Selling Shovels in the AI Gold Rush
Capital Allocation in the Age of AI Infrastructure
Executive Summary
Public market narratives around artificial intelligence have concentrated on a narrow set of frontier model providers and the hyperscalers deploying them. This framing has produced concentrated capital flows into a small number of names and limited investor attention on the physical layer that underwrites the entire AI economy.
This memorandum argues that the durable investment opportunities of the next decade sit not in frontier software, but in the atoms that power, move, and compute for AI at scale. As the AI economy scales, value migrates from the model layer to the infrastructure layer. Every incremental unit of AI capability depends on real-world systems that must be funded, built, permitted, connected, cooled, and operated. The capital intensity of these systems is now visible in near-term supply constraints: transformers, gas turbines, advanced lithography tools, liquid cooling units, and grid interconnection queues have all become binding.
We examine four categories of public market exposure to this buildout: (1) energy and power, including generation, grid, nuclear, and natural gas infrastructure; (2) applied automation and the execution layer, including industrial robotics and specialised physical AI; (3) physical compute and the infrastructure stack, including semiconductors, data centres, and cooling; and (4) adjacent industrials and connectivity.
We also examine where capital should not flow. As cognitive outputs commoditise, certain legacy service architectures face structural compression. These are the "sand" of the thesis: categories where AI erodes margin rather than amplifies it.
This memorandum is not prescriptive. Individual allocation decisions depend on mandate, horizon, risk profile, and liquidity. The goal is to offer a framework for distinguishing durable capital exposure from narratives that have outrun the underlying unit economics, and to identify where public market capital can underwrite the physical transition that software alone cannot fund.
Scope boundary: This memorandum examines public market capital allocation. A separate forthcoming memorandum will examine how operators and private capital can deploy AI inside existing businesses to generate operating leverage — the private market complement to this thesis.
Glossary of Terms
The Capital Migration from Bits to Atoms
The dominant AI investment narrative has centred on a vertically integrated picture of frontier model providers capturing the economic returns of artificial intelligence through software and platform distribution. This picture is incomplete.
As AI moves from research and training into continuous production and inference, the binding constraints shift from the intellectual to the physical. We argued this in prior work with respect to inference geography. [1] The same structural dynamic applies to capital allocation. Where the India memorandum considered which jurisdictions are positioned to host AI infrastructure, this memorandum considers which physical layers can be funded in public markets by investors who want exposure to the AI economy without concentrated bets on a small number of model providers.
The layer beneath frontier AI is not glamorous. It is electrical substations, gas turbines, nuclear restarts, transformer manufacturing lines, advanced lithography systems, precision cooling plants, interconnection queues, and industrial automation platforms. It is not the layer that dominates technology press coverage. It is, however, the layer that hyperscalers are now committing hundreds of billions of dollars to build. Alphabet, Amazon, Meta and Microsoft are expected to spend more than $650 billion in 2026 to expand AI capacity. [2] Of this, a material share cannot be delivered on time because the underlying physical infrastructure does not exist at the pace the industry requires. [2][3]
This is the central structural fact of the AI economy as of April 2026. Capital is abundant. Physical delivery is constrained. The gap between the two creates a multi-year window during which the owners and suppliers of scarce physical capacity earn outsized returns.
For capital allocators, this framing has three practical implications.
First, exposure to AI does not require concentrated bets on frontier model providers. The physical layer offers multiple durable exposures at more attractive entry valuations and with stronger unit economics than many of the software-layer names that dominate AI narratives.
Second, the physical layer has a longer duration. Model providers face a future of commoditising cognition, compressing pricing power at the output layer. The physical layer faces a multi-decade buildout with structurally tight supply chains. Duration of exposure matters for allocators with long horizons.
Third, the physical layer is globally distributed in ways that frontier model development is not. This has implications for portfolio diversification, geopolitical hedging, and access to capital-efficient growth outside the narrow group of dominant US software firms.
Section 2 begins with the heaviest constraint: electrical power.
Power: The Binding Constraint
In prior work we argued that power constraints shape where inference can be hosted. [1] This memorandum examines the public market capital stack that funds the physical layer itself, and the power layer is where capital most obviously meets physical reality.
The scale of AI-driven power demand is now well-documented. Global electricity consumption from data centres is projected to more than double to approximately 945 terawatt-hours by 2030, equivalent to the current total electricity consumption of Japan. [4][5] In the United States, data centres are expected to account for almost half of the growth in electricity demand between now and 2030, consuming more electricity for processing data than for manufacturing all energy-intensive goods combined, including aluminium, steel, cement and chemicals production. [4] AI-optimised data centre electricity demand alone is projected to more than quadruple by 2030. [4]
The constraint is not demand visibility. The constraint is physical delivery. As of April 2026, approximately half of all planned United States data centre builds are projected to be delayed or cancelled, not because of capital or demand shortfalls, but because the electrical grid cannot support them in the required timeframe. [2][3] Interconnection queue backlogs have grown substantially, and only a small share of projects in the queue ultimately reach completion, stretching timelines and increasing uncertainty. [6]
Three investable sub-categories emerge within the power layer.
2.1 Generation: Nuclear and Natural Gas
Nuclear power has transitioned from a declining industry into a central component of hyperscaler energy strategy. Big tech companies signed contracts for more than 10 gigawatts of potential new nuclear capacity in the United States in the year to late 2025. [7] Microsoft committed to a 20-year, 835-megawatt power purchase agreement to restart the Three Mile Island site; Google ordered up to 500 megawatts of small modular reactors from Kairos Power; Amazon invested more than $20 billion converting the Susquehanna site into a nuclear-powered AI data centre campus; Meta issued a request for proposals targeting 1 to 4 gigawatts of new nuclear generation. [7][8] The pipeline of conditional offtake agreements between data centre operators and small modular reactor projects has grown from 25 gigawatts at the end of 2024 to 45 gigawatts by late 2025. [9]
The investable exposure here spans traditional utilities with nuclear assets, specialist SMR developers, enrichment and fuel-cycle operators, and the engineering firms that will build and retrofit these facilities. SMR commercial deployment remains years away, with 2028–2030 the earliest realistic window for initial units. [10] Allocators should therefore distinguish between near-term nuclear exposure, concentrated in existing fleet operators and restart candidates, and long-duration exposure to the SMR category, which remains an option-value play with execution risk.
Natural gas has emerged as the fastest-moving near-term answer to the power gap. Microsoft is working with Chevron and Engine No. 1 to build a natural gas power plant in West Texas designed to scale to 5 gigawatts. [11] Google has confirmed a 933-megawatt gas plant project with Crusoe in North Texas. [11] Oracle has agreed to purchase up to 2.8 gigawatts of power from Bloom Energy fuel cell systems to supply AI data centres. [12] Energy Transfer is supplying gas to Oracle sites, Enbridge anticipates new data centre demand near its infrastructure, and GE Vernova has received turbine orders for multiple data centre projects. [13] According to market intelligence, approximately 30 percent of anticipated new data centre capacity will be supplied by on-site generation, up from effectively zero a year earlier. [14] Data centre power demand alone is forecast to drive a 10 to 15 percent increase in United States natural gas production over five years. [15]
The capital exposure here extends beyond the obvious midstream and upstream names. Gas turbine manufacturers face multi-year order backlogs, with turbine delivery timelines now stretching into 2028. [16] Fuel cell specialists, on-site power integrators, and engineering contractors involved in "behind-the-meter" generation are material beneficiaries.
2.2 Grid, Transmission, and Hardware
The less-obvious exposure lies in grid hardware. United States data centre expansion is constrained less by chips, servers or funding than by the electrical hardware needed to connect new facilities to reliable power. [17] According to Wood Mackenzie analysis, the United States faced a supply deficit of approximately 30 percent for power transformers and 10 percent for distribution transformers in 2025. [18] Imports of high-power transformers from China surged from fewer than 1,500 units in 2022 to more than 8,000 in 2025. [17][18]
This is not a cyclical shortage. The United States power grid was designed for a load profile that no longer applies. Approximately 70 percent of the existing grid is approaching the end of its operational life. [19] Data centres can be built in under three years, but new gas generation takes around six years, solar or wind three to six, and new nuclear more than ten. [20] Transmission line permitting and transformer manufacturing operate on lead times measured in years in a sector where deployment cycles now run under 18 months.
Exposure to this category includes listed power equipment manufacturers (transformers, switchgear, grid hardware), transmission utilities with growing rate bases, and specialised engineering and manufacturing firms in the grid modernisation supply chain. This is, in many cases, a capital-intensive and slower-moving set of businesses than the software-layer alternatives. That profile is a feature, not a defect. These are the businesses whose long-duration cash flows have re-rated as the physical constraint has become undeniable.
2.3 Geographic Arbitrage: The Gulf
Where United States and European markets face grid constraints, Gulf jurisdictions have moved quickly to deploy capital and compress time-to-power. Saudi Arabia's Humain, launched in May 2025 and owned by the Public Investment Fund, has broken ground on 100-megawatt data centres in Riyadh and Dammam with initial launch planned for 2026, and aspires to deliver 6.6 gigawatts of capacity within a decade. [21][22] Humain and Qualcomm have announced a partnership targeting 200 megawatts of AI compute deployment from 2026. [23] The NEOM megaproject has been redesignated around data centre capacity, with Red Sea seawater cooling and integration with DataVolt's AI factory campus at Oxagon, Humain's involvement from November 2025, and approximately $40 billion committed across Saudi AI investment vehicles. [24][25][26]
The capital implication is not that Gulf public markets offer the most liquid exposure to this buildout. It is that Gulf demand is materially expanding the addressable market for equipment manufacturers, specialist engineering firms, and global infrastructure platforms. The Gulf is a source of capex rather than a destination for portfolio allocation, and exposure is most cleanly captured through the global suppliers that serve it.
Physical Compute and the Infrastructure Stack
Beneath the power layer sits the physical compute stack: semiconductor manufacturing and the specialised equipment that enables it, data centre real estate, thermal management systems, and the specialty materials that underpin advanced node production.
3.1 Semiconductors and Wafer Fab Equipment
The semiconductor layer is the most widely covered element of the AI buildout and, partly for that reason, the most expensive at the point of entry. It remains a meaningful part of the picks-and-shovels thesis nonetheless, because the multi-year order book for advanced equipment is now visible.
Taiwan Semiconductor Manufacturing Company raised its 2026 capital expenditure guidance to $52 billion to $56 billion, substantially above prior market expectations and materially above the $40.5 billion spent in 2025. [27] The sole supplier of extreme ultraviolet lithography systems, ASML, raised its 2026 revenue guidance to €36 to 40 billion and reported net sales of €8.8 billion in Q1 2026 against market estimates of €8.5 billion. [28][29] Wafer fab equipment revenue reached approximately $104 billion in 2024, is forecast to rise to $115.7 billion in 2025, and is projected to reach $135.2 billion by 2027. [30] ASML has projected $71 billion in revenue by 2030 on the basis of AI-driven demand for EUV lithography. [30]
Capital exposure here runs from the foundries themselves to equipment manufacturers, memory specialists, packaging and test providers, specialty materials firms, and the ecosystem of component and sub-system suppliers. Unlike the model providers, these businesses have multi-year contracted order books and capital-intensive moats that are difficult to disintermediate.
3.2 Data Centre Real Estate
The data centre real estate category sits at the intersection of real assets and technology capex. Construction spending on data centres in the United States reached $41.1 billion in Q2 2025, marking the tenth consecutive quarter of growth and the highest increase on record. [31] The three largest US-listed data centre real estate investment trusts (Digital Realty, Equinix, and Iron Mountain) collectively expanded their active development pipelines to approximately $9.8 billion, up from $1.5 billion in 2021, representing a compound annual growth rate of approximately 65 percent. [31] Equinix has announced plans to roughly double its capacity by the end of 2029, with 52 development projects in progress across major markets. [32] Digital Realty has 769 megawatts under construction in major metro areas, with announced Americas future capacity of 499 megawatts at 79 percent pre-leased. [32]
Despite the growth, the major REITs have traded with mixed performance over the past year, reflecting market scepticism about where value capture concentrates in the AI stack. [33] This creates an asymmetry for allocators. The demand visibility for data centre capacity is among the strongest in public markets; the pricing reflects doubt about whether landlords or chipmakers capture the economic profits. Allocators with longer horizons and appetite for real asset exposure can use this asymmetry.
Construction activity is concentrated in land- and power-constrained major markets, where the two largest REITs have long track records and deep utility relationships. [32] This concentration is an advantage for existing operators against new entrants. New development in tier-one markets increasingly depends on pre-secured power allocations and interconnection rights that take years to assemble.
3.3 Thermal Management and Cooling
As AI workloads push rack densities from traditional 10-15 kilowatts into 30-50 kilowatts and beyond, traditional air cooling becomes inadequate. Liquid cooling, once a niche approach, is now required for new high-density AI deployments. The global data centre liquid cooling market is forecast to grow from approximately $6.65 billion in 2025 to approximately $8.17 billion in 2026, and to approach $29.46 billion by 2033 at a compound annual growth rate of approximately 20 percent. [34] Specialist direct-to-chip and immersion cooling segments are growing even faster, with immersion cooling projected to grow from $668.6 million in 2026 to approximately $3.4 billion by 2033. [35]
The dominant listed beneficiary in this category is Vertiv Holdings, which reported a backlog of approximately $15 billion and orders growing substantially in early 2026. [36] Vertiv's strategic partnership with NVIDIA on reference architectures for high-density rack systems positions it as a preferred supplier to the hyperscaler buildout. [37] Other beneficiaries span thermal management specialists, cooling component manufacturers, and the broader industrial firms adapting their product lines to AI data centre requirements.
The broader point is that thermal management has graduated from a commodity service to a strategic bottleneck. This is the pattern to watch across the physical compute stack: as AI intensity increases, previously unremarkable components become binding constraints and their suppliers earn pricing power.
Applied Automation and the Execution Layer
Software artificial intelligence has advanced faster than the physical world can absorb. The bridge between them is applied automation, the layer where AI decisions translate into physical action in the real economy.
This is the layer we described in prior work as the domain of durable scarcity in the services economy: physical execution under regulated, accountable conditions where cognition alone cannot deliver. [41] Where that prior memorandum examined services business models, this section considers the public market exposures to the underlying automation hardware and systems that enable AI to operate in physical environments.
4.1 Industrial Automation
Industrial automation is a less crowded and arguably more investable category than humanoid robotics, and it is further along in its deployment cycle. The global industrial robotics market is forecast to grow from approximately $21.94 billion in 2025 to approximately $77.36 billion by 2034, a compound annual growth rate of approximately 15.5 percent. [38] The broader industrial automation market is forecast to reach approximately $149.3 billion by 2030, growing at a compound annual growth rate of approximately 7.2 percent. [39] The global robotics market as a whole is forecast to reach approximately $205 billion by 2030, with industrial robots, mobile robots, and collaborative robots accounting for the majority of deployed units and revenue. [40]
The investable exposure spans established industrial robotics incumbents, collaborative robot specialists, warehouse automation platforms, and the systems integrators that deploy them into end markets. Demographic pressure (aging workforces in developed economies and tightening labour markets in manufacturing) provides a durable demand backdrop that is independent of AI narrative cycles. AI is the coordination layer that makes these systems materially more capable, but the underlying capex was already accelerating.
4.2 Specialised-Form Robotics and the Humanoid Question
Within applied automation, the humanoid robotics category has attracted the most venture and media attention. Tesla has announced Optimus production targets of 5,000 units with potential scaling to 12,000 in 2025, and price targets around $20,000 per unit at scale. [42][43] BYD aims to deploy 1,500 humanoid robots in 2025, scaling to 20,000 by 2026. [42] The broader humanoid robot market is forecast to grow from approximately $1.55 billion in 2024 to approximately $4.04 billion by 2030 at a compound annual growth rate of approximately 17.5 percent. [44]
The humanoid wave reflects a specific thesis about form factor: that the built environment is designed around the human body, and that a robot with roughly human form is the cheapest path to AI labour without retrofitting the world. This is a legitimate argument and it underwrites the capital flowing into the category.
The thesis is not, however, universally applicable. In environments where the human design constraint does not hold (subsurface infrastructure, offshore assets, confined industrial spaces, extreme temperature or pressure environments, precision agricultural settings, and large-format warehouses where the environment can be redesigned around the robot) specialised-form factors are often superior. Human form is fragile, energy-inefficient relative to purpose-built machines, and limited in speed and payload. Public market exposure to humanoid robotics today is concentrated in a small number of listed equity names and the supply chain feeding into them. The specialised-form category is less concentrated, less narratively compelling, and in several sub-segments substantially more advanced in commercial deployment.
Compress
The humanoid-form bet is a bet on the cost of retrofitting the world to robots.
Expand
The specialised-form bet is a bet on the cost of building purpose-fit robots for environments where the human constraint does not apply.
Both are investable categories. Public markets currently overweight the humanoid bet relative to the specialised category's addressable scope.
For capital allocators, the implication is not to avoid humanoid exposure. It is to recognise that the humanoid category is one form factor within a broader applied automation opportunity, and that the narrative concentration in this sub-segment obscures more durable exposures elsewhere in the stack.
4.3 Middle East and the Physical AI Corridor
Gulf jurisdictions are extending their infrastructure buildout beyond data centres into physical AI deployments. Saudi Arabia has structured its AI strategy around what functions as a full-stack deployment model: AI infrastructure, generation, cooling, and increasingly the applied automation layer, including NEOM's AI operating layer for urban management and robotics manufacturing partnerships under the Vision 2030 framework. [26][45] This creates demand signals for global automation suppliers and opens one of the few greenfield markets where physical AI deployments can be built without the constraints of legacy infrastructure.
Adjacent Industrials and Connectivity
Beyond the obvious shovels sit the adjacent industrials whose products and services the entire buildout depends on. These are less glamorous categories, often overlooked by technology-sector capital allocators, and among the cleanest expressions of the picks-and-shovels thesis.
The adjacent industrial layer includes the mechanical, electrical, and plumbing (MEP) engineering contractors that build and fit out data centres; the specialty materials and gases firms that supply semiconductor manufacturing; the connectivity and networking hardware companies supplying optical and interconnect solutions to hyperscale campuses; the water infrastructure firms supplying and treating the substantial volumes of water required for cooling; and the specialty chemicals producers supplying photoresists, precursors, and high-purity materials.
Several features make this category attractive from a capital allocation standpoint.
First, these are often businesses with long-standing operational moats, established supply relationships, and high switching costs. The barriers to competitive entry are meaningful, particularly for regulated or safety-critical sub-segments.
Second, these businesses tend to trade at more modest valuation multiples than their technology-sector equivalents, often reflecting their industrial classification rather than their exposure to AI-driven demand growth.
Third, these businesses are globally distributed in ways that the frontier AI software layer is not. Specialty materials and equipment suppliers concentrated in Japan, Germany, the Netherlands, and Korea offer allocators non-US exposure to the AI buildout through operationally established firms with decades of track record.
Fourth, the demand signal is durable and extends beyond AI-specific workloads. The same firms serving the AI buildout also serve electric vehicle manufacturing, reshoring initiatives, and the broader industrial electrification cycle. This diversification reduces exposure to any single AI narrative cycle.
The practical implication for capital allocation is not that adjacent industrials should displace the more obvious shovel categories. It is that allocators constructing balanced exposure to the AI buildout should not omit them. The adjacent industrial layer often offers superior risk-adjusted returns over multi-year horizons because its participation in the buildout is less narratively priced in.
Sand, Not Shovels
Not every business adjacent to AI benefits from the buildout. Some face structural compression as AI commoditises the cognitive outputs that underwrite their economics. These are the "sand" of the thesis: categories where capital is best underweighted rather than overweighted.
Four categories warrant attention.
6.1 Pure Cognitive Wrappers
Businesses whose core value proposition is a thin layer of user interface over a general-purpose foundation model face sustained margin compression. As model providers continue to improve their own application layers, and as open-weight and alternative models proliferate, the pricing power of commodity cognitive tools will erode. This includes many consumer AI chatbot apps, generic copilot products, and commodity AI-generated content platforms. Allocators should assume that defensibility in this category requires proprietary data, workflow integration, distribution, or regulatory moats that most pure wrapper businesses do not possess.
6.2 Legacy Business Process Outsourcing
Traditional business process outsourcing (BPO), particularly in call centre and low-end back-office segments, faces sustained structural pressure. The category is not collapsing; established operators such as Teleperformance and Concentrix have significantly invested in AI integration and hybrid human-plus-automation delivery models. [46][47] But the economic model of headcount-based revenue is under clear pressure as automation compresses the cognitive cost of routine interactions. Concentrix's 2023 $5 billion merger with Webhelp created a combined workforce of more than 400,000 agents, which frames both the scale of the category and the magnitude of potential compression. [47]
Allocators should distinguish between BPO operators that are credibly rebuilding their delivery model around AI-integrated workflows and those that are layering AI tools over unchanged processes. The first category may emerge as durable operators. The second faces sustained margin erosion.
6.3 Commodity Content and Stock Media
Commodity content producers, including stock photography, stock video, generic editorial content, basic translation services, and template-driven creative services, face direct substitution pressure from generative AI. The addressable market for these services does not disappear; it contracts as the unit cost of AI-generated alternatives approaches zero. Defensibility in this category requires proprietary content libraries, rights ownership, brand, or specialised domain expertise that most commodity providers lack.
6.4 Low-End Professional Services
Junior-tier professional services, particularly those built around standardised output production such as basic legal research, commodity accounting, routine regulatory filings, and formulaic marketing content, face pressure at the task level even if firm-level revenues remain resilient. The firms likely to compress are those that scaled on labour arbitrage rather than on defensible judgment or regulated accountability. This was the central argument of our prior memorandum on the services economy. [41] The public market implication is less direct, because most affected firms are privately held, but listed professional services firms whose revenue is concentrated in commoditising task types warrant scrutiny.
The common thread across these categories is a dependence on cognitive output as the primary economic unit. Where cognition was scarce, these businesses scaled. As cognition commoditises, they compress. The question for allocators is not whether these categories disappear, but whether current valuations adequately reflect a sustained reset in unit economics.
Three Horizons of Shovel Exposure
Exposure to the AI buildout operates on three distinct horizons, and the relative weight of each horizon depends on allocator mandate.
7.1 Tactical (12 to 24 months)
The near-term horizon is dominated by earnings-cycle dynamics, catalyst-driven sentiment, and supply chain visibility. The binding physical constraints (transformers, gas turbines, advanced lithography, high-density cooling) create supply-constrained pricing environments that flow directly into near-term earnings for the relevant suppliers. Allocators with tactical mandates can concentrate positioning around companies whose order books are visibly expanding and whose delivery capacity is structurally constrained.
Near-term risks include macro-cyclical demand sensitivity in specific sub-segments, permitting and regulatory timing, and idiosyncratic execution risk in capital-intensive buildouts.
7.2 Strategic (3 to 5 years)
The strategic horizon corresponds to the bulk of the current capex cycle and the realistic deployment window for most of the infrastructure being contracted in 2025 and 2026. Over this window, the economic returns of the buildout accrue to the operators and suppliers that successfully deliver at scale. Strategic exposure favours businesses with durable competitive positions, strong balance sheets, and demonstrated capacity to ramp delivery. For most infrastructure categories this favours established incumbents with existing manufacturing footprints over emerging specialists dependent on single-customer relationships.
7.3 Macro (5 to 10 years and beyond)
The macro horizon captures the longer arc of the AI economy's physical build. Over this horizon the central question is which categories become the permanent floor of the global economy (analogous to how electrification in the early twentieth century became an irreducible economic layer).
Our view is that the power layer is the most clearly structural. Electricity demand growth, once reset upward by AI, is unlikely to revert to the flatlining profile that characterised advanced economies for two decades. [19] Grid modernisation, transmission buildout, nuclear restart, and the long-run supply chains for electrical hardware are durable at macro horizon.
Applied automation is also structural but with more dispersion across sub-segments. Industrial automation appears durable at macro horizon. Humanoid-form robotics remains a more uncertain category at this horizon, because its long-run market size is sensitive to form-factor unit economics that have not yet been demonstrated at scale.
Allocators should layer exposure across horizons rather than choosing one. Tactical positioning captures the supply constraint premium; strategic positioning captures the capex cycle; macro positioning captures the structural reset.
Capital Implications and the Bridge to Value Creation
This memorandum has argued that public market capital deployed into the AI economy is best allocated along four shovel categories (energy and power; applied automation; physical compute; and adjacent industrials and connectivity) and underweighted across four sand categories (pure cognitive wrappers; legacy BPO; commodity content; and low-end professional services).
Three implications follow.
First, exposure to the AI economy is broader and more investable than the dominant narrative implies. Allocators concentrated in a small number of model providers and hyperscalers are likely underexposed to the physical layer that is capturing the majority of near-term capital deployment.
Second, the picks-and-shovels thesis is not a single trade. It is a layered framework for exposure across different time horizons, risk profiles, and sub-segments. Different allocator mandates will emphasise different layers. An infrastructure-focused mandate will concentrate exposure in power and data centre assets; a growth mandate will favour applied automation and equipment manufacturers; a value mandate will find depth in adjacent industrials.
Third, the sand categories warrant more scrutiny than they have received. Portfolios constructed over the past decade on the assumption that cognitive-output business models are durable may require re-underwriting as the cost of cognition continues to compress.
The Private Market Complement
Public markets give allocators exposure to the shovels. They do not directly monetise the value that AI creates inside existing operating businesses. That value accrues primarily in private markets, where operators and private capital deploy AI-enabled tooling and workflow redesign to generate operating leverage in acquired or held companies.
The private market thesis is structurally distinct from the public market thesis. It is not about who supplies the shovels, but about who uses them. It is less about who sells AI and more about who absorbs it. A subsequent memorandum will examine how operators and private capital can deploy AI inside existing businesses to generate operating leverage: which business types are most receptive, what the playbook looks like in practice, how value creation compounds in fragmented verticals, and where the returns concentrate. This framework complements the public market picks-and-shovels thesis developed here.
Public markets fund the physical transition. Private markets determine which operators capture the operating leverage that AI unlocks in existing businesses. Both are necessary for a complete view of capital allocation in the AI age.
Conclusion
Artificial intelligence is reshaping global capital allocation in ways that are not fully reflected in current market positioning. The dominant narrative has concentrated attention on a small number of frontier model providers and hyperscalers. The physical layer that underwrites the entire AI economy has been comparatively under-followed, even as supply constraints across power, semiconductors, cooling, and industrial automation have become increasingly binding.
For capital allocators, this mispricing is an opportunity. Exposure to AI does not require concentrated bets on model providers. It can be constructed through layered exposure to the physical infrastructure that the AI buildout depends on, with more attractive entry valuations and in many cases stronger unit economics than the software layer offers.
The shovel categories (energy and power, applied automation, physical compute, and adjacent industrials) are broadly durable at macro horizon. The sand categories warrant re-underwriting as cognitive outputs continue to commoditise.
This memorandum offers a framework, not a portfolio. Allocation decisions remain the domain of investors operating within their own mandates, risk profiles, and time horizons. The objective is to contribute to structured dialogue around capital allocation in the AI age, and to identify where public market capital can underwrite the physical transition that software alone cannot fund.
