US Companies That Supply to AI Data Centres: List of Hottest US Stocks that Benefit From AI

Nvidia may make the engine of the AI boom, but the next phase will be decided by three things a chip cannot create on its own: electricity, cooling and fast data movement. As AI campuses move from tens of megawatts to gigawatts, the companies controlling these physical bottlenecks can become just as important as the chipmakers themselves.

Let's break down how an AI data centre is built, which listed companies get paid at each layer, and where the strongest business economics and biggest risks sit.

Which Companies Benefit From the AI Data Center Boom?

The AI data center boom is no longer just about GPUs. Five of the world’s largest technology companies spent more than $400 billion on capital expenditure in 2025, and the International Energy Agency expects this spending to increase by another 75% in 2026. A large part of this money is going towards building the infrastructure needed to train AI models and serve millions of users.

The scale of this expansion becomes clearer when we look at electricity demand. Global data center electricity consumption is expected to more than double to around 945 terawatt-hours by 2030. In the US alone, data centers could account for 6.7% to 12% of total electricity consumption by 2028, compared with 4.4% in 2023.

But building an AI data center requires much more than purchasing GPUs. It needs servers to house the chips, networks to connect them, fibre to move data, electrical equipment to deliver power, cooling systems to control heat, and contractors to build the facility. This creates nine connected layers of companies that can benefit from the AI infrastructure expansion:

The useful investor question is not simply, “Does this company serve data centres?” It is, “Does AI spending materially reach its financial statements, and can the company keep attractive economics when supply catches up?” 

That distinction matters because a server company may report explosive revenue while keeping only a small margin, while a network or electrical supplier can grow more slowly but retain far more profit from each dollar of revenue.

Start With the AI Company: How an AI Data Center Is Built From the Ground Up

Imagine OpenAI, Anthropic, Meta or xAI wants to train a larger model and serve it to millions of users. Or imagine Microsoft, Amazon, Google or Oracle wants more cloud capacity for enterprise AI workloads.

The company first estimates how many accelerators it needs. But the real design starts with the power available at the site. A developer cannot plan 100,000 GPUs and then discover that the local utility can supply only half the electricity or that the building cannot remove the heat.

Think of an AI data centre like a Formula 1 car. The GPU is the engine, but the car does not finish a race without fuel delivery, cooling, tyres, electronics and a pit crew. The fastest engine only makes every supporting system work harder.

That is why the value chain begins with chips but does not end there.

Layer 1: AI Data Center Chips

NVIDIA - NVDA

NVIDIA remains the centre of AI compute because it combines accelerators, networking, interconnects and a mature software ecosystem. In the quarter ended April 26, 2026, NVIDIA reported $81.6 billion in revenue, up 85% year on year. Data Centre revenue reached $75.2 billion, including $60.4 billion from compute and $14.8 billion from networking. The networking figure nearly tripled from a year earlier, showing how NVIDIA increasingly earns from the connections around its GPUs as well as the GPUs themselves.

Our view: NVIDIA still has the strongest integrated platform in AI infrastructure. Its risk is not a lack of demand but the size of expectations, export restrictions and the possibility that hyperscalers shift a growing share of workloads to their own custom chips. Its advantage is that even this shift can leave NVIDIA exposed to networking, software and complete systems.

Advanced Micro Devices - AMD

AMD is the clearest large alternative in merchant AI accelerators. Its Instinct GPUs, EPYC server CPUs and Helios rack-scale systems give customers a second architecture and reduce dependence on one supplier.

AMD's Q2 2026 revenue reached $11.5 billion, up 50% year on year. Data Centre revenue more than doubled to $6.7 billion and represented about 58% of company revenue. That is important because AMD is no longer only an “option on future AI share”. The shift is already visible in reported numbers.

Our view: AMD offers meaningful competitive tension in AI compute, but its real test is not a benchmark win. It is whether customers can deploy its complete rack, networking and software stack at large scale with similar reliability and developer ease. Watch Data Centre revenue growth, gross margin and the pace at which Helios deployments move from announcements to repeat orders.

Broadcom - AVGO

Broadcom is unusual because it earns at two critical points: custom AI accelerators designed with hyperscalers and the networking silicon connecting huge clusters. Fiscal Q2 2026 AI semiconductor revenue was $10.8 billion, up 143% year on year, driven by custom accelerators and AI networking. Management expected the figure to reach about $16 billion in the following quarter. Broadcom also generated free cash flow equal to 46% of quarterly revenue, which shows the difference between merely participating in AI and capturing high-quality economics from it.

Our view: Broadcom may be the strongest hedge against a world in which hyperscalers build more proprietary AI chips. The trade-off is customer concentration and very high expectations around future custom-silicon programmes.

Simple linkage: More AI usage -> more computation -> more accelerators and custom silicon -> more chip and networking revenue for NVDA, AMD and AVGO.

However, these are already recognised AI leaders. The less obvious opportunity sits one or two layers further down the chain.

Layer 2: AI Servers and GPU Racks

Thousands of GPUs cannot be placed on shelves and switched on. They need CPUs, memory, storage, networking, power supplies, cooling manifolds, cables and management software assembled into tested systems.

Dell Technologies - DELL

Dell's AI Factory combines servers, storage, networking, software and services into a deployable platform. More than 5,000 customers were using the Dell AI Factory by May 2026. The financial scale has also changed quickly: Dell closed more than $64 billion of AI-optimised server orders during FY2026, shipped more than $25 billion and entered FY2027 with $43 billion in AI server backlog.

But investors should separate volume from economics. Dell's Infrastructure Solutions Group produced $60.8 billion in FY2026 revenue and $7.1 billion in operating income, an operating margin of roughly 11.7%. That is healthy for a systems business but far below the economics of NVIDIA, Broadcom or Arista.

Our view: Dell has the scale, customer relationships and cash generation to turn AI infrastructure into a large, durable business. The main question is how much profit remains after expensive components are passed through to customers.

Super Micro Computer - SMCI

Supermicro specialises in fast product launches, high-density racks and direct liquid cooling. Its latest Data Center Building Block Solutions can cover compute, storage, networking, power and cooling from 5 MW building blocks to a 1 GW campus. A standard unit can include 1,152 Rubin GPUs, while Supermicro says it has experience deploying clusters with more than 100,000 GPUs.

The opportunity is large, but the earnings quality is less stable. Gross margin fell to 6.3% in fiscal Q2 2026 before preliminary fiscal Q4 margin rebounded to an estimated 15% to 17%. The company also warned that some of the more than $60 billion of new fiscal Q4 orders could face cancellation or delays.

Our view: Supermicro offers some of the most direct AI rack exposure, but it also carries greater margin volatility, customer concentration, working-capital pressure and execution risk. This is the clearest example of why fast revenue growth does not automatically equal a high-quality AI business.

Simple linkage: More accelerators -> more complete servers -> more racks -> more systems revenue for DELL and SMCI.

Layer 3: AI Data Center Networking that Connects Thousands of GPUs

Training a large model involves constant communication across accelerators. If one part of the network is slow or unreliable, expensive GPUs sit idle. A network failure is therefore not just an IT problem. It is wasted compute capital.

Arista Networks - ANET

Arista provides high-speed Ethernet switches and the EOS software that manages them. Its new 1.6-terabit platforms are designed for clusters scaling from thousands to hundreds of thousands of accelerators. Meta, Microsoft and Oracle have publicly discussed working with Arista on next-generation AI fabrics. Some new platforms also use Linear Pluggable Optics, which Arista says can reduce interconnect power by about 60%.

The business economics are exceptional. Q2 2026 revenue rose 37.7% to $3.04 billion, while GAAP operating margin reached 45.4%. Few physical infrastructure suppliers combine that pace of growth with that level of profitability.

Our view: Arista has one of the strongest quality-growth combinations in the entire chain. Its biggest risks are heavy reliance on large cloud customers, competition from NVIDIA networking and the chance that hyperscalers use more internally designed hardware with open software.

Celestica - CLS

Celestica, headquartered in Canada and listed on the NYSE, designs and manufactures networking hardware for cloud companies. Its 1.6TbE switches use Broadcom Tomahawk 6 chips, and it is working with AMD on switches for the Helios rack architecture.

In Q2 2026, Celestica's Connectivity and Cloud Solutions revenue rose 84% to $3.81 billion. Hardware Platform Solutions revenue increased 58% to about $1.9 billion, while CCS segment margin was 8.7%.

Our view: Celestica is a strong volume beneficiary of open, Ethernet-based AI systems. But its lower margin shows that a meaningful share of the value still belongs to the silicon and software suppliers. Its investment case depends on maintaining design wins and gradually improving the mix of higher-value hardware platforms.

