As we pass the mid-year mark, it is perhaps useful to take stock of current themes that occupy the mind of the markets.

Fiscal sustainability

Global public debt reached just under 94% of GDP in 2025 and is now projected to hit 100% by 2029, according to the IMF. The accumulation is driven primarily by the world’s major economies, with mounting spending pressures across social needs, defence, and strategic autonomy compounding rising interest burdens.

In the US, the CBO projects a federal deficit of 5.8% of GDP, well above the 50-year historical average of 3.8%. The One Big Beautiful Bill Act made matters worse. Higher average interest rates on government debt, partly driven by policy changes, push debt further still and stresses debt service.

China’s general government debt crossed 100% of GDP in 2026 on IMF projections (107%, counting on-balance-sheet local government financing vehicles), with debt projected to reach 127% by 2031, the second-largest rise after the US. Fiscal policy should remain expansionary until the economy reflates durably, but over the medium term, ensuring sustainability will require significant fiscal consolidation, reforms, and a restructuring of LGFV debt to tackle the local government debt overhang.

India is the relative bright spot among major EMs. Growth expanded 7.8% in the first quarter of FY2025/26. Fiscal consolidation has advanced and the current account deficit has been contained. State-level finances remain weak, however. Consolidated state debt stood at 28.2% of GSDP in FY2022/23, up from 22.2% in FY2012/13. The growing interest payment burden adds to already-high non-discretionary expenditure including salaries and pensions, explaining high-debt states’ difficulties in consolidating their deficits.

Europe: The aggregate deficit ratio in the euro area is expected to increase to 3.2% of GDP in 2025 and 3.3% in 2026, partly as rising defence spending pushes up government expenditure. Public debt is forecast to rise to 88.8% of GDP in 2025 and 89.8% in 2026.

Germany’s 2026 federal budget envisages expenditure of €524.5 billion, with defence spending rising to €83 billion plus an additional €25.5 billion from the off-budget Bundeswehr Special Fund. Public debt is projected to reach 80.25% of GDP by 2029, and Germany risks being placed under an excessive deficit procedure in 2026 and 2027. France remains the euro area’s most acute sovereign risk. A compromise budget should allow the deficit to reach 5% of GDP in 2026, compared with 5.4% in 2025, pushing public debt to 118% of GDP.

UK: The OBR projects public sector net borrowing to fall from 5.2% of GDP in 2024/25 to 4.3% this year and then to 1.6% by 2030/31. Public sector net debt is expected to be broadly stable and settle at around 95% of GDP in the early 2030s. The public debt position remains fundamentally unsustainable, and a serious medium-term plan to bring debt down as a share of the economy will be needed.

Brazil is the most immediate concern in the emerging market universe. The consolidated public sector ran a primary deficit of nearly R$25 billion in the first five months of 2026, reversing a surplus of around R$69 billion in the same period of 2025, a YOY swing of roughly R$94 billion. Gross general government debt reached 81.1% of GDP, the highest in five years.

Argentina: Milei’s reform agenda of reducing public spending and subsidies, partially liberalising the exchange rate, and introducing the RIGI framework for large investments has helped put Argentina back on the map for international investors.

US and Brazil occupy the two most concerning positions, the US by sheer systemic weight and trajectory, Brazil by the combination of high interest rates, political cycle dynamics, and a structural spending-revenue mismatch. China’s official numbers are manageable, but the augmented balance sheet is not. The EU is muddling through with a new framework that is already bending under political pressure. India is the cleanest picture: high growth providing the denominator effect, a credible anchor, and consolidation progress,  though state-level debt deserves watching. The UK is stable but structurally boxed in with almost no headroom. Across Latin America, the election-year Brazil risk is the dominant variable for the region’s sovereign credit story.

These conditions are not robust against significantly rising interest rates thus making inflation a key factor in fiscal sustainability. Interest rates will be influenced by fiscal policy, confidence, and inflation. There is not much room for debt monetization to tame interest rates as that would only serve as an accelerant to inflation. Central banks face a narrow policy window, keeping inflation and thus rate expectations under control, while avoiding directly elevating debt service for sovereigns. Inflation can come from other sources, such as industrial policy and geopolitics, as we have seen. The impact of AI on productivity and thus inflation is yet another factor.

To do:

• Monitor the trajectory of interest rates and exchange rates. The importance of interest rates on funding and valuations will be significant. Maintain a short duration posture in the credit book, favouring floating rate product.
• Given the relative tensions and constraints in US fiscal policy, maintain an underweight position in USD.
• Be cautious long duration assets such as core or core plus real estate and infrastructure.

AI

The productivity case for AI is real but unevenly distributed. Task-level evidence is consistent: the BIS notes that task-level studies consistently show productivity gains of 20% to 50% in time savings.  But there is a well-documented gap between task-level gains and total factor productivity or TFP the “productivity paradox” that attended computing, the internet, and every prior general-purpose technology.

The labour market implications are asymmetric. AI automates cognitive, white-collar tasks. Roles involving routine cognitive work such as coding, translation, basic analysis face the highest displacement risk. A “reasoning wage premium” is emerging that benefits workers who can orchestrate models, while a “growth without hiring” trend sees companies expanding output without proportional headcount increases. Physical jobs remain largely unaffected: automating fine motor skills is harder than automating high-level reasoning. The conditions under which a rapid AI productivity boom can coexist with, and indeed cause, a macroeconomic contraction are real and underappreciated. If productivity gains accrue to capital owners while labour income stagnates, the consumption multiplier is lower than historical cycles, and fiscal stabilisers face greater strain.

The scale of the AI investment cycle is unprecedented. Wall Street estimates total AI capex could exceed US$1 trillion in 2027. The five largest hyperscalers are on pace to spend more than US$1 trillion on AI-related capex across 2025 and 2026 combined, a sum that is already outpacing earnings and free cash flow, forcing some to issue debt to cover the gap. Somewhat concerning is the circular nature of the AI ecosystem’s financing. Hyperscalers take equity stakes in AI labs, which in turn commit to multi-year purchases of chips or computing power from those same hyperscalers. This creates a closed, recursive financing loop.

AI intersects with monetary policy in two ways. First, the productivity question directly affects the neutral rate of interest. If AI delivers a Solow-style productivity boost, the neutral rate rises because the return on capital rises. Central banks would need to tighten more. Second, the AI capex boom is itself inflationary in the near term. Massive investment in data centres, power infrastructure, and chips puts upward pressure on prices in those specific markets. Power demand from data centres is already a material constraint in the US, Europe, and parts of Asia.

AI creates a potential fiscal dividend, higher growth, broader tax base, but only if the productivity gains are large, fast, and distributed broadly enough to raise labour income and corporate profits in taxable forms. This is not guaranteed.

The bull case is essentially the post-war US analogy: a genuine general-purpose technology that takes time to diffuse but ultimately raises TFP broadly, lifts real growth sustainably above neutral rates, and makes current debt loads manageable.

The bear case is the dot-com analogy. Returns disappoint relative to the capex committed, the circular financing loop unwinds, equity valuations crash, financial conditions tighten, and at a moment when sovereign balance sheets have no room for fiscal expansion.

The stagflation case: AI delivers real productivity gains, but they accrue to capital rather than labour, suppressing aggregate demand while the AI capex boom sustains inflationary pressure in factor markets. Central banks face rising prices and weak demand simultaneously.

To do:

• Underweight the hyperscalers and their ecosystem, particularly those within their circular financing loops.
• Just avoid the mania.
• Seek zero cost, levered, net short trade expressions referencing these companies.

Energy

Global data centre electricity consumption was approximately 415 TWh in 2024, representing about 1.5% of global electricity use, growing at 12% per year over the last five years. The IEA’s base case has global electricity consumption for data centres is projected to reach around 945 TWh by 2030, representing just under 3% of total global electricity consumption. That is roughly equivalent to the entire current electricity consumption of Japan. In the IEA’s “Lift-Off” scenario, demand exceeds 1,700 TWh by 2035. In the US specifically, data centres currently consume the equivalent of 4.4% of total US electricity, projected to rise to between 6.7% and 12% by 2028. A single AI task can consume up to 1,000 times more electricity than a traditional web search, and AI-focused facilities now require 80 MW of power, more than double the 32 MW standard data centres consume.

The energy shortage is not primarily a problem of insufficient generation capacity in aggregate. It is a problem of the wrong capacity, in the wrong places, connected through infrastructure with lead times that cannot match the pace of AI deployment. The bottlenecks are a) transmission and interconnection, b) ageing grid architecture, c) geographic concentration, and d) hardware supply chains (can’t make enough equipment quickly enough.)

Constrained by slow grid connections, data centre developers in the United States are pushing forward projects with onsite natural gas-based power generation. Total power generation for renewables is projected to grow 22% per year until 2030, meeting nearly half the anticipated growth of data centre electricity demand. However, renewable intermittency conflicts with data centre 24/7 uptime requirements. The first commercial SMR-powered data centres will come online by 2030 at the earliest, with 22 GW of SMR projects in global development. Capital expenditure of just five hyperscaler technology companies is now larger than global investment in oil and natural gas production. The problem is not in generation or storage but in evacuation and distribution.

To do:

• Grid infrastructure is the most direct and least speculative investment thesis. High-voltage substations, transformers, cables, and transmission expansion face structural demand that is independent of which AI model wins or whether productivity gains materialise. The constraint is physical, the demand is contractual, and the regulatory moats are durable.

Climate

Global investment in climate mitigation reached a record $2.3 trillion in 2025, up 8% from 2024. The largest drivers were electrified transport ($893 billion), renewable energy ($690 billion), and grid investment ($483 billion). Against the need, it is less than half of what is required. To keep the world on a net-zero pathway, investment in climate mitigation must rise to between $6.2 trillion and $9.5 trillion per year by 2030, leaving an annual gap of $4.5-7.8 trillion compared to 2023 flows of $1.7 trillion. The geographic mismatch is severe. Since 2015, renewable energy investment in emerging markets excluding China has nearly tripled to $140 billion in 2024, but developing economies’ share of global clean energy spending has averaged just 18% over the past decade. In contrast, developed economies and China together capture 82% of total funding.

UNEP’s 2025 Adaptation Gap Report finds that adaptation finance needs in developing countries by 2035 are over $310 billion per year, 12 times as much as current international public adaptation finance flows. UNEP estimates the private sector’s realistic contribution at $50 billion per year toward national public adaptation priorities, ten times greater than current private flows, but still only one-seventh of total need.

The mitigation/adaptation asymmetry in private capital flows is not accidental. It reflects fundamentally different return structures.

Mitigation investments such as solar, wind, EVs, grid storage, generate revenue through energy or transport services. They have clear offtake mechanisms, liquid secondary markets, bankable cash flows, and increasingly competitive economics against fossil fuel alternatives. The private sector has learned how to price, structure, and exit these assets.

Adaptation is structurally harder because most of its value is in avoided loss rather than generated revenue. The beneficiary and the payer are often different parties, creating a public-goods problem. Many adaptation investments deliver broad public benefits but do not generate clear or predictable financial returns, making them less attractive to investors.

Opportunities:

Physical losses are becoming significant insurance events. Insurers are re-pricing or withdrawing coverage from high-risk geographies. Closing the insurance protection gap is one of the most specific short-term priorities from the COP30 with the insurance industry directed to work with climate-vulnerable developing countries to reduce the financial protection gap. Taxonomy and measurement infrastructure is improving. COP30 launched a set of common principles to make the world’s more than 50 sustainable taxonomies interoperable, with the potential to cut barriers to sustainable finance and reduce transaction costs for cross-border adaptation investments. Blended finance is scaling into adaptation. Southeast Asia faces particularly acute exposure with Vietnam, the Philippines, Indonesia, Thailand and Cambodia among the most climate-vulnerable economies globally.

To do:

Direct origination towards

1. adaptation,
2. insurance,
3. blended finance and
4. vulnerable regions such as Southeast Asia.

Private credit

Private credit has reached an inflection point, not in crisis, but entering its first genuine “test” as a major asset class.  The market has grown to an estimated US$1.5–2 trillion in assets and, at its current size and scope, has not been tested during a severe economic downturn. Signs of stress are emerging.

Headline default rates understate true stress. While commonly cited default rates in private credit often remain around 2–3%, more comprehensive measures indicate higher levels of distress, circa 5.8% for the trailing 12 months through January 2026. The widespread use of PIK interest is a signal that cash flows are under stress.

The growing use of private ratings, sometimes from lesser-known providers, to facilitate investment by rating-reliant investors such as insurers warrants monitoring.

Interconnections between private credit funds and banks, insurers, and private equity firms are deepening, raising potential vulnerabilities.

Semi-liquid vehicles for the wealth channel now command almost a third of the $1 trillion US direct lending market. Liquidity issues stem from the growing popularity of funds offering redemption options to investors, which may heighten the procyclicality of private credit.

Over the past decade, major private equity firms have either acquired life insurers outright or built their own. The structure creates a three-layered conflict: the PE firm originates assets, the affiliated insurer provides the permanent capital to fund them, and captive offshore reinsurers reduce the regulatory capital burden. Each layer individually passes regulatory scrutiny; the compound effect is a largely opaque self-dealing ecosystem where policyholders’ retirement assets are effectively funding the PE firms’ own credit books. The system has not yet been tested through a genuine credit cycle at this scale.

To do:

Tighten fund due diligence standards to account for potentially directly or indirectly conflicted GPs.

AI and Fiscal Reality

The history of human development is replete with technological advances that have extended and leveraged human ability, allowing us to accelerate growth and development. Some technologies are disruptive and set back significant proportions of the population in terms of employment and income and thus welfare. Until AI, no other technology has had the potential to replace or displace human ability more than its potential to extend or amplify it.

As such, a policy that continues to tax labour in the form of personal income taxes, will be insufficient to balance the fiscal books. AI’s displacement of human ability will also introduce new interventions and mitigations, policies which will likely more funding. Combine this with ageing populations and rising dependency rates, and a fiscal imbalance looms.

What alternatives and options are there in terms of raising tax revenue to replace income tax revenues and indeed to raise tax revenues beyond current levels to fund AI displacement mitigation policies? A number of technical solutions present, but ultimately, the solution lies in the purview of political economics.

AI displacement of labour and the consequences for income tax revenue is a truly hard fiscal design problem, and the historical analogies only go so far, the shift from agricultural to industrial labour took generations and was partly absorbed by demographic expansion. AI displacement, if it proceeds at pace, compresses that timeline dramatically while coinciding with demographic contraction in most advanced economies.

The core problem is that income tax is a tax on human productive activity. As AI substitutes for that activity, not just in routine tasks but progressively in cognitive and creative work, the tax base erodes just as fiscal demands rise (displacement mitigation, ageing populations, healthcare, potentially universal basic income-adjacent transfers). Capital income, meanwhile, accrues disproportionately to AI owners and deployers. So the political economy of taxation has to shift from taxing the application of human effort to taxing the ownership and use of productive capital.

Fiscal Architecture for a Post-Labour Economy

The erosion of labour income as a tax base is a logical consequence of capital substituting for human productive capacity. The emotional responses this provokes are understandable but irrelevant to the design problem. What follows is an assessment of alternatives, evaluated on their merits.

The Structural Problem

Income tax is a levy on human effort. As artificial intelligence systematically displaces that effort, the base shrinks. Capital returns, meanwhile, concentrate among those who own the displacing systems. Fiscal demand rises precisely as the conventional revenue mechanism weakens. The logical response is to shift taxation from the application of human effort to the ownership and deployment of productive capital. Several options are available.

Option 1: Restructured Corporate Profit Taxes

Tax the entities that capture AI-generated value more directly and at higher effective rates, using unitary taxation to prevent jurisdictional arbitrage.

Pros: Targets value where it concentrates. The OECD Pillar Two framework establishes a precedent for international coordination. Administratively familiar.

Cons: Current rates and structures are demonstrably inadequate. Multinational profit-shifting has outpaced regulatory capacity for decades. Political resistance from large technology firms is substantial.

Practicality: Moderate. The architecture exists; the political will does not. Incremental progress is likely; transformative reform is not, absent a coordinated multilateral agreement considerably more robust than what currently exists.

Option 2: Automation or Robot Tax

A levy on firms equivalent to the income and payroll taxes foregone when human workers are replaced by automated systems.

Pros: Logically consistent with the problem it addresses. Directly links the displacement to its fiscal consequence. Creates a mild moderating effect on displacement velocity, which may be socially useful.

Cons: Defining the taxable unit is non-trivial. Software-based AI embedded in ordinary business processes resists clean categorisation. Risk of discouraging productivity-enhancing deployment.

Practicality: Low to moderate. No major economy has implemented this seriously. Definitional and administrative complexity is genuine, not merely a political excuse. Requires significant regulatory innovation before it becomes workable.

Option 3: Wealth Taxes

An annual levy on net financial wealth above a defined threshold, capturing returns to AI ownership without waiting for realisation events.

Pros: Directly addresses concentration of capital returns. Relatively immune to the labour displacement dynamic. Captures value that income and capital gains taxes miss when gains are deferred.

Cons: Capital flight is real, though manageable with multilateral coordination. Valuation of illiquid assets is administratively demanding. Several European experiments, notably France and Sweden, were repealed due to practical difficulties.

Practicality: Low without international coordination; moderate with it. The political obstacles are substantial. Technically implementable but historically fragile.

Option 4: Capital Gains Reform

Taxing capital gains at income-equivalent rates, and potentially taxing unrealised gains above a threshold.

Pros: Corrects an existing asymmetry that has no principled justification. Broad base. Administratively integrated with existing systems.

Cons: Unrealised gains taxation creates liquidity problems for asset-rich, cash-poor individuals. Realisation-based reform is less contentious but also less effective at capturing concentrated AI equity wealth.

Practicality: Moderate. Rate equalisation between capital gains and income is achievable and has been done in various jurisdictions. Unrealised gains taxation faces stronger resistance and implementation complexity.

Option 5: VAT Expansion and Progressive Consumption Taxes

Broadening consumption taxation, either through VAT rate increases or progressive consumption tax structures that exempt net saving.

Pros: Efficient, broad-based, and relatively difficult to avoid. A progressive consumption tax elegantly addresses the dynamic where capital owners accumulate without consuming.

Cons: Standard VAT is regressive in distributional terms, requiring corrective transfers. Progressive consumption tax is administratively novel; no large economy has fully implemented it.

Practicality: High for VAT expansion. Moderate for progressive consumption tax design. The political obstacle to VAT increases is manageable; the distributional optics require careful handling.

Option 6: Financial Transaction Taxes

A small levy on trades of equities, bonds, and derivatives.

Pros: AI-driven high-frequency trading generates enormous transaction volumes. Even a minimal rate produces substantial revenue. Broad application reduces avoidance.

Cons: Market liquidity effects, though often overstated in theoretical models, are real at scale. Jurisdictional arbitrage is possible if not applied broadly.

Practicality: Moderate. The EU has debated this for over a decade without resolution. Revenue potential is genuine; coordination requirements are significant.

Option 7: Land Value Taxation

A tax on the unimproved value of land, capturing socially generated appreciation rather than private effort.

Pros: Economically efficient: land cannot be relocated or reduced in supply in response to taxation. Progressive in effect. AI-driven agglomeration will likely increase land value concentration, making this more rather than less relevant over time.

Cons: Requires revaluation infrastructure that most tax authorities lack. Politically contentious among property-owning constituencies. Transition costs are real.

Practicality: Moderate in the long run. The economic case is exceptionally strong. The political case requires sustained effort. Several jurisdictions (Australia, Taiwan, Denmark) have partial implementations that demonstrate feasibility.

Option 8: Sovereign AI Fund

Public acquisition of equity stakes in AI infrastructure entities, either through a sovereign wealth fund or through licensing requirements on large AI operators.

Pros: Aligns public fiscal interest with AI productivity gains structurally rather than through periodic tax extraction. The Norwegian sovereign wealth fund demonstrates long-run viability. Revenue scales with AI value creation rather than against it.

Cons: Requires either significant public capital for acquisition or regulatory leverage for mandatory equity issuance. Political opposition from private sector would be intense. Governance complexity is substantial.

Practicality: Low in the near term; potentially high as a long-term structural response. The logic is sound. The implementation pathway is unclear and politically demanding.

Option 9: Data Extraction Levies

A royalty on commercial use of data generated by the population, analogous to natural resource extraction fees.

Pros: Targets a genuine and underpriced input to AI value creation. Principled: the data is not created by the firms that monetise it.

Cons: Valuation methodology for data does not exist at the required level of precision. Definitional boundaries are contested. Administratively novel with no established precedent at scale.

Practicality: Low currently. The conceptual foundation is sound; the administrative infrastructure does not yet exist. A candidate for medium-term development rather than near-term implementation.

Summary Fiscal Assessment

The logical sequence for any jurisdiction serious about this problem is: first, expand and reform consumption taxation, as this is available now; second, pursue coordinated corporate tax reform more aggressively than current frameworks allow; third, introduce land value taxation and capital gains reform as medium-term structural corrections; and fourth, begin designing the sovereign fund and data levy frameworks now, as they will take a decade to implement properly.

Emotional attachment to the existing income tax architecture is not a rational basis for fiscal policy. The evidence that it will be insufficient is already accumulating. The instruments to replace it exist. What is lacking, as is frequently the case in human affairs, is the application of logic to political will.

However

A purely fiscal response to AI displacement, however well designed, addresses the symptom. The symptom is a revenue shortfall. The condition is an ownership structure that was not designed for a world in which capital can substitute for human labour at scale and at speed.

The question of who owns the means of production in an AI economy is not a secondary question to be resolved after the fiscal architecture is designed. It is the primary question, from which the fiscal architecture should follow.

The deeper issue is one of structural legitimacy. If the productive capacity of an economy is increasingly owned by a narrow class, and that capacity displaces the labour income of the broader population, then taxation and transfer become an indefinitely escalating contest between concentration and redistribution.

The Ownership Question

The means of production in an AI economy are not factories or land in the conventional sense. They are compute infrastructure, trained models, proprietary data, and the network effects that accrue to dominant platforms. These are currently owned almost exclusively by a small number of corporations and their shareholders. There are three coherent structural responses to this, each with distinct implications.

Redistribution through taxation and transfer. The approach implicitly assumed by all nine options above. Ownership remains concentrated; the state extracts and redistributes. It is familiar and does not require restructuring property rights. Its weakness is that it is permanently reactive, politically unstable under democratic pressure, and vulnerable to the lobbying and avoidance capacity of those being taxed. It also does nothing to address the legitimacy deficit that arises when a population experiences itself as a recipient of transfers rather than a participant in production.

Distributed ownership by design. Structural mechanisms that ensure broader population ownership of AI productive assets from the outset, rather than attempting redistribution after concentration has occurred. A sovereign wealth fund acquiring equity is one version. Mandatory profit-sharing or equity issuance to employees or citizens is another. Alaska’s Permanent Fund, which distributes resource rents as a citizen dividend, is a small but instructive precedent. The logic is that if the population cannot sell its labour to AI systems, it should instead own a share of them.

Public or common ownership of AI infrastructure. The most structurally radical option. Treating foundational AI infrastructure, large language models, compute networks, and data repositories as public utilities or commons, subject to public governance rather than private accumulation. This does not require nationalisation of every application layer, but it does require a deliberate decision that the foundational layer is not appropriately governed by private ownership alone. Historical analogies include public roads, the internet protocol stack, and the electromagnetic spectrum, all of which were treated as public infrastructure despite generating enormous private value on top of them.

The current energy transition presents us with a number of practical issues in search of solutions.

Managing the Compute-Energy Coupling

The physical baseline of data processing remains unalterably anchored to Landauer’s Principle. Universal law dictates that changing a single bit of information must inevitably dissipate a minimum threshold of thermal energy:

Where k represents the Boltzmann constant and T denotes the absolute temperature of the circuit. While this core physical link cannot be broken, macro-efficiency is pivoting toward software optimizations and architectural overhauls. Think of these as micro fixes and patches.

• Algorithmic Pruning & Quantization: Short-to-medium-term efficiency gains are driven by a shift from raw scaling to specialized model compression. Low-bit quantization (reducing mathematical precision from 16-bit to 4-bit or 2-bit operations) and Sparse Mixture-of-Experts (MoE) architectures, which activate only a specific mathematical fraction of a network per prompt, effectively eliminate 90% of parameters. This yields a tenfold increase in compute capacity per watt. It is a highly logical allocation of resources.
• Alternative Hardware Architectures: True structural decoupling requires bypassing traditional von Neumann architectures. Neuromorphic computing utilizes event-driven, analog-digital pathways that mimic the cognitive architecture of organic brains. Simultaneously, optical (photonic) computing leverages photons rather than electrons to execute matrix multiplications. Because photons pass through mediums without resistance, optical computing eliminates structural Joule heating losses, lowering the power-to-compute ratio to a remarkable degree.

Note to self. Why photons instead of elections: An autocorrelation, a concept in matrix multiplication is computationally analogous to a convolution.  A convolution in one domain is multiplication in the Fourier domain. Optics naturally perform Fourier transforms in their interference calculus. Photons are therefore more efficient than electrons at massive volumes of matrix multiplications.

Extraterrestrial Data Centers

Orbital data centers in low Earth orbit have progressed from fiction into actual venture portfolios, capitalizing on unique atmospheric and cosmic advantages. Despite these benefits, orbital clusters will not serve as primary processing hubs for mainstream terrestrial AI demand due to two severe friction points:

1. The speed of light dictates a fixed latency floor on routing data from Earth to orbit and back. This structural lag isolates orbital compute to non-latency-critical workloads, such as deep-space communications, localized edge-processing of satellite telemetry, or highly asynchronous batch training of foundational models.
2. Terrestrial hardware undergoes an aggressive 3-to-5-year obsolescence cycle. Launching massive capital assets into space, only for the underlying accelerators to become obsolete within 48 months, is economically illogical. Until autonomous in-space robotic servicing and modular chip hot-swapping mature, orbital facilities will remain a highly specialized sovereign niche.

Impact on Other Energy Users

The purchasing power of hyperscalers deeply disruptive to the broader merchant power market. Because data centers demand highly reliable, 24/7 baseload power tech companies are executing massive, long-term PPAs effectively sweeping the premium, zero-carbon baseload assets off the market. This creates a dual penalty for industrial and residential consumers:

• The Intermittency Burden: As corporate buyers lock down stable, clean baseload power, public grids are forced to rely on older, highly volatile, or fossil-heavy generation mixes to balance residential demand peaks. (Note to self: the harmonics of the grid requires active grid management to avoid catastrophic volatility in transmission. Consider the Spanish grid blackout 2025.)
• Capital Cost Socialization: Connecting gigawatt-scale data clusters requires extensive transmission grid upgrades, transformer procurements, and advanced substation construction. Utilities frequently socialize these infrastructure outlays, leaving residential and small-business ratepayers with structurally higher utility bills to subsidize the grid modernization demanded by hyperscale operators.

The Power Generation Forecast: Solar/Wind vs. Nuclear/SMRs

The global energy matrix over the next two decades will settle into a co-dependent relationship between variable renewables and localized nuclear power.

The Near-Term Volume Leaders: Solar, Wind, and Hydro

Solar PV and onshore wind will drive the volume of capacity additions through 2040. They have effectively won the LCOE race, a megawatt-hour of solar paired with utility-scale BESS represents the cheapest form of power generation to date. Hydroelectric power remains a critical regional anchor, but is limited by geography and geopolitical water-rights.

However, variable renewables cannot solve the data center bottleneck independently due to compounding land-use constraints, protracted transmission permitting timelines, and the threat of the dunkelflaute (extended periods of zero wind and solar output). Note to self: refer to grid harmonics above.

Small Modular Reactors (SMRs) and Deep Fission

SMRs represent the ultimate structural power solution for the computational age. By shifting nuclear construction from bespoke, multi-billion-dollar civil engineering projects to factory-standardized, modular assemblies, SMRs significantly shorten construction times and bypass traditional financing bottlenecks.

• SMRs are clearing regulatory hurdles in advanced industrial hubs (including the US, China, and parts of Europe). Mainline commercial installations operating behind-the-meter at major data campus sites are projected to scale meaningfully between 2032 and 2035.
• SMRs are not a direct substitute for solar or wind; they are a complementary resource that commands a significant price premium because they deliver a zero-carbon, non-intermittent baseline, allowing hyperscales to operate their clusters at almost full utilization without relying on the public grid.

Commercial Nuclear Fusion

Nuclear fusion remains a post 2050 possibility. Despite significant VC inflows and technical advances in achieving net energy gain (Q > 1) at core level, translating lab net energy into a grid-synchronized, commercial power plant requires solving signifincat materials-science challenges. These include tritium breeding containment and mitigating neutron-resistant structural blanket degradation. Consequently, fusion will not likely play a measurable role in resolving the current generation of AI driven grid constraints. Finally, there are doubts if terrestrial conditions will ever support a practical fusion reactor.

Mitigating Cognitive Effects and Atrophy

Offloading structural cognitive processing to automated algorithmic systems introduces a profound evolutionary risk to human capability. If unchecked, the species faces intellectual degradation. Mitigating this shift requires intentional institutional interventions across education and healthcare.

Remedial Action in Education

Education systems are undergoing a rapid pendulum swing away from total digital integration and back toward raw cognitive resistance training:

• Curriculums are shifting backward to evaluate the process of thought rather than the artifact of production. This includes a return to closed-book, handwritten examinations, viva voce (oral defense) testing, and real-time, unassisted problem-solving.
• Educational frameworks are beginning to classify AI tools not as baseline infrastructure, but as an advanced cognitive privilege, similar to mastering basic arithmetic long before utilizing a graphing calculator. The focus is moving toward training students in “first-principles synthesis,” forcing the human mind to construct internal semantic models before engaging with an external generative interface.

Therapeutic and Psychological Ecosystems

A specialized vertical within the wellness and clinical psychology sectors is emerging to treat the psychological fallout of automated cognitive displacement:

• Noetic Therapy: Clinical frameworks are being actively developed to address “existential obsolescence”, the profound psychological disorientation individuals experience when their core intellectual or creative skill sets are duplicated by an algorithm.
• Cognitive Rehabilitation and “Digital Fasts”: Similar to physical therapy treating muscular atrophy, cognitive wellness centers are commercializing structured programs to rebuild human focus, long-term memory retention, and deep text synthesis. These programs enforce prolonged periods of high-friction analog processing to restimulate synaptic density and dopamine pathway regulation outside of hyper-optimized digital feedback loops.

The Labor Market: Robustness, Vulnerability, and Derailment

The displacement vector of AI is uniquely inverted compared to past industrial revolutions: it is aggressively attacking the upper-middle tiers of the cognitive labour market before automating dexterous physical labor.

Vulnerable Sectors

The most exposed roles are cognitive-routine positions—jobs that involve moving, synthesizing, or translating information between standardized digital systems:

• Mid-Tier Professional Services: Junior corporate lawyers (due diligence, contract generation), entry-level software developers (routine code generation, debugging), financial analysts (data aggregation, standard market modelling), and traditional corporate middle management.
• Administrative and Creative Production: Content marketing, technical translation, routine graphic design, and basic customer operations.

Robust Sectors

Robustness in the modern labor market is determined by two human-centric attributes: hyper-dexterity in high-entropy physical environments, and high-stakes emotional architecture:

• The Precision Trades: Electrical grid technicians, specialized plumbers, HVAC installers, and specialized construction engineers. The physical world is infinitely chaotic, and building a robotic actuator that can navigate a non-standardized crawlspace safely remains orders of magnitude harder than training a trillion-parameter LLM.
• Acute Human Care: Palliative care physicians, surgical nurses, physical therapists, and early-childhood developmental specialists. These roles depend entirely on the bi-directional transfer of high-fidelity human empathy and trust, which cannot be automated without destroying the efficacy of the service.
• High-Stakes Discretionary Capital Allocators: Senior corporate strategists, trust attorneys, and elite fund managers. When a decision involves systemic risk, shifting regulatory landscape navigation, and deep intuition regarding un-quantifiable black swan risks, capital owners will always demand a human throat to throttle.

Derailing or Altering the Displacement Trajectory

The trajectory of mass labour displacement can be fundamentally altered through structural changes to the employment architecture:

• The Mutual-Assurance Regulatory Model: Governments can enforce a framework where AI cannot act as a direct replacement, but must act as a legal co-pilot. For example, maritime aviation did not eliminate pilots; it changed them into system managers. By mandating that high-stakes outputs (legal briefs, medical diagnoses, architectural blueprints) carry strict, non-delegable human liability, public policy can ensure that humans remain legally anchored to the centre of the production loop.
• Taxing the Electronic Worker: If automation accelerates too rapidly for societal absorption, states may deploy “compute taxes” or “automation levies”—tying corporate tax rates directly to the ratio of human payroll to computational energy consumption, artificially slowing the displacement curve to match human retraining cycles.

Demographic Conflicts: Aging Populations and the AI Safety Valve

The convergence of the Fourth Energy Transition with the global demographic winter (sharply falling fertility rates and rapidly aging populations across developed and transition economies) creates a profound macroeconomic paradox. Far from being a crisis, AI is part of the essential structural solution for an aging civilization.

Support Ratios and Fiscal Dependency

In an economy with a collapsing support ratio (where the number of active workers per retiree drops from 4:1 down to 1.5:1), traditional pension systems and fiscal balance sheets face structural insolvency. AI serves as a powerful productivity multiplier, allowing a shrinking pool of young workers to maintain a massive per-capita GDP output. By automating the administrative and routine cognitive overhead of society, a smaller labor force can generate the tax revenues and economic surplus required to fund the fiscal liabilities of an aging demographic.

Looking After the Old

The labour shortage in eldercare is acute. AI and advanced robotics provide a two-pronged solution:

• Administrative De-burdening: AI can automate up to 40% of the bureaucratic and regulatory paperwork that currently consumes the time of nurses and physicians, structurally returning human caregivers to direct patient-facing bedside care.
• Ambient Health Monitoring: Hyperscale, low-power AI models integrated into living spaces can monitor cognitive decline, detect gait degradation to predict fall risks, and manage complex, multi-drug pharmaceutical regimens without requiring continuous, manual human oversight.

Educating the Young and Putting Them to Work

This is the area of highest tension. While AI can act as the ultimate hyper-personalized tutor, democratizing elite, structured, 1:1 education for every child regardless of background, it threatens to destroy the apprenticeship tier of the economy. Historically, the young entered the workforce by doing the routine, low-level cognitive work (summarizing documents, writing basic scripts, sorting data) under the supervision of seniors.

If AI completely displaces this entry-level tier, the ladder of professional development is broken. The young cannot jump from school directly into senior discretionary roles. The primary structural challenge of public policy will be artificially creating or subsidizing “learning positions” within corporations to ensure the continuous transmission of institutional knowledge.

The Public Policy Playbook for the Transition

Sovereign states will be forced to move away from the laissez-faire digital governance models of the early internet era. The physical realities of power grids clashing with the societal shocks of automated cognition require robust, structural public policy intervention.

Grid Sovereignty and Zoning Regimes

Governments will increasingly treat computational capacity as a critical national resource, tightly bound to energy security:

• Algorithmic Curtailment Laws: Public utility commissions will enact regulations that subject hyperscale data centres to mandatory curtailment clauses during periods of acute grid stress. If a heatwave threatens municipal power stability, data centers will be legally required to down-throttle their non-essential inference or training workloads. Basically, this is conditional priority very much like a capital structure for energy.
• Energy-Pairing Mandates: Zoning approvals for new data campuses will be legally contingent upon “Bring Your Own Power” (BYOP) frameworks. Hyperscalers will not be permitted to plug directly into public grids unless they simultaneously co-locate or finance equivalent, dedicated baseload generation capacity (such as co-developing an adjacent SMR or utility-scale battery storage network).

Sovereign Data Hydrology and Anti-Trust

Just as nations established strategic oil reserves in the 20th century, 21st-century states will establish Sovereign Compute and Data Reserves. To protect against the cognitive flattening driven by foreign commercial monopolies, governments will directly fund and maintain localized, open-source foundational models trained on culturally specific, high-integrity regional datasets. This ensures that domestic legal systems, public education, and state administration are not outsourced to the private server architectures of external corporate sovereigns.

Fiscal Re-Alignment: Shifting from Income to Resource Taxation

As the corporate wage bill shrinks relative to computational output traditional fiscal tax systems which rely overwhelmingly on personal income taxes and payroll contributions will face structural deficits.

By structurally taxing the material and energetic inputs of the algorithm rather than human labour, the state may be able to generate the revenues necessary to fund the social safety nets, retraining programs, and public services of a highly automated, low-entropy civilization.

Energy is one of the prime factors of human evolution and civilisation. While we confront the costs of the fossil fuel era, we might consider that this is not the first energy transition in the history of our species.

The first energy transition was perhaps Neolithic Agrarian Transition. Not industrial agriculture but the basic act of planning and organising the cultivation, storage and transport of calories on the small scale.

It took us from the calorific calculus of the hunter, weighing the calorie yield of the hunt against the expenditure of energy obtaining it, to an economy based on delayed consumption and storage. This structural shift fundamentally altered human relationships and social hierarchies. In hunter-gatherer society, mobility acts as a natural leveling mechanism since you can’t own more than you can carry. Agriculture changed the economic calculus by creating storable surpluses (grains like wheat, rice, and barley). The ability to monopolize this surplus led directly to social stratification, creating distinct classes of elites, priests, warriors, and peasants. Also, surplus assets had to be defended or could be stolen, giving birth to organized, large-scale warfare and the construction of fortified cities. Land transitioned from a shared resource into private or state property. Women’s roles were increasingly restricted to high-frequency domestic reproduction and intensive domestic labor. Living in close, permanent proximity to garbage, human waste, and domesticated animals created a perfect storm for pathogens. Foragers ate an incredibly diverse diet of hundreds of species of plants, nuts, insects, and wild game. Farmers became heavily reliant on a single “monocrop” starch. Skeletal remains of early agriculturalists display severe osteoarthritis in the spine, knees, and toes, the direct result of hours spent bending over fields or kneeling to grind grain on stone metates. Foragers had virtually no cavities because their diets were low in simple carbohydrates and sugars. Grains, however, break down into fermentable carbohydrates. When combined with the fine granite grit that wore off stone mills into the flour, early farmers suffered from catastrophic dental decay, abscesses, and premature tooth loss.

The energetic (calorific) properties of agriculture must have seemed myriad and mysterious to the populations in the transition.
Much as we struggle to understand the complex options and interrelations of electric power generation, evacuation and distribution.
Why is the cost of power often zero or negative? Why did the Spanish grid fail? Why does it take so long to obtain approval for a new energy project? What are the relative costs of each type of power when electrons are all the same?
The demand, supply and pricing of power depends on more than science and engineering, the availability of resources or capital, but on regulations that mediate the interests of a host of stakeholders. Producers, users including households and businesses, municipalities and states, conservationists and activists have interests and voices.

Transport by wind at sea was dominant for thousands of years peaking in the late 1800s. This was thermal energy of solar driven convection harnessed by sail. At the same time that sail was peaking, steam driven railways were rising. Sail and rail were the main modes of transporting people and goods until

The last major energy transition was the regime of hydrocarbon energy in the form of crude oil later joined by natural gas. Post WW1, oil helped to propel automobiles, trucks, tanks, planes, and it was post WW2 when it demand skyrocketed. The discovery of massive, cheap, and easily extractable oil fields in the Middle East flooded the market with cheap crude. By the late 20th century, the transition was complete. Civilization was no longer reliant on a single dominant fuel (like wood in the agrarian era or coal in the Victorian era), but on a diversified triad of fossil fuels, split roughly equally between crude oil, nat gas and coal.

The fourth transition, the age of electricity.

In the first energy transition, the Agrarian regime, nearly all energy was directed toward sustaining human and animal life. The vast majority of a civilization’s energy was spent on agriculture (producing calories to fuel muscle power) and heating (burning wood for cooking and basic survