The Architecture of Economic Power

Executive thesis

The defining question of the next fifteen years is not simply which laboratory produces the most capable artificial intelligence. It is who owns the infrastructure through which machine intelligence becomes productive, commercial and politically consequential. Technological leadership, economic value capture and geopolitical influence are related, but they are not identical.

The economic intuition is straightforward: when a capability becomes easier to reproduce, some of the value it creates can move toward complementary inputs that remain difficult to expand. In artificial intelligence, these inputs include advanced manufacturing, electrical capacity, secure data access, customer distribution, enterprise integration and regulatory permissions. Yet scarcity alone does not guarantee investor returns; prices paid, reinvestment needs and political constraints matter.

Six influential works help frame this question, although they do not support one uniform prediction.

Bitcoin (2008) explored peer-to-peer transfer of digital value; Attention Is All You Need (2017) established the Transformer architecture; Situational Awareness (2024) argued for rapid AI progress and major infrastructure requirements. AI 2027 (2025) described an aggressive scenario for autonomous systems, AI 2040: Plan A proposed international coordination, and The Commanding Heights (2026) advanced a thesis about durable economic chokepoints.

This paper develops an independent interpretation from those different materials.

It distinguishes technical demonstrations from forecasts, normative proposals from empirical observations, and attractive sectors from attractive securities. Its purpose is not to predict a precise AGI date but to identify mechanisms that could reorganize capital allocation and state power.

The big idea — explained to a seven-year-old

Imagine a town where everyone suddenly gets a very clever robot helper. At first, having a robot makes you special. But when almost everyone has one, the robot alone is no longer what makes you powerful.

The robots still need electricity, computer chips, buildings and internet connections. Someone must also decide what each robot is allowed to do. The people who own these important things may earn money whenever the robots are used.

Think of a toy train. The clever robot is the train. But the train cannot go anywhere without tracks, power and stations. If you own the only safe bridge on the route, many trains may need to cross it.

What could happen next?

1. Electricity and computer equipment may become more valuable because more robots need them. But if people build too much equipment, its price could fall.
2. Some businesses may earn money by helping robots work safely together, just as traffic lights help cars avoid crashes.
3. Countries may compete to build the best computer networks and secure enough energy. They may also make rules about where powerful robots can be used.
4. Money may move more easily through digital payment systems. But those systems still need rules, safety checks and trustworthy operators.
5. Some jobs may change. People could use robots to do more work, while some tasks may no longer need to be done by humans. Nobody knows exactly how many new jobs will appear.

The lesson for someone investing money

Do not buy a company just because it says “AI.” Ask what it owns that others really need, whether customers will keep paying for it, and whether the business earns more than it spends.

And remember: owning a bridge is useful only while people need that bridge. Someone might build another route. That is why putting all your money into one idea can be risky.

In one sentence: the biggest winners may not be the people who make the smartest robots, but the people who provide what the robots need to work — although nobody can know the winners in advance.

Reader’s guide: This analogy simplifies the paper’s argument. It is not a prediction that AI will become universally cheap or that infrastructure owners will necessarily earn high returns.

I. The age of increasingly accessible intelligence

For much of the digital era, firms gained advantage by accumulating information and building software to process it. Artificial intelligence potentially lowers the cost of producing analysis, content, code and increasingly sophisticated decisions. That does not make all forms of intelligence interchangeable, nor does it establish that frontier model performance will converge completely.

The Transformer paper established a more parallelizable architecture for sequence processing within the tasks studied by its authors. It did not predict the economics of modern data centers or establish a permanent trajectory of falling inference costs. Those later developments depend on hardware, engineering, competition, demand and organizational adoption.

Situational Awareness adds a different type of argument: algorithmic progress, compute investment and AI-enabled research could reinforce one another. AI 2027 explores a faster and more disruptive sequence, but explicitly presents a scenario rather than a settled forecast. Both works are valuable as stress tests for strategic planning precisely because their assumptions can be challenged.

The implication for investors is to separate three questions: will a technology improve, will spending on it increase, and will the owners of particular assets earn returns above their cost of capital? An affirmative answer to the first does not settle the second or third. In rapidly expanding industries, intense competition and high capital expenditure can transfer value to customers rather than shareholders.

An alternative world is also plausible: frontier intelligence remains expensive and concentrated while lower-cost models serve narrower use cases. Under that configuration, the strongest laboratories could retain substantial pricing power. A credible investment thesis must therefore work under multiple model-market structures rather than assume commoditization as an accomplished fact.

