VEERAN ADVISORY
← All research

№ 02 · Apr 2026 · ALLOCATORS · OPERATORS · OBSERVERS

Selling Shovels in the AI Gold Rush

As AI collapses the cost of cognition, durable value migrates to the physical: the infrastructure that powers intelligence, and the execution businesses that deliver it. Where capital should flow, and where it should not.

Executive Summary

Public market narratives around artificial intelligence concentrate on a narrow set of model providers and the hyperscalers deploying them. This memorandum argues the durable returns sit elsewhere: in the atoms beneath the models, and in the operating businesses that absorb them.

The argument runs on two floors. On the first floor, as AI scales from research into continuous production, value migrates from the model layer to the physical layer: power, grid hardware, semiconductors, data centres, cooling, and automation. Capital is abundant there and physical delivery is not, a gap that hands outsized returns to the owners of scarce capacity. On the second floor, as AI collapses the cost of cognition, scarcity migrates inside the services economy too: away from businesses that sell cognitive output, toward businesses defined by physical execution, regulated accountability, and presence. Public markets fund the first floor. Private capital and operators capture the second.

The argument adds a third element between the floors: the energy-first entry, where the scarcest asset in the entire stack, time-to-power, becomes a product in its own right.

It also examines where capital should not flow. As cognition commoditises, four categories compress. They are the sand of the thesis.

At Three Altitudes

For allocators. Exposure to AI does not require concentrated bets on model providers. The physical layer offers multiple durable exposures with contracted order books and capital-intensive moats, and the binding variable across all of it is time-to-power: the years between a project’s approval and reliable grid electricity. Assets that compress that variable, from grid hardware to power-ready land, earn the scarcity premium. The cleanest private entry sits upstream of the data centre entirely, in energy and land origination, where a qualified site with secured power is worth multiples of the same land without it.

For operators. The second floor is yours. In execution-heavy service businesses, AI is not a substitute but a margin multiplier: it strips scheduling friction, admin drag, reporting, and coordination cost, so throughput per operator rises without headcount. The advantage goes to operators who redesign workflows around it and capture proprietary operational data through delivery, not to those who layer tools over unchanged processes. In fragmented, regulated verticals, that is a structural consolidation advantage.

For observers. The picks-and-shovels frame comes from the gold rush: most prospectors lost money, the suppliers of picks and shovels did not. Watch the physical signals rather than the model benchmarks. Transformer lead times, turbine backlogs, interconnection queues, and cooling order books tell you where the AI economy actually is. When an unremarkable component becomes a bottleneck, its supplier earns pricing power. That pattern now repeats across the whole stack.

01 · The Migration

Two compressions drive one migration.

The first is physical. As AI moves from training into continuous inference, the binding constraints shift from the intellectual to the physical, an argument this practice first made about geography. [1] Alphabet, Amazon, Meta and Microsoft are expected to spend more than $650 billion in 2026 expanding AI capacity, and a material share of it cannot be delivered on time because the underlying infrastructure does not exist at the required pace. [2] Roughly half of planned US data centre builds are projected to be delayed or cancelled, not for want of capital or demand, but because the grid cannot support them in the timeframe. [3] Capital is abundant. Delivery is constrained. The gap is the opportunity.

The second compression is cognitive. Generative AI performs discrete knowledge tasks, drafting, summarising, structured analysis, at collapsing marginal cost, and better tooling strips out layers of coordination. McKinsey estimates generative AI alone could add 0.1 to 0.6 percentage points to annual global labour productivity growth through 2040, rising to as much as 3.4 points combined with wider automation. [25] Services are roughly two thirds of global GDP, so this is not a sector story, it is a repricing of how most economic output is produced. [27] Adoption is wide but shallow, which means the leverage is still mostly unrealised. [26]

Both compressions push value to the same place: the physical world. That is the thesis in one line. What follows maps where it lands.

02 · The First Floor: Infrastructure

Power is the binding constraint. Global data centre electricity consumption is projected to more than double to roughly 945 terawatt hours by 2030, about the current consumption of Japan, and in the United States data centres are expected to account for almost half of all electricity demand growth to 2030. [4]