References
- [1] Veeran Advisory (2026) India as a Global Inference Hub. https://veeranadvisory.com/research/india-global-inference-hub
- [2] European Business Magazine (2026) Data Centre Power Crisis Is Choking the AI Revolution. April 2026. https://europeanbusinessmagazine.com/business/technology-data-centre-power-crisis-ai-growth-2026/
- [3] Tom's Hardware (2026) Half of planned US data center builds have been delayed or canceled. https://www.tomshardware.com/tech-industry/artificial-intelligence/half-of-planned-us-data-center-builds-have-been-delayed-or-canceled-growth-limited-by-shortages-of-power-infrastructure-and-parts-from-china-the-ai-build-out-flips-the-breakers
- [4] International Energy Agency (2025) Energy and AI: Special Report. https://www.iea.org/reports/energy-and-ai
- [5] S&P Global (2025) Global data center power demand to double by 2030 on AI surge: IEA. https://www.spglobal.com/energy/en/news-research/latest-news/electric-power/041025-global-data-center-power-demand-to-double-by-2030-on-ai-surge-iea
- [6] World Resources Institute (2025) Powering the US Data Center Boom: The Challenge of Forecasting Electricity Needs. https://www.wri.org/insights/us-data-centers-electricity-demand
- [7] Introl (2026) Nuclear power for AI: inside the data center energy deals. https://introl.com/blog/nuclear-power-ai-data-centers-microsoft-google-amazon-2025
- [8] Trellis (2025) Amazon, Google, Meta and Microsoft go nuclear. https://trellis.net/article/amazon-google-meta-and-microsoft-go-nuclear/
- [9] International Energy Agency (2026) Data centre electricity use surged in 2025, even with tightening bottlenecks driving a scramble for solutions. https://www.iea.org/news/data-centre-electricity-use-surged-in-2025-even-with-tightening-bottlenecks-driving-a-scramble-for-solutions
- [10] World Wide Technology (2025) Big Tech's Nuclear Bet: Key Small Modular Reactors for Cloud Power. https://www.wwt.com/blog/big-techs-nuclear-bet-key-small-modular-reactors-for-cloud-power
- [11] TechCrunch (2026) AI companies are building huge natural gas plants to power data centers. https://techcrunch.com/2026/04/03/ai-companies-are-building-huge-natural-gas-plants-to-power-data-centers-what-could-go-wrong/
- [12] AI Power Weekly (2026) AI Power News 4/20/2026. https://www.aipowerweekly.com/p/ai-power-news-4202026
- [13] IndexBox (2026) AI Data Center Boom Fuels Energy Sector Growth & Investment in 2026. https://www.indexbox.io/blog/energy-stocks-surge-as-ai-data-center-demand-strains-power-grids/
- [14] Tech Insider (2026) The AI Data Center Power Crisis. https://tech-insider.org/ai-data-center-power-crisis-2026/
- [15] American Oil & Gas Reporter (2025) Powering AI Will Require Rapid Increases In Natural Gas Production. https://www.aogr.com/web-exclusives/exclusive-story/powering-data-centers-will-require-rapid-increases-in-natural-gas-production
- [16] IndexBox (2026) AI Data Centers Drive Tech Giants' Natural Gas Power Investments in 2026. https://www.indexbox.io/blog/tech-giants-expand-natural-gas-power-for-ai-data-centers-amid-equipment-shortages/
- [17] Digital Watch Observatory (2026) Power hardware shortages are delaying AI data centre expansion, despite record investment. https://dig.watch/updates/power-hardware-shortages-are-delaying-ai-data-centre-expansion-despite-record-investment
- [18] PandaYoo (2026) Why Chinese Transformer Makers Are Benefiting From the Global Grid and AI Power Crunch. https://pandayoo.com/post/why-chinese-transformer-makers-are-benefiting-from-the-global-grid-and-ai-power-crunch-en/
- [19] Data Center Knowledge (2026) 2026 Predictions: AI Sparks Data Center Power Revolution. https://www.datacenterknowledge.com/operations-and-management/2026-predictions-ai-sparks-data-center-power-revolution
- [20] Transformer Magazine (2026) Power bottleneck slows AI data centres. https://transformers-magazine.com/tm-news/power-bottleneck-slows-ai-data-centres/
- [21] Al Arabiya (2025) Saudi Arabia's Humain to open data centers with US chips in early 2026. https://english.alarabiya.net/News/saudi-arabia/2025/08/26/saudi-arabia-s-humain-to-open-data-centers-with-us-chips-in-early-2026
- [22] CNBC (2025) Saudi AI firm Humain is pouring billions into data centers. https://www.cnbc.com/2025/08/27/saudi-arabia-wants-to-be-worlds-third-largest-ai-provider-humain.html
- [23] Computer Weekly (2025) Saudi Arabia partners with Humain and Qualcomm to power next global AI frontier. https://www.computerweekly.com/news/366633624/Saudi-Arabia-partners-with-Humain-and-Qualcomm-to-power-next-global-AI-frontier
- [24] Data Center Dynamics (2026) Saudi Arabia's Neom megaproject could be redesignated as a data center hub. https://www.datacenterdynamics.com/en/news/saudi-arabias-neom-megaproject-could-be-redesignated-as-a-data-center-hub/
- [25] Abhishek Gautam (2026) The Line Is Cancelled: Saudi Arabia's $50B NEOM Project Becomes an AI Data Centre. https://www.abhs.in/blog/saudi-arabia-the-line-neom-cancelled-ai-data-center-pivot
- [26] Middle East Insider (2026) Saudi Arabia AI Strategy 2026: From Data Centers to Global Tech Power. https://themiddleeastinsider.com/2026/03/29/saudi-arabia-ai-strategy-2026-data-centers-tech-investment/
- [27] CNBC (2026) TSMC and ASML post-earnings stock moves could be a sign of what's to come from chip companies. https://www.cnbc.com/2026/04/16/taiwan-semi-tsm-asml-stock-earnings-ai-chips.html
- [28] Digital Today (2026) ASML raises annual revenue forecast amid chip market recovery. https://www.digitaltoday.co.kr/en/view/48434/asml-raises-annual-revenue-forecast-amid-chip-market-recovery-expects-capex-from-tsmc-samsung-sk-hynix
- [29] Semiconductor Things (2026) The Semiconductor Boom is Real and Prolonged. https://www.semiconductorthings.com/p/the-semiconductor-boom-is-real-and-profound-tsmc-2026-taiwan
- [30] Tom's Hardware (2025) Sales of chip production equipment to reach $156 billion by 2027. https://www.tomshardware.com/tech-industry/semiconductors/sales-of-chip-production-equipment-to-reach-usd156-billion-by-2027-china-taiwan-and-korea-lead-intense-demand
- [31] Capright (2025) Data Center Market Update - October 2025. https://www.capright.com/data-center-market-update-october-2025/
- [32] S&P Global (2025) Digital Realty, Equinix ramp up datacenters as AI drives demand. https://www.spglobal.com/market-intelligence/en/news-insights/articles/2025/6/digital-realty-equinix-ramp-up-datacenters-as-ai-drives-demand-90542889
- [33] Bisnow (2026) Built For This Moment: Data Center REITs Equinix, Digital Realty See AI Boom Hitting Inflection Point. https://www.bisnow.com/national/news/data-center-capital-markets/built-for-this-moment-data-center-reits-equinix-digital-realty-hitting-ai-stride-133220
- [34] Grand View Research (2025) Data Center Liquid Cooling Market | Industry Report, 2033. https://www.grandviewresearch.com/industry-analysis/data-center-liquid-cooling-market-report
- [35] Coherent Market Insights (2026) Data Center Liquid Immersion Cooling Market Forecast, 2033. https://www.coherentmarketinsights.com/market-insight/data-center-liquid-immersion-cooling-market-1767
- [36] Seeking Alpha (2026) Vertiv Stock: The $15 Billion Backlog, Liquid Cooling Dominance, And The AI Trade. https://seekingalpha.com/article/4890719-vertiv-holdings-the-15-billion-backlog-liquid-cooling-dominance-and-the-ai-infrastructure-trade-wall-street-is-still-underpricing
- [37] EnkiAI (2026) Vertiv Liquid Cooling: Dominating AI Data Centers in 2025. https://enkiai.com/data-center/vertiv-liquid-cooling-dominating-ai-data-centers-in-2025/
- [38] Fortune Business Insights (2026) Industrial Robots Market Size, Share | Industry Report, 2034. https://www.fortunebusinessinsights.com/industry-reports/industrial-robots-market-100360
- [39] IndustryARC (2024) Industrial Automation Market Size, Share | Industry Trend & Forecast 2030. https://www.industryarc.com/Research/Industrial-Automation-Market-800985
- [40] ABI Research (2025) The Global Robotics Market Outlook. https://www.abiresearch.com/blog/global-robotics-market-outlook
- [41] Veeran Advisory (2026) The AI-Augmented Services Economy. https://veeranadvisory.com/research/ai-augmented-services-economy
- [42] IDTechEx (2025) Humanoid Robots 2025-2035: Technologies, Markets and Opportunities. https://www.idtechex.com/en/research-report/humanoid-robots/1093
- [43] Humanoid Robotics Technology (2025) Tesla Unveils Ambitious Optimus Humanoid Roadmap. https://humanoidroboticstechnology.com/industry-news/tesla-unveils-ambitious-optimus-humanoid-roadmap/
- [44] Grand View Research (2025) Humanoid Robot Market Size & Share | Industry Report, 2030. https://www.grandviewresearch.com/industry-analysis/humanoid-robot-market-report
- [45] Crowell & Moring (2025) The Middle East's Big Bet on Artificial Intelligence and Data Security. https://www.crowell.com/en/insights/client-alerts/the-middle-easts-big-bet-on-artificial-intelligence-and-data-security
- [46] 1840 & Co (2025) 8 Top Concentrix Competitors & Global Alternatives in 2026. https://www.1840andco.com/blog/concentrix-competitors
- [47] ShyftOff (2026) Best Call Center Outsourcing Companies 2026: Ultimate Guide. https://www.shyftoff.com/blog/top-call-center-outsourcing-companies