II. The return of physical scarcity

AI appears immaterial to its users but depends on physical systems. Advanced chips require complex fabrication and packaging; data centers require land, electrical connections, cooling, networking and operating expertise. Even where computational hardware is available, an unavailable grid connection can delay deployment.

The relevant unit of scarcity is often not electricity in the abstract but deliverable power at a particular site and time. A project with permits, grid access, resilient connectivity and a creditworthy customer can have a different economic profile from an otherwise similar undeveloped site. The value of these attributes depends on contract duration, regulatory conditions and the cost of securing them.

Cooling creates another layer of constraint. Water consumption varies with local climate, cooling design and electricity generation, so blanket statements about the water intensity of AI are misleading. Water stress nonetheless can alter the economics and social acceptability of projects in particular regions.

The resulting capital chain runs from equipment makers and engineering firms through power producers, grid operators, real-estate owners and data-center operators. It is tempting to label all of them winners. That would confuse demand growth with margin durability and overlook the possibility that capacity additions eventually erase scarcity premiums.

Investors should distinguish temporary bottlenecks from defensible economic franchises. A transformer manufacturer may enjoy an order backlog without possessing permanent pricing power. An infrastructure asset may be strategically indispensable while its regulated return limits upside; the same state that protects it may constrain its economics.

III. Where the rents may migrate

The Commanding Heights, published by Will Manidis and Joe Roberts in October 2026, places control of scarce positions at the center of its thesis. Its argument is not that intelligence ceases to matter, but that economic gains may be appropriated by actors who own hard-to-bypass routes to customers, capital and essential resources. This is a powerful hypothesis, not an empirical proof of future equity performance.

The distinction between producing value and capturing value is central. A model provider may create substantial productivity gains that are captured by enterprise customers through lower costs. A distributor may collect fees without owning the underlying technology, while an infrastructure owner may earn stable revenue but little upside after financing and maintenance costs.

A useful framework separates the AI economy into production, distribution and authorization. Production includes chips, compute and models; distribution includes software, customer relationships and embedded workflows. Authorization includes identity, policy enforcement, auditability, payment permissions and the ability to determine whether a proposed AI action can actually be executed.

These layers overlap in practice, and the boundaries are contested. Hyperscalers can operate across all three; enterprise software providers may integrate orchestration directly into existing applications. The strategic question is whether a particular firm controls a durable bottleneck or merely resells a feature that others can reproduce.

The strongest opportunities may lie where switching costs, regulatory obligations and trusted workflows intersect. Yet even mission-critical software can be displaced by new standards or customer consolidation. No architecture should be treated as a permanent monopoly before the economics of interoperability, procurement and customer ownership are understood.

IV. The geography of compute

Compute is becoming a component of national industrial strategy because advanced AI depends on cross-border supply chains. Design, fabrication, advanced packaging, networking, energy and cloud deployment are geographically distributed. That creates both efficiency and vulnerability.

The United States benefits from a deep ecosystem of software, capital markets, research and major technology companies. China has significant industrial capacity and a strategic incentive to reduce dependence on foreign technologies. Europe combines industrial competencies and regulatory influence with persistent challenges in financing scale, infrastructure deployment and market integration.

The central geopolitical issue is not a simple ranking of nations by model performance. It is the resilience of each national system under export controls, energy shocks, technology restrictions and capital-market disruption. A country may possess excellent researchers yet lack the physical capacity or financial incentives to deploy systems at scale.

AI 2027 frames this competition through a scenario of accelerated technological rivalry. AI 2040: Plan A considers a sharply different possibility: negotiated limits, verification and international cooperation. The difference is analytically useful because it shows how policy choices could alter both the pace of technical progress and the identity of economic beneficiaries.

The Eurasian landmass matters through energy, minerals, logistics and industrial connectivity. Central Asian economies, including Kazakhstan, may have opportunities to connect resource endowments with European and Asian demand. However, resources alone are not equivalent to a competitive AI ecosystem: power reliability, connectivity, skills, governance and access to capital remain decisive.

V. Digital money and the geopolitics of settlement

Bitcoin's original contribution was a protocol for electronic value transfer without reliance on a trusted central intermediary to prevent double spending. It is not an empirical demonstration that Bitcoin must appreciate or that all future machine payments will use it. Its relevance here is architectural: it established one route toward digital settlement outside conventional intermediated systems.

Dollar-denominated stablecoins follow a different logic. Their design generally depends on identifiable issuers, reserve assets, operational infrastructure and legal arrangements. They seek a relatively stable unit of account rather than the floating price exposure associated with Bitcoin.

The Commanding Heights connects dollar stablecoins to the possibility of renewed monetary influence. If issuers hold short-term US government obligations as reserves, larger circulation can create demand for those instruments. The scale and persistence of that effect depend on regulation, reserve composition, interest rates and the extent to which activity displaces rather than adds to existing dollar demand.

AI agents could increase interest in programmable payments, but automated commerce requires more than a token. It requires identification of the principal, spending limits, fraud controls, revocation rights, dispute mechanisms and reliable audit trails. Payment rails and authorization systems may therefore become complementary infrastructures.

For investors, Bitcoin exposure, equity in a stablecoin issuer and ownership of payment infrastructure are distinct propositions. Their return drivers, regulatory exposures and downside risks differ materially. Treating them as interchangeable expressions of a single digital-money thesis would obscure rather than diversify risk.

VI. The capital–intelligence nexus

Industrial-scale AI can be capital intensive even when its end-user applications reduce the cost of individual tasks. Semiconductor plants, data centers and power systems require financing long before their ultimate utilization and profitability become clear. The speed of technological change creates an additional problem: assets may depreciate economically faster than their financing structures assume.

An investment boom can be socially productive and financially disappointing. If many providers expand capacity simultaneously, utilization or prices may fall below underwriting assumptions. The same dynamic has appeared in other infrastructure and technology cycles, although the timing and magnitude of any AI-related overcapacity cannot currently be confirmed.

Capital allocation should therefore begin with cash-flow mechanics. How much incremental investment is required for each unit of revenue? What share of cash flow is consumed by replacement hardware, grid upgrades and customer acquisition? Can contracts pass through power-price risk, and who bears the cost of technical obsolescence?

Private equity has a potentially different opportunity from venture capital. Rather than financing new frontier models, it can seek established companies that own specialized workflows, industrial components or customer relationships. AI can improve such businesses, but only if implementation costs, security obligations and customer adoption permit durable margin or revenue gains.

This also reframes the capital–labor debate. AI may substitute for some cognitive tasks while complementing other forms of work; effects on wages and employment will vary by sector and institution. Ownership of productive assets could become more important to the distribution of income, but broad productivity growth does not imply uninterrupted equity-market appreciation.

VII. The control layer

As organizations move from AI systems that recommend actions to systems that initiate them, the right to act becomes a separate source of value. An agent may propose a payment, contract amendment or credit decision. Whether the action is permissible depends on authority, policy, legal obligations and sometimes human approval.

An AI orchestration layer can coordinate agents, data access, workflows and monitoring. Governance and authorization may be implemented within that layer or through distinct security and enterprise systems. The architecture is not settled, and claims of a single inevitable operating system would be premature.

For a bank, the distinction between generating a recommendation and executing a transaction is fundamental. The same applies to medical records, public procurement and industrial control. Systems that can prove who authorized an action, under which rules and with what evidence may enable wider automation without relinquishing accountability.

The commercial opportunity is attractive in theory because trust and integration can create switching costs. The competitive risk is equally real: cloud providers, identity platforms, ERP vendors and open standards may absorb much of the functionality. Investors should evaluate whether a product owns a defensible control point or is a temporary integration wrapper.

The more consequential question is institutional. Who is accountable when multiple autonomous systems interact, when permissions conflict or when an action crosses jurisdictions? The answers will be shaped as much by law and governance as by model capability.

VIII. Three futures, 2026–2041

Scenario one is rapid acceleration.

AI capability and deployment improve quickly, stimulating large investments in compute, energy and automation. Economic gains could be substantial, but overbuilding and concentrated ownership may also increase financial and political instability.

Scenario two is gradual industrialization.

Models improve while enterprise adoption is constrained by integration, security, skills and institutional inertia. Under this trajectory, incumbents with trusted distribution and durable cash flows may capture a significant portion of the gains.

Scenario three is fragmentation and coordinated restraint.

Export restrictions, security incidents, resource constraints or international agreements alter the pace and geography of deployment. Some assets benefit from regional redundancy while others lose scale economies or market access.

These scenarios are not assigned numerical probabilities because the six source documents do not justify reliable estimates. They also are not mutually exclusive: different regions and industries could follow different paths simultaneously. Their value lies in exposing hidden assumptions in investment and policy decisions.

