. Unit prices for new orders in the first half of 2026 were 10%-20% higher than in the fourth quarter of 2025. According to third-party estimates, gas turbine prices could rise to about $600/kW by the end of 2027, nearly three times the 2019 level. In other words, large gas turbines are no longer equipment that can be bought whenever needed; they are scarce products for which customers have to compete for production slots.

Production slots running into 2030 have forced three very practical alternatives, each of which is a new investment direction. The first is "bridge power": before large gas turbines arrive, smaller aeroderivative gas turbines, or even reciprocating gas generators (recip engine), are used to fill the gap. xAI used this type of unit to bring a data center online within months, benefiting engine manufacturers such as Cummins and Caterpillar. The second is "behind-the-meter" self-generation: building natural gas power plants directly inside the campus, bypassing the grid queue, with even a "physical infrastructure as a service" model emerging in which power plants and development land are packaged and sold to data centers. Behind this is another layer of policy pressure: starting in 2026, the United States is requiring technology companies to pay out of pocket for their own incremental electricity demand rather than pushing up residential power prices. The third, and the change most worth watching over the past two years, is fuel cells.

The fuel cell path deserves a separate discussion. Bloom Energy (BE) and Oracle expanded their solid oxide fuel cell (SOFC) collaboration to as much as 2.8GW. Project Jupiter in New Mexico will use up to 2.45GW of fuel cells as the campus's primary power source, replacing the entire originally planned gas turbine and diesel generator package and consolidating it into a microgrid campus. Bloom also signed an approximately $5 billion financing partnership with Brookfield, and in just the first 90 days of early 2026 secured about $7.6 billion of data-center-related contracts. This matters in two ways: first, fuel cells have truly moved from "pilot projects" into "campus-scale primary power"; second, the shift also pulls along the old diesel backup-power market. It also echoes the earlier industrialization thesis - in Bloom's own words, AI infrastructure needs to be built like a factory, at scale and at speed.

Further out is nuclear power, but the timing needs to be separated clearly. As of mid-2026, almost all hyperscale cloud providers had signed nuclear power agreements, with the public tally already exceeding 13 deals and nearly 10GW: Microsoft signed a 20-year contract with Constellation to restart Three Mile Island Unit 1 (Constellation is investing about $1.6 billion and plans to restart it in 2028; the contract value has not been disclosed. The "$16 billion" figure often cited externally is actually the Brattle Group's estimate of the project's contribution to Pennsylvania GDP, not the contract price), Amazon is building a campus next to the Susquehanna nuclear plant, Google and Meta are betting on small modular reactors (SMRs), and even NVIDIA has invested in an SMR company. But the key judgment is this: long-term contracts with existing nuclear plants and restarts are a "next two to three years" story (the first AI nuclear power supplies begin roughly in 2027-2028), while most truly new SMR builds will have to wait until after 2030. Nuclear power's unit construction cost is several times that of natural gas, and the math works only under the premise of "carbon neutrality." So nuclear is a clear but longer-term theme; in the short term, it has not yet translated into hardware orders.

Putting these threads together, a new scarce asset is taking shape: "power certainty." A site that can deliver "power included, no grid interconnection queue" within a compressed construction schedule is valuable in its own right. In the market, these "power-certain" sites can command rents about 15%-25% higher than sites constrained by the grid. So the key to this theme is not only who sells generation equipment, but who can turn "power" into a deliverable, financeable, repeatable product: gas turbines and bridge-generation units (GEV, Cummins, Caterpillar), fuel cells (Bloom), and developers and integrators that package generation, storage, and distribution into the campus. Also note the electrification theme discussed earlier: GEV's electrification segment (transformers, switchgear, high-voltage direct current) received more data center orders in a single quarter than in all of last year, while large transformers and switchgear themselves are going through a supply-demand "supercycle."

Chapter 8: Power and Compute Are Becoming Financialized

If we push the previous lines of analysis further, a larger picture comes into view, and it is also the deepest trend in this cycle: AI data centers are moving from an "IT system" into an "integrated power-thermal-compute industrial factory." Their output is tokens, their constraint is megawatts (MW), their risk is heat flow, and the core of their cash flow is power dispatch.

Along this trajectory, several things will happen in sequence.

First, compute scheduling and power dispatch will begin to converge. AI workloads - especially deferrable training, batch inference, video processing, and embedding updates - can be scheduled across time and geography according to power prices, carbon intensity, and grid interconnection constraints. Google has already been incorporating machine learning workloads into demand response - actively reducing AI power consumption when the grid is under stress - which shows that hyperscalers are turning "compute loads" into dispatchable power assets. Looking further ahead, researchers are already experimenting with "treating inference as a flexible load": by adjusting inference batch sizes and coordinating with energy storage to manage the power ramp of training, they can significantly reduce discharge pressure on storage systems. This means the future is not simply about adding more batteries, but about jointly optimizing "compute scheduling + battery dispatch."

Second, dynamic power pricing will enter AI's cost structure. The API prices users see in the future may already embed GPU depreciation, peak-valley electricity rates, storage, grid interconnection capacity fees, and even carbon and heat-rejection costs. When compute becomes a relatively stable commodity while power costs fluctuate meaningfully, the market is likely to develop financialized arrangements such as "compute forwards," "inference cost hedges," and "token capacity reservations." These are not trading products that will launch in the short term, but they indicate the direction of the industry structure - the output of AI factories will increasingly resemble a form of "financialized capacity."

Taken together, the end state is not a simple data center, but a chain: power procurement -> storage buffering -> thermal system constraints -> GPUs producing tokens -> tokens becoming revenue -> power prices, loads, failures, and carbon emissions forming new risk markets. Whoever can dispatch this entire chain will be close to what might be called an "AI Power Operating System (Power OS)." This Power OS is not a single piece of software, but a cross-layer system: it connects workload scheduling, energy storage, cooling control, and grid compliance, ultimately serving the optimization of the token economy.

This brings us back to the main thread at the beginning: the industry's pricing logic is shifting from "selling equipment" to "selling control." But "control" must be disaggregated, so equipment companies do not capture a "software premium" that does not really belong to them. It can be divided into four layers: workload scheduling rights (scheduling training, inference, storage, power prices, and cooling together) - a layer mainly controlled by hyperscalers, model platforms, and GPU platforms; power market participation rights (demand response, ancillary services, peak-valley arbitrage, and behind-the-meter power sources); facility power control rights (UPS, BESS, power distribution, peak shaving, and dynamic control); and thermal system control rights (CDUs, flow, temperature, fluid quality, and fault isolation). The latter two layers are where publicly traded equipment companies can truly enter - they can capture value from interfaces, services, and operating data, but the highest control rights over workload scheduling and the token economy will most likely remain with cloud providers and compute platforms. Therefore, Power OS should be treated more as a long-term hypothesis than as the natural property of any single listed equipment company. The public-market opportunity is to identify companies that are already embedded in this architecture, can collect rent from the "interfaces/services/data" layer, but have not yet been fully priced by the market as holders of "control."