Nuclear has moved from a declining industry to a pillar of hyperscaler strategy. Big tech signed contracts for more than 10 gigawatts of potential new US nuclear capacity in the year to late 2025: Microsoft’s 20-year agreement to restart Three Mile Island, Google’s order for up to 500 megawatts of small modular reactors, Amazon’s $20 billion-plus Susquehanna campus, Meta’s request for up to 4 gigawatts. [5] The SMR offtake pipeline grew from 25 to 45 gigawatts during 2025, though commercial SMR deployment remains a 2028 to 2030 story at the earliest. [6] Gas is the faster answer: Microsoft is building toward 5 gigawatts in West Texas with Chevron, Google has confirmed a 933 megawatt plant with Crusoe, Oracle has contracted up to 2.8 gigawatts of fuel cells, and roughly 30 per cent of new data centre capacity is now expected to come from on-site generation, up from effectively zero a year earlier. [7][8][9] Turbine deliveries now stretch into 2028. [11]

The grid is the quieter trade. US data centre expansion is constrained less by chips or funding than by the electrical hardware connecting facilities to power. The US ran a supply deficit of roughly 30 per cent for power transformers in 2025, imports of high-power transformers from China rose from under 1,500 units in 2022 to more than 8,000 in 2025, and about 70 per cent of the existing grid is approaching the end of its operational life. [12][13][14] A data centre goes up in under three years. New gas takes six, renewables three to six, nuclear more than ten. [15] That mismatch is the pricing power of everyone who makes, builds, or connects grid hardware.

The compute stack has visible order books. TSMC raised 2026 capital spending guidance to $52 to 56 billion, ASML guided 2026 revenue to €36 to 40 billion, and wafer fab equipment revenue is projected to reach $135 billion by 2027. [16][17][18] US data centre construction hit $41.1 billion in a single quarter in 2025, and the three largest data centre REITs grew development pipelines from $1.5 billion in 2021 to $9.8 billion. [19][20] Cooling has graduated from commodity to bottleneck: the liquid cooling market is forecast to grow from $6.65 billion in 2025 toward $29 billion by 2033, and the category leader carries a backlog around $15 billion. [21][22]

Automation is the execution hardware. Industrial robotics is forecast to grow from roughly $22 billion in 2025 to $77 billion by 2034, with demographic pressure providing demand independent of AI narrative cycles. [23] The humanoid category attracts the attention and the venture capital, but its bet is specific: that retrofitting robots to a human-shaped world beats building purpose-fit machines. In environments where the human constraint does not hold, subsurface, offshore, confined industrial spaces, extreme conditions, specialised forms are superior and commercially further along. Public markets currently overweight the humanoid bet relative to the specialised category’s scope. [24]

And the adjacent industrials are the cleanest expression of the whole thesis. The MEP contractors, specialty gases and materials firms, optical connectivity suppliers, and water infrastructure companies that the buildout cannot proceed without. Long-standing moats, industrial-sector valuations, global distribution, and demand that also serves electrification and reshoring. Less narratively priced, which is precisely the point.

03 · The Energy-First Entry

There is a position in this stack that public markets barely touch, and it may be the purest expression of the thesis: owning time-to-power itself.

The template is Crusoe Energy, which built compute on flared natural gas, energy that was being burned off as waste because it was stranded from any market. The insight generalises. Wherever energy is stranded, curtailed renewables, gas without pipeline access, grid regions with surplus generation and no load, there is an arbitrage between the near-zero value of that energy and the premium AI demand places on delivered power. The data centre is not the scarce asset. The powered, permitted, connected site is.

India makes the case concretely, and it connects this memorandum to this practice’s inference-hub thesis. [1] India curtails meaningful volumes of solar and wind in states like Tamil Nadu and Rajasthan, renewable generation with no load to serve, the flared gas of the AI era. At the same time, the capital has started arriving: Microsoft has committed $17.5 billion to India, Google is building a gigawatt-scale hub at Visakhapatnam, OpenAI and Tata have announced 100 megawatts with a path to a gigawatt, and infrastructure investors have declared multi-gigawatt pipelines through 2030. [31] Every one of those projects needs what almost none of them can assemble quickly: qualified land near substations with secured power, clear title, water, and approvals.

That defines an investable ladder with rungs at different capital intensities. At the bottom, land and power origination: qualifying sites through title, grid proximity, utilities, approvals, and community standing, then packaging them for developers and funds. A qualified, power-ready site trades at a multiple of the same land unqualified, and the work is diligence and coordination rather than construction. Above it, captive generation paired to sites. Above that, campus joint ventures with developers and hyperscalers. At the top, operating inference capacity itself, the position requiring the most capital and earning the thesis’s full premium. Allocators can enter at any rung. The scarcity being manufactured at every rung is the same: compressed time-to-power.

04 · The Second Floor: Execution

Public markets fund the shovels. The second question is who uses them, and that value accrues mostly in private markets, inside operating businesses.

Services are labour architectures: structures that convert human time, judgment, and coordination into billable output. Revenue tracks billable hours, growth tracks headcount. AI pressures all of it at once, and the evidence says the transition is underway rather than theoretical: BIS research across European firms finds AI adoption associated with roughly 4 per cent higher labour productivity, driven by capital deepening, without significant short-run employment effects. [28] The pattern is not collapse. It is repricing, from labour arbitrage to leverage arbitrage.

As cognition gets cheap, scarcity migrates to what software alone cannot deliver: regulated sign-off and liability, physical presence and site access, safety-critical execution, trusted local relationships, coordination under real-world constraints. In execution-heavy services the deliverable is not the report. It is the accountable act of doing.

That is why the strongest near-term application of AI in services is not substitution but margin expansion. In businesses where physical execution is the core deliverable, AI strips the back-office drag, scheduling, admin, reporting, compliance paperwork, coordination, so throughput per operator rises without proportional headcount. The field service management software market alone is heading toward roughly $12 billion by 2030. [29] In fragmented, regulated verticals with recurring demand, this is a structural consolidation advantage: standardised workflows, embedded automation, and proprietary operational data let platforms scale without the linear cost curve that always constrained services. The moat is not the model, models will be universally available. The moat is integration discipline plus the data captured through delivery.

The direction of travel reinforces it. Agentic systems that execute multi-step workflows autonomously are commoditising the coordination layer itself. When coordination is cheap and everywhere, advantage concentrates in operators who bolt it into defensible systems: proprietary data, regulatory compliance, accountable execution. Integration discipline, not invention, decides who wins.

For allocators the re-underwriting questions are simple. How much portfolio revenue depends on repeatable cognitive work, and how much on regulated or physical execution? Is proprietary operational data captured through delivery? Is margin driven by headcount or by process leverage? Those proportions now determine resilience, and the transition is creating valuation gaps between AI-enabled operators and structurally exposed incumbents that disciplined capital can exploit.

05 · Sand, Not Shovels

Four categories warrant underweighting rather than exposure.

Pure cognitive wrappers. Thin interfaces over general-purpose models face sustained compression as the model providers improve their own application layers and open alternatives proliferate. Defensibility requires proprietary data, workflow integration, distribution, or regulatory moats that most wrappers do not have.

Legacy business process outsourcing. The headcount-based revenue model is under structural pressure as automation compresses the cost of routine interaction. The category will not vanish, and operators credibly rebuilding delivery around AI-integrated workflows may endure, but those layering tools over unchanged processes face sustained margin erosion.

Commodity content and stock media. Stock imagery, generic editorial, basic translation, template creative: the market contracts as the unit cost of generated alternatives approaches zero. Survivors will own rights, libraries, brands, or genuine domain depth.

Low-end professional services. Standardised output production, basic legal research, commodity accounting, routine filings, formulaic marketing, compresses at the task level. The firms that scaled on labour arbitrage rather than defensible judgment or regulated accountability are the exposed ones.

The common thread: dependence on cognitive output as the primary economic unit. Where cognition was scarce these businesses scaled. As it commoditises, they compress. The question is not whether they disappear but whether current pricing reflects the reset.

06 · Horizons

Exposure runs on three clocks, and allocators should layer rather than choose. Tactically, over twelve to twenty-four months, the supply-constrained categories, transformers, turbines, lithography, high-density cooling, convert scarcity directly into earnings. Strategically, over three to five years, the capex cycle rewards incumbents with the balance sheets and manufacturing footprints to actually deliver at scale.

At macro horizon, the power layer is the most clearly structural: electricity demand, once reset upward by AI, does not revert, and grid modernisation becomes a permanent economic layer the way electrification did a century ago. Applied automation is durable with dispersion across sub-segments. Humanoid robotics remains the least certain long-horizon category, because its unit economics are not yet demonstrated at scale.

Conclusion

The AI economy is being built twice. Once in public, in the power stations, fabs, transformers, cooling plants, and robots that the buildout cannot proceed without, where supply constraints hand pricing power to the suppliers. And once in private, inside the operating businesses that absorb AI and convert it into margin, where the winners are execution-heavy platforms with regulated accountability and proprietary data. Between the two sits the scarcest asset of all, time-to-power, and the emerging trade of manufacturing it: stranded energy, qualified land, compressed delivery.

Capital positioned only in the model providers holds the narrowest and most crowded exposure to all of this. The shovels, the executors, and the energy-first entry offer the broader claim on the same transition, at better prices, with longer duration. The sand should be re-underwritten while the repricing is still early.

This memorandum is a framework, not a portfolio, and it is not investment advice. Allocation decisions belong to investors operating within their own mandates.

References

All research →