AI Infrastructure Investing: Chips to the Grid

When investors discuss artificial intelligence, attention usually goes to the model, the chip and the cloud. Yet an AI system also depends on power quality, data movement between racks, heat removal, grid access, permissions for software agents and the financing that connects all of those pieces.

This Guide turns those less visible layers into a dated investment thesis. It is not a list of stocks that are expected to rise. It records the assumptions we are making today, the evidence that can test them and the places where the thesis may prove wrong.

The short answer

As AI capacity expands, economic value may migrate not to the company selling the most visible model but to the system layer that becomes indispensable and difficult to replace. In the first phase, those layers may be power quality, data movement, optical connectivity, cooling and grid access. Over a longer horizon, identity, permissions, observability, workflow and industrial control may matter more.

This is a system thesis, not automatically a stock thesis. Demand pressure in a layer does not mean every company in that layer will win, or that its stock is cheap. To identify durable economic surplus, we need to see scarcity, multi-customer use, recurring revenue and margins that survive the cost of capital expansion.

Why invisible systems?

The “AI winners” narrative often treats product visibility as evidence of economic value. That makes the story easy to tell, but it can point the research in the wrong direction. A better model does not make the power-quality equipment, optical link, liquid-cooling loop or identity layer irrelevant. It may make them more important.

Invisible systems have a distinctive economic feature. They are not usually sold directly to consumers. They sit inside the data centre, server or enterprise workflow. Public attention can therefore be low. But if failure is expensive or substitution takes a long time, invisibility can coexist with pricing power.

The narrow version of the thesis is:

Binding bottleneck + slow substitution + multi-customer use + surplus after capacity costs = an investable system layer to investigate

If one of those conditions is missing, the thesis narrows. A product can be essential but commoditised. Orders can grow but reflect a single customer's capacity reservation. Revenue can rise while factory, energy, warranty and debt costs erase the economic surplus.

The 2026 baseline: the bottleneck is moving beyond chips

The most important independent system source is the International Energy Agency's work on AI and energy. The IEA projects that data-centre electricity consumption could rise from 485 TWh in 2025 to 950 TWh in 2030, while AI-focused consumption could roughly triple over the same period. This is not a company revenue forecast. It is a system baseline showing that the AI economy is constrained by more than the supply of accelerators.

The IEA also estimates that an advanced AI server rack could have peak power demand equivalent to roughly 65 households by 2027. It says AI server power density rose 11-fold from 2020 to 2025 and discusses the possibility of 20-25 GW of battery storage in data centres by 2030. The investment question is not which battery manufacturer to buy. It is whether managing fast-changing AI loads creates a more valuable system layer than conventional backup power.

1. Power quality: from the amount of electricity to its behaviour

AI racks do not merely consume a lot of electricity. They can change their load quickly. That makes voltage, frequency and transition behaviour important alongside the total amount of power. Medium-voltage UPS systems, synchronous condensers, flywheels, battery storage, microgrid controls and power electronics become relevant here.

ABB's second-quarter 2026 release makes this layer concrete. The company describes synchronous condensers and flywheel technology with VoltaGrid as a “shock absorber” for rapid AI load changes. It also highlights medium-voltage UPS, prefabricated eHouse systems and control panels for data-centre grid connections and power quality. This does not establish AI-only revenue. It is first-party evidence that AI factories may require a new power-equipment design.

A reference architecture from Siemens, NVIDIA and Fluence points in the same direction. Its 136 MW facility design includes a 100 MW IT load and batteries for voltage and frequency ride-through, black start, demand response and AI-load smoothing. This is not proof of a global standard. It is an interesting signal that power, cooling, controls and data-centre management are beginning to be designed as one system.

2. Data movement: the problem may be distance between chips

An AI system can have enough compute units and still fail to operate at full capacity if data cannot move at the right time. That makes PCIe, CXL, UALink, NVLink Fusion, fabric switches, optical I/O and advanced packaging potential bottleneck layers.

Astera Labs reported second-quarter 2026 revenue of $392.4 million and described the production ramp of its Scorpio X-Series 320-lane fabric switch as an important inflection point for the next period. Its platform spans PCIe, CXL, Ethernet, NVLink Fusion and UALink. This does not mean Astera will win. The narrower conclusion is that data movement can be identified as a separate product and software layer.

Coherent says AI data-centre architectures are increasingly moving from copper to optical connectivity and that the transition is expanding manufacturing capacity. Fabrinet, which provides advanced optical packaging and precision manufacturing services, reported FY2026 revenue of $4.6 billion, up 36% year on year. These figures are not AI-only revenue. They show why optical connectivity is a research area, while customer concentration and rapid price declines remain central risks.

The real test is not whether optical or fabric products are technically necessary. It is whether multiple customers use them across multiple racks and product generations at an acceptable margin.

3. Time-to-power: who gets the site running faster?

A completed data-centre building cannot generate revenue if it cannot connect to the grid. Cable, transformers, medium-voltage distribution, interconnection, permitting, prefabricated power systems and commissioning time can therefore become a separate bottleneck.

Prysmian reported 18% organic growth in Digital Solutions and 13% growth in Power Grid for the first half of 2026. The company describes North American data-centre demand as an important driver. NKT says demand in distribution remains robust but capacity constraints limit growth. It says additional medium-voltage capacity in Denmark was completed while Portugal capacity was expected by the end of 2026.

These examples show both the strength and the risk of the time-to-power thesis. If cable and distribution capacity is scarce, orders and backlog can rise. But expanding capacity can pressure margins and cash conversion. Scarcity does not automatically leave a high and durable economic rent with the supplier.

4. Heat: liquid cooling may be mandatory, but who keeps the margin?

Higher rack power means more heat. Liquid cooling, coolant distribution units, pumps, heat exchangers, sensors, leak detection and maintenance may become critical parts of data-centre architecture.

nVent reported second-quarter 2026 sales of $1.471 billion, up 53% on a reported basis. It said Systems Protection sales rose 70% and that it was expanding liquid-cooling manufacturing capacity for data-centre growth. Modine announced a $4 billion long-term capacity agreement through 2029 for Airedale by Modine cooling solutions.

These are strong signals, but they are also a warning. A long-term capacity agreement can improve demand visibility. It does not prove that all of the revenue has already been realised, that margins will hold or that the customer will use the capacity profitably. The cooling thesis needs to be tested through installation, repeat orders, service revenue and multi-customer use.

The one-, three-, five- and ten-year map

One year: the physical installation bottleneck

By 2027, our highest-confidence view is that AI racks will add power-quality and thermal requirements to existing data-centre systems. The first research universe therefore includes ABB, Astera Labs, Coherent, Fabrinet, Prysmian, NKT, nVent and Modine, with orders, backlog, capacity, customer diversity and margins as the leading indicators.

The one-year forecast is not a share-price forecast. It is a system forecast: power quality, liquid cooling, optical connectivity and time-to-power will become more visible in AI-factory designs. If that happens, the first research threshold has been crossed. It does not mean any stock is cheap.

Three years: which standard and data-movement architecture wins?

By 2029, the question is which connectivity and power architectures move from pilot projects to default designs. CXL, UALink, NVLink Fusion, fabric switches and optical I/O may compete while different standards gain or lose ground. Technical adoption matters, but so does whether the standard makes suppliers replaceable.

In the positive three-year case, system suppliers sell more than a component. They build longer customer relationships through software, management, service, design and multi-generation compatibility. In the negative case, the products standardise, competitors multiply and data-centre customers push prices down.

Five years: does the physical system meet a control plane?

By 2031, an AI factory may be more than a collection of racks. It may have an integrated control layer that decides when to connect to the grid, when to use batteries, which load to reduce, how to tune cooling and which agent can access which system.

ServiceNow says AI ACV exceeded $1 billion in the second quarter of 2026, reported cRPO of $13.2 billion and described AI Control Tower as a layer for discovery, observation, governance, security and measurement. Okta positions human and AI-agent identity within the same platform framework. Datadog is extending into agent observability and reinforcement-learning operations.

These companies are worth tracking in the long-term part of the thesis. But the evidence today is not that agent control is already mandatory. It is that companies have started to productise the need. Agents must enter high-permission enterprise workflows, generate renewable contracted revenue and justify the cost of security and rollback before this becomes a durable economic layer.

Ten years: where does the durable relationship sit outside the model?

It is not possible, or particularly useful, to make a precise company forecast for 2036. The more defensible question is where AI will be embedded in the economy.

One possibility is the industrial digital thread. Models, digital twins, control systems, robots, machines and maintenance data could become part of the same production cycle. ABB describes adding analytics and AI to existing automation systems, while FANUC discusses Physical AI and autonomous machine tools. These statements do not yet establish a broad and measured profit pool. They do show that AI value may emerge on the factory floor rather than only inside the data centre.

At a ten-year horizon, confidence should be placed in the measurement method rather than a company name. A production deployment, lower downtime, better cycle economics and recurring software or service revenue are evidence. A pilot and a promise are not.

The research universe: how the candidates are separated

This file separates companies into three groups:

Group	Companies and layer	Why
Deepen research	ABB; Astera Labs; Coherent and Fabrinet; Prysmian and NKT; nVent and Modine	Direct product, order or capacity evidence for a specific physical bottleneck
Wait for a trigger	Fluence; Itron; ServiceNow; Okta; Datadog; Siemens; Celestica	Interesting mechanism, but AI-only revenue, multi-customer use, margins or production evidence are not yet clean enough
Benchmark	NVIDIA; Broadcom; TSMC; ASML; Vertiv	Core reference points for the chain, not undiscovered names

This is not a performance ranking. Current valuation, expectations, debt structure and share price are outside this first version. A system thesis can be right while a stock offers a poor return because expectations are already high. The next phase should compare valuation, expectations, returns on capital, debt and customer concentration.

The thesis falsification contract

A thesis becomes useful when it says not only what it believes but when it will give way. The main hypotheses are:

H1-PQ: Power-quality equipment becomes a durable AI data-centre layer. If orders remain limited to pilots by 2028 or products commoditise with lasting margin loss, the thesis narrows.
H2-DM: Data movement through CXL, UALink, fabric and optical I/O becomes a separate economic layer. If production deployments remain pilot-scale in 2029 or standardisation erodes margins, the thesis is falsified.
H3-TP: Cable, transformer and grid access remain the AI site's time-to-power bottleneck. If lead times, backlog and capacity pressure normalise by 2028, the thesis weakens.
H4-THERM: Liquid cooling and thermal systems become recurring revenue layers. If orders do not translate into installation, service revenue and multi-customer use, the thesis narrows.
H5-CTRL: As agents spread, identity, permissions, observability and rollback become a required control plane. If agents remain low-permission assistants and AI-only contracted revenue stays invisible, the thesis weakens.
H6-PHYS: Physical AI and the industrial digital thread become durable layers connecting model output to production. If pilots do not reach production or show measurable customer returns by 2032, the thesis weakens.

These are not immutable scientific thresholds. Writing them down now makes it harder to rewrite the story after the outcome is known.

What could go wrong?

The strongest counter-thesis is that AI investment is absorbed by efficiency. Better software, smaller models, custom silicon and lower energy use per task could slow total demand. The IEA reports rapidly falling energy use for simple AI tasks, while video, reasoning and agentic use can be far more energy-intensive. The outcome is not one-way: efficiency can reduce demand while new use cases increase it.

The second risk is over-investment. The BIS describes how AI infrastructure can be financed through debt, leases, offtake agreements and guarantees that create economically debt-like risks. Its investment-race model also produces over-investment relative to the socially efficient level under particular assumptions. This does not prove that AI demand is fictional. It says that revenue growth must be tested alongside financing quality and capacity utilisation.

The third risk is bundling by large customers. A cloud provider may combine power, networking, cooling and software into one offer. An independent supplier may remain technically necessary while leaving most of the economic surplus with the customer. For each company, the question is therefore not only “is the product necessary?” but also “can the company set the price?”

How will we revisit this?

This article should be reread in 2027, 2029, 2031 and 2036 using the same metrics. Each update should use one of four outcomes:

Held: The pre-written indicator moved in the expected direction and counterevidence did not break the thesis.
Narrowed: The mechanism was right, but the expected layer or company universe was smaller than assumed.
Falsified: A pre-written failure condition occurred and the thesis is no longer defensible.
Insufficient data: There is not enough measurement or comparable disclosure to reach a conclusion.

For example, liquid-cooling installations may rise in 2028 while supplier margins fall. In that case, it would be wrong to say cooling was unnecessary. The system thesis held, while the company-level economic-surplus thesis narrowed. Likewise, an agent-control product may be used without producing renewal or profitability evidence. Usage and investment return must not be treated as the same thing.

Methodology and limitations

This Guide combines company disclosures, institutional sources such as the IEA and BIS, and selected technical system examples. Direct facts, management claims, institutional projections, model results and author inferences are labelled separately. Platform-level or company-level revenue is not presented as AI-only revenue.

In power quality, cooling and optical connectivity, much of the evidence comes from company statements. Those statements can show that a product, order or capacity plan exists. They do not independently prove a durable competitive advantage or future share returns. IEA projections are scenarios, not outcomes. BIS financing and over-investment work is used to understand system risk, not as an accusation or forecast about a named company.

The next phase should examine valuation, expectations, debt, customer concentration, returns on capital and free cash flow for each company. This Guide does not do that work. Its more basic purpose is to fix the question, the evidence and the date to which we will return.

Sources and further reading

Not investment advice; for research and educational purposes.