Agentic AI Is Moving Beyond Software and Into the Physical World. Industrial Companies Are the Quiet Winners
The AI investment narrative has been dominated by a relatively narrow cast of technology companies — Nvidia, TSMC, Samsung, SK Hynix, the hyperscalers — whose products are directly identifiable in the AI compute stack. Globe and Mail market analysis from July 7, 2026 captures an important but underappreciated dimension of the AI infrastructure investment wave: "Massive AI data centre growth is benefiting several nuclear power generator and reactor makers, construction giants, cooling and water purifying companies and industrial manufacturers." The AI trade, the analysis notes, "is expanding from chips to memory and storage devices as well as servers and racks. Moreover, agentic AI is expanding the scope of AI infrastructure providers in the physical layer." These are not peripheral beneficiaries of the AI cycle — they are structural enablers whose capacity constraints are as real as the chip shortages that make headlines.
Agentic AI — AI systems that can take sequences of actions autonomously to complete complex tasks without step-by-step human direction — requires substantially more inference compute per user interaction than conventional AI query-response models. An agentic AI system that browses the web, writes and executes code, manages files, sends emails, and coordinates across multiple tools in response to a single user request generates 5 to 20 times the compute demand of a conventional chatbot query. As enterprise adoption of agentic AI accelerates — with over half of major pharmaceutical companies classified as "heavy AI users" as of mid-2026, and enterprise AI adoption rates growing across financial services, logistics, and manufacturing — the inference compute infrastructure required to serve those agentic workloads is multiplying beyond what the AI training compute buildout alone would have required. The physical infrastructure that supports this expanding compute demand — power, cooling, construction, and water — is where the industrial sector's AI exposure lives.
Nuclear Power: From Nuclear Anxiety to Nuclear Necessity
The hyperscalers' commitment to nuclear power purchase agreements for AI data centre energy supply represents one of the most significant reversals in energy market positioning in recent memory. Microsoft's agreement with Constellation Energy to restart Three Mile Island Unit 1 as the Crane Clean Energy Centre, Amazon's $650 million acquisition of a nuclear-powered data centre campus from Talen Energy, and Google's agreement with Kairos Power for small modular reactor deployment at its data centre sites all signal that the world's largest technology companies have concluded that nuclear is the only power source that can deliver the combination of continuous baseload availability, zero carbon, and physical co-location with data centres that AI compute operations require at the scale their infrastructure programmes demand.
The small modular reactor market is the specific industrial technology category receiving the most commercially actionable investment from the nuclear-for-AI thesis. SMRs — reactor designs with electrical output below 300 megawatts — can be manufactured in standardised factory components and assembled on-site in significantly shorter timelines than conventional large-scale nuclear plants, making them feasible for deployment at or near specific data centre campuses rather than requiring data centres to be located near existing grid-connected nuclear facilities. NuScale, TerraPower, X-energy, and Oklo are the leading U.S. SMR developers with active licensing processes at the Nuclear Regulatory Commission, and each has received documented commercial interest from at least one major technology company. The SMR market's transition from development-stage technology to contracted commercial deployment is happening substantially faster than the regulatory timelines that critics of nuclear power have historically cited as prohibitive for the technology's relevance to near-term clean energy deployment.
Cooling and Water: The Hidden Constraint in AI Infrastructure
Data centre cooling represents the second major physical infrastructure constraint on AI compute scaling that industrial sector investors are tracking. A modern AI training cluster running at full utilisation generates heat at densities — kilowatts per rack — that conventional data centre cooling infrastructure was not designed to manage. Traditional air cooling, which uses chilled air circulation to remove heat from server racks, becomes thermally inadequate and energy-inefficient at the rack densities that Nvidia's H100 and H200 GPU clusters require. Liquid cooling — where coolant fluid is circulated directly through server hardware in contact with heat-generating components — is the engineering solution for high-density AI compute, but it requires fundamentally different data centre mechanical infrastructure than air cooling and creates water consumption volumes that raise site selection constraints in water-stressed geographies.
Vertiv Holdings, Eaton, and Schneider Electric — the industrial companies that supply the power distribution, uninterruptible power supply, and thermal management infrastructure that data centres require — are reporting order backlogs that reflect hyperscaler data centre construction programmes running at capacity. Vertiv's share price has appreciated substantially through 2026 as its data centre power and cooling infrastructure becomes a recognised AI infrastructure investment rather than a conventional industrial equipment stock. The company's order backlog metrics provide real-time visibility into hyperscaler data centre construction commitments that is complementary to the chip and memory supply data that dominate AI infrastructure market analysis — and in many respects more reliable as a forward indicator because cooling and power infrastructure is ordered before chip installation rather than simultaneously with it.
Construction Giants and the Data Centre Building Boom
Data centre construction is one of the fastest-growing segments in commercial construction globally, absorbing engineering, materials, and skilled labour at rates that are generating above-market revenue growth for the construction companies with data centre specialisation. Turner Construction, Hensel Phelps, DPR Construction, and Mortenson are executing AI data centre campuses of 200 to 500 megawatts — among the largest civil construction projects currently underway in the United States. The construction demand from AI data centre programmes is competing for the same skilled electricians, mechanical engineers, and construction project managers as the CHIPS Act-funded semiconductor fabrication plant construction — creating labour market pressure that is extending construction timelines and elevating project costs beyond what hyperscaler initial budget models anticipated.
SoftBank's Roze initiative — using robotics to accelerate data centre construction — is a direct response to this physical construction bottleneck. If robotic construction systems can reduce the labour intensity and sequential dependencies of data centre construction, they address the binding constraint on AI infrastructure deployment pace in ways that additional capital investment alone cannot. Boston Dynamics, Built Robotics, and Dusty Robotics are positioned at the intersection of two capital-intensive waves: the AI infrastructure buildout and the manufacturing near-shoring programme.
What This Means for Market Participants
Industrial sector investors and market research professionals should treat AI data centre physical infrastructure as an investment category with specific, identifiable, and currently undervalued components relative to the semiconductor and software companies that receive the majority of AI investment attention. Nuclear power developers with data centre offtake agreements, cooling infrastructure specialists with data centre-specific products, construction companies with AI campus experience, and water treatment technology providers serving data centre operations all have direct, contractually supported exposure to the $750 billion annual hyperscaler AI capex programme. The industrial physical layer of AI infrastructure is where the AI investment cycle's most durable earnings growth opportunities may lie for investors who have found the semiconductor and software AI investment categories already fully valued relative to their earnings delivery timelines.