Simple linkage: More GPUs -> more communication -> more switch capacity -> more demand for ANET, AVGO and CLS.

Layer 4: AI Connectivity and Data Movement

Networking moves data across servers. Connectivity chips also solve shorter-distance traffic problems between the GPU, CPU and memory inside systems and racks.

Astera Labs - ALAB

Astera makes products that extend, retime and switch high-speed connections. In plain language, it acts like a traffic controller that keeps data flowing correctly as signals travel through increasingly complex systems.

Q2 2026 revenue reached $392.4 million, up 104% year on year, with a 73.3% GAAP gross margin and 22.7% GAAP operating margin. Growth is broadening from Aries signal-conditioning products into Scorpio fabric switches.

Our view: Astera has attractive semiconductor economics and unusually direct exposure to the complexity created by larger AI systems. The risks are customer concentration, rapid standards changes and powerful competitors such as Broadcom and Marvell.

Marvell Technology - MRVL

Marvell supplies custom silicon, electro-optical components, data-centre switches and digital signal processors. After acquiring Celestial AI and XConn, it also has more exposure to optical interconnect and advanced switching.

Fiscal Q1 2027 revenue rose 28% to a record $2.42 billion, and management said exceptional AI bookings led it to raise its fiscal 2027 and 2028 outlook. More than three-quarters of revenue came from data centres.

Our view: Marvell has broad AI connectivity exposure, but that breadth also makes execution harder to analyse. Investors should track data-centre revenue, custom-chip ramps, optical revenue and whether acquisition costs convert into stronger cash earnings.

Broadcom appears again in this section because the same Tomahawk switching silicon that powers high-performance networks also benefits from the growing amount of data moving through each cluster.

Layer 5: Optical Networking for AI Data Centers that Moves Data Using Light

Copper remains efficient over short distances, but signal loss, heat and power consumption become harder to manage as speed and distance rise. Optical systems convert an electrical signal into light, send it through fibre and convert it back.

Think of copper as a crowded city road and fibre as a high-speed railway. Both are useful, but the railway becomes more valuable as the distance and number of passengers rise.

Coherent - COHR

Coherent manufactures lasers, transceivers and optical components. NVIDIA's strategic actions show how critical this capacity has become. In March 2026, NVIDIA committed $2 billion to Coherent and entered a non-exclusive multiyear agreement containing a multibillion-dollar purchase commitment and future capacity rights.

Coherent's fiscal Q3 2026 revenue rose 21% to $1.81 billion. Data Centre and Communications revenue reached about $1.36 billion, up from roughly $969 million a year earlier, while GAAP operating margin improved to 11.1% from 4.8%.

Our view: Coherent combines real manufacturing scarcity with improving financial performance. The main risks are the capital required to expand capacity, raw-material exposure, customer concentration and the possibility of double ordering during a shortage.

Corning - GLW

Corning provides fibre, cable and connectors, the physical paths through which optical signals travel. Meta agreed to purchase up to $6 billion of Corning products for US data-centre expansion. NVIDIA separately agreed to a long-term partnership under which Corning plans to expand US optical-connectivity capacity tenfold and fibre capacity by more than 50%.

Q2 2026 Optical Communications revenue rose 32% to $2.07 billion. Enterprise Networks revenue grew 65%, with generative-AI products growing faster still. The segment's net income increased 77% to $438 million.

Our view: Corning may be the most durable “physical fibre” exposure because its glass science, manufacturing scale and customer agreements are difficult to reproduce quickly. However, it is a diversified company, so consumer-electronics and other segments can dilute the AI story.

Lumentum - LITE

Lumentum makes laser chips, optical components and switching systems. Like Coherent, it received a $2 billion NVIDIA investment plus a multibillion-dollar purchase commitment and future capacity rights in March 2026.

Fiscal Q3 2026 revenue rose 90% to $808 million. Non-GAAP operating margin expanded from 10.8% to 32.2%, showing strong operating leverage as scarce laser capacity and a better product mix flowed through earnings.

Our view: Lumentum currently shows the highest earnings torque in optics. That also makes it sensitive to any easing of shortages, customer delays or faster-than-expected capacity additions. Investors should watch margin durability, not just revenue growth.

Simple linkage: Larger clusters -> more data traffic -> more optical links -> more lasers, fibre and connectors -> more demand for COHR, GLW and LITE.

Layer 6: AI Data Center Power Infrastructure

Electricity must move through substations, transformers, switchgear, uninterruptible power supplies, busways and rack-level distribution before it reaches a GPU. Every conversion creates heat and some energy loss, so power architecture affects both reliability and operating cost.

Eaton - ETN

Eaton supplies switchgear, circuit protection, busways, UPS equipment and energy storage. Its equipment sits between the grid connection and the rack, making it a “grid-to-chip” supplier. Eaton is also working with NVIDIA on power and cooling designs for the Vera Rubin generation.

In Q2 2026, Eaton's total sales rose 21% to $8.5 billion. Electrical-sector data-centre orders increased about 85% and revenue rose about 65%. Electrical backlog increased 43% year on year, providing visibility beyond a single quarter.

Our view: Eaton offers one of the most durable business models in the chain. Its products are mission-critical, qualification cycles are long, and its exposure extends beyond data centres to utilities, industrial facilities and aerospace. That diversification reduces purity but improves resilience.

Vertiv - VRT

Vertiv supplies UPS systems, busways, power distribution, racks, chillers, coolant distribution units and services. This creates two-sided exposure: GPUs require more power, and nearly all that power becomes heat that must be removed.

Q2 2026 sales rose 24% to $3.27 billion. Adjusted operating margin increased 410 basis points to 22.6%, and adjusted free cash flow reached $925 million. Full-year guidance implied about 31% organic growth at the midpoint. Vertiv's reference architecture for NVIDIA's GB200 NVL72 supports up to 132 kW per rack, compared with the much lower rack densities common before generative AI.

Our view: Vertiv is the purest large listed exposure to the physical systems inside an AI data centre. The key risk is project timing. Large deployments can shift between quarters, while a very strong market narrative can create demanding expectations.

Simple linkage: More compute -> more electricity -> more protection and distribution equipment -> more demand for ETN and VRT.

Layer 7: AI Data Center Cooling Which Becomes Mission-Critical

An NVIDIA GB200 NVL72 rack contains 72 GPUs and can draw about 120 kW at full load. At that density, traditional air cooling becomes difficult. Direct liquid cooling brings coolant closer to the hottest components, transfers heat more efficiently and supports denser racks.

Modine - MOD

Modine's Airedale business supplies chillers, cooling units and related thermal systems. In May 2026, one strategic customer reserved more than $4 billion of cooling capacity for 2027 to 2029 and provided a $165 million upfront payment to support expansion. That commitment is larger than Modine's entire FY2026 company revenue of $3.2 billion.

Our view: Modine has powerful earnings potential because cooling is becoming a larger part of each AI campus. But the $4 billion agreement also exposes the central risk: one customer can drive a very large share of future growth. Capacity expansion must be completed on time, at the expected cost and with healthy margins.

Vertiv appears again in this section because it sells both the power equipment that feeds racks and the liquid-cooling systems that remove the resulting heat. This double exposure is a genuine differentiator.

Layer 8: AI Data Center Power Generation and Backup Power

A 100 MW campus already resembles a major industrial plant. A 1 GW campus requires dedicated thinking about generation, transmission and backup. Utility interconnection can take years, so developers are increasingly considering “bring your own power”.

Bloom Energy - BE

Bloom's solid-oxide fuel cells convert natural gas or other fuels into electricity onsite without conventional combustion. That can let a developer add power in blocks while a grid connection is delayed.

Brookfield expanded its framework for Bloom-powered AI infrastructure from $5 billion to $25 billion in June 2026. Bloom then reported Q2 revenue of $1.07 billion, up 166% year on year, and raised full-year revenue guidance to $3.9 billion to $4.2 billion.

Our view: Bloom offers the highest growth and one of the clearest solutions to grid delay. It also carries higher technology, financing and policy risk than a diversified electrical supplier. Investors should track cash conversion, service economics, backlog quality and whether fuel-cell deployments remain competitive after grid access improves.

GE Vernova - GEV

For a very large campus, the answer may be a dedicated gas power plant plus grid equipment. GE Vernova supplies gas turbines, generators, transformers and transmission technology.

Q2 2026 orders rose 88% organically to $24.2 billion, revenue increased 22% to $11.1 billion, and total backlog reached $176 billion. Gas Power equipment backlog and slot reservations rose from 100 GW to 116 GW in one quarter, while first-half data-centre orders exceeded $5 billion.

Our view: GE Vernova is one of the strongest long-duration ways to capture the power shortage because turbine and grid capacity cannot be created quickly. Its weak wind business, long project cycles and execution on a huge backlog remain important offsets.

Caterpillar - CAT and Cummins - CMI

Data centres require backup power even when their main supply is reliable. Caterpillar and Cummins manufacture large diesel and gas generators that can start quickly during an outage. Increasingly, these systems can also support prime or distributed power.

Caterpillar's Q2 2026 revenue rose 24% to a record $20.5 billion. Energy and Transportation revenue increased 17%, supported by data-centre demand. Cummins reported record Q2 revenue of $9.5 billion, while Power Systems revenue rose 19% and segment EBITDA margin expanded to 24.5%. Both results show that backup power is already material, not just a future possibility.

Our view: Cummins provides more visible earnings sensitivity through its Power Systems segment, while Caterpillar offers broader exposure to generators plus the construction equipment used to build campuses. Both are diversified industrial companies, so AI is a growth driver rather than the entire thesis.

Simple linkage: Grid delay or outage -> onsite or backup generation -> power reaches the campus -> demand for BE, GEV, CAT and CMI.

Layer 9: AI Data Center Construction Building the Grid and the Data Centre

Even after the equipment is ordered, skilled workers must build transmission lines, substations, electrical rooms, cooling loops, plumbing and miles of cable. Labour availability and project management can become bottlenecks in their own right.

Quanta Services - PWR

Quanta builds transmission, substations and utility infrastructure. A 1 GW campus may require significant reinforcement of the local grid before it can operate.

Q2 2026 revenue reached $9.6 billion, remaining performance obligations were $33.6 billion and total backlog hit a record $53.4 billion. This is broader than AI, but the same grid expansion required for electrification is now being accelerated by data-centre demand.

Our view: Quanta is a “no electrons, no AI” business. Its exposure is less direct than Vertiv's, but its skilled workforce and self-perform capabilities are difficult to replicate at scale.

Comfort Systems USA - FIX

Comfort Systems installs mechanical, electrical and plumbing systems and also builds modular systems offsite. Q2 2026 revenue rose 50% to $3.27 billion, while backlog increased 73% year on year to $14.06 billion. Technology projects were a major source of demand.

Our view: Comfort Systems is one of the strongest construction-layer operators because it combines local execution with prefabrication, which can shorten schedules and improve labour productivity. The main risks are a construction slowdown, project concentration and margin pressure if labour or materials become more expensive.

EMCOR - EME

EMCOR provides electrical and mechanical construction and ongoing facility services. Q2 2026 revenue rose 19.8% to $5.15 billion, operating margin reached 10.6%, and remaining performance obligations increased 43.9% to a record $17.14 billion.

Our view: EMCOR offers a balanced combination of data-centre construction, diversified end markets and recurring service activity. It may show less explosive growth than a pure cooling name, but its revenue base is more balanced.

The Complete AI Data-Centre Value Chain: From Chips to Construction

What does a 1 GW AI campus imply?

Here is an illustrative model, not a forecast. Assume a 1 GW facility, a power usage effectiveness or PUE of 1.2, and GB200 NVL72 racks drawing 120 kW at full load. PUE measures total facility power divided by IT equipment power.

The model explains the second-order opportunity. Roughly half a million GPUs would also require nearly 7,000 rack systems, a vast switch fabric, optical links, busways, UPS capacity, cooling for hundreds of megawatts of heat and power equipment capable of supporting a continuous industrial-scale load.

The exact mix will vary, and future chips should deliver more computation per watt. But efficiency gains may not reduce total infrastructure demand if developers use the savings to deploy more compute. This is similar to wider roads attracting more traffic: each car becomes easier to move, yet total traffic can still rise.

Which AI Data Center Stocks Have the Strongest Exposure?

Direct exposure and business quality are not the same. We use a simple Constraint-to-Cash framework with four questions:

1. Scarcity: Is the product a real bottleneck?
2. Economics: How much profit does the supplier retain?
3. Visibility: Are there firm backlogs, capacity reservations or repeat orders?
4. Risk: Could competition, customer concentration or new capacity weaken the economics?

If the priority is business quality, Arista, Vertiv and Eaton stand out. If the priority is earnings sensitivity to a physical shortage, Lumentum, Coherent, Modine and Bloom offer more torque but also greater downside if supply, timing or customer behaviour changes. If the priority is decade-long power buildout, GE Vernova and Quanta have the deepest infrastructure visibility.

This is not a prediction of short-term share-price performance. A strong company can still disappoint if its valuation already assumes flawless execution. The framework is meant to identify what must be monitored after the excitement fades.

Where Could the Biggest AI Bottlenecks Emerge?

1. Power

Power is the hardest constraint to solve quickly. Turbines, transformers, substations and transmission projects can have multiyear lead times. The main listed beneficiaries are ETN, GEV, BE, CAT, CMI and PWR.

What to monitor: data-centre power orders, transformer and turbine lead times, utility interconnection queues, backlog cancellation terms and whether onsite generation remains economical once grid connections arrive.

2. Cooling

Higher rack density pushes the industry from room-level air cooling toward direct liquid cooling. VRT and MOD have the clearest exposure.

What to monitor: liquid-cooled share of new deployments, cooling revenue per megawatt, manufacturing capacity, customer concentration and warranty performance. Cooling failures are expensive, so reliability will matter as much as price.

3. Networking

More GPUs create value only if they work as one coordinated system. ANET, AVGO and CLS benefit from faster Ethernet and larger switch fabrics.

What to monitor: AI networking revenue, adoption of Ethernet versus proprietary fabrics, switch speeds, optical content and whether customers use branded systems or internally designed hardware.

4. Optics

Faster links and longer distances increase the amount of optical content per cluster. COHR, GLW and LITE benefit, but optics can also be cyclical.

What to monitor: capacity expansion, indium-phosphide supply, pricing, customer inventory and any evidence of double ordering. A shortage can create exceptional margins, but it can also encourage customers to order more than they ultimately need.

5. Skilled construction

This is the missing bottleneck in many AI stock discussions. Equipment is useless until electricians, pipefitters, welders and project managers install it. PWR, FIX and EME convert AI plans into operating facilities.

What to monitor: backlog growth, labour productivity, prefabrication, project delays and cash flow. Construction revenue can rise rapidly while cash gets trapped in working capital, so operating cash flow is a useful quality check.

Why These Companies Matter to Investors

The first phase of the AI story asked, “Who makes the accelerators?” The next phase asks, “What must exist around every accelerator?”

That second question can have a longer shelf life. Chips are refreshed quickly, but substations, cooling plants, fibre routes and power-generation assets can operate for years or decades. Suppliers may therefore benefit from three separate waves:

1. Initial construction: Equipment, engineering and installation.
2. Technology refresh: New racks, faster networks and denser cooling.
3. Lifecycle service: Maintenance, spare parts, software and upgrades.

The most attractive models do not rely on only one wave. Vertiv combines equipment and service. Arista combines hardware with an operating system. Eaton serves data centres as well as the grid around them. GE Vernova can earn from equipment and long-term service agreements. These recurring or follow-on revenues can make the economics more durable than a one-time construction spike.

Investors should also remember the reverse logic. If hyperscaler spending slows, the impact will not be equal. Server and optical orders can move quickly. Grid projects and power plants may be slower to cancel once permitted and contracted. Diversified suppliers may feel less pain, but their AI contribution is also less pure.

Bottom Line

AI data centres are becoming industrial-scale factories that convert electricity into computation. The complete chain is:

Chips -> Servers -> Networking -> Connectivity -> Fibre -> Electrical Systems -> Cooling -> Power Generation -> Construction -> AI Services

NVIDIA provides the most visible engine, but Arista, Vertiv, Eaton, Coherent, Corning, Lumentum, Modine, Bloom Energy, GE Vernova, Quanta, Comfort Systems and many others make that engine usable at scale.

Our central view is simple: the next major AI infrastructure winners are more likely to be found by tracking constraints than by chasing labels. Power has the longest lead time. Cooling becomes more valuable as rack density rises. Networking determines whether expensive GPUs work efficiently. Optics grows as electrical links reach their physical limits. Skilled construction decides when the entire campus can finally switch on.

The best way to study these stocks is therefore not to ask which company has the loudest AI story. Ask which one controls a scarce layer, converts orders into cash at attractive margins and can defend those economics after the shortage ends.