A robust strategy would not require the most aggressive AGI timeline to materialize. It would emphasize balance-sheet resilience, contractual protections, customer diversification and the ability to redirect capital when technology or regulation changes. Strategic optionality may be worth more than maximal exposure to a single forecast.

IX. The investment test

An investor considering AI-linked assets should ask first whether the firm owns a resource or relationship that remains hard to replace as intelligence becomes cheaper. That question must be followed by an assessment of whether the firm can retain the resulting economic surplus. Control is valuable only if customers, competitors and regulators permit attractive returns.

Return on invested capital should be evaluated against the cost of capital through the cycle. Free cash flow should be measured after necessary reinvestment rather than inferred from revenue growth or adjusted earnings alone. Valuation should reflect plausible utilization, pricing and technological depreciation, not only the most optimistic market narrative.

Concentration deserves special attention. Broad market portfolios already have exposure to leading technology companies, and thematic funds may add overlapping positions. A nominally diversified AI strategy can conceal heavy dependence on the same semiconductor cycle, capital-spending plans or US monetary conditions.

The same discipline applies to digital assets. A protocol's historical innovation does not establish its future return, and stablecoin transaction growth does not automatically translate into equity value. Each position requires its own cash-flow, legal and liquidity analysis.

The most promising assets may ultimately be unglamorous. Grid equipment, industrial software, compliance infrastructure and specialized distribution can be economically central without appearing at the frontier of model research. The challenge is identifying which positions are truly durable before the market prices in that durability.

X. The strategic imperative

For governments, the priority is to connect technology policy to industrial capacity. Talent, research, power infrastructure, capital formation and secure supply chains must be treated as interacting systems. Rules without implementation capacity risk limiting domestic deployment without materially changing global technology leadership.

For Europe and Italy, a plausible route is selective specialization rather than imitation of every frontier laboratory. Industrial automation, power systems, enterprise software, cybersecurity and regulated-sector integration offer domains where existing capabilities can matter. Success would still require faster execution, investable project pipelines and sufficient scale.

For executives, the key question is where AI enters the organization's authority structure. Pilot projects can demonstrate productivity, but enterprise value depends on process redesign, data quality, permissions and measurable outcomes. The shift from assistance to execution should be governed deliberately rather than left to individual vendors.

For private capital, the priority is to distinguish assets that merely participate in a fast-growing market from assets that can capture cash flows through several technology cycles. That requires technical diligence, operational understanding and geopolitical analysis in a single investment process. The objective is not to forecast every innovation but to own resilient positions in the systems that innovations must use.

The central paradox of the coming period is that increasingly accessible intelligence may make certain forms of scarcity more valuable. Yet scarcity can migrate, dissolve or become regulated. The winners will not necessarily be those who announce the most advanced technology, but those who combine productive capability with durable economics and legitimate institutional authority.

Conclusion — Power after abundance

The digital economy is being reorganized along two axes: the production and distribution of intelligence, and the transfer and authorization of value. Their convergence could change the geography of investment and the instruments of state power. Neither outcome is predetermined.

The six works examined here provide useful but different lenses: technical foundations, rapid-development scenarios, coordinated-governance proposals and a theory of economic chokepoints. They should not be flattened into a single prediction. The most credible synthesis is conditional: if intelligence becomes more accessible, scarce complementary assets and control rights may capture a greater share of the gains.

The defining investment challenge is therefore not simply to identify the fastest-growing technology. It is to identify the ownership structures, contractual rights and institutional positions that can convert technological progress into sustainable cash flow. The defining geopolitical challenge is to understand which states can assemble the physical, financial and regulatory foundations of that progress.

The new commanding heights may be less visible than the models that attract public attention. They may lie in grid connections, industrial capacity, payment permissions, embedded workflows and governance systems. Understanding who controls them is an essential part of understanding the next global economy.

Sources and methodological notes

• Satoshi Nakamoto, Bitcoin: A Peer-to-Peer Electronic Cash System (2008)
• Vaswani et al., Attention Is All You Need (NeurIPS, 2017)
• Leopold Aschenbrenner, Situational Awareness (2024)
• AI Futures Project, AI 2027 (2025)
• AI Futures Project, AI 2040: Plan A and Economics of Plan A
• Will Manidis and Joe Roberts, The Commanding Heights (2026)

Method: This essay is an original synthesis and does not claim that its six sources agree on a single future. The first two are technical works; the remaining materials include speculative scenarios, policy proposals and economic arguments. Forward-looking statements are hypotheses, not verified forecasts. No investment recommendation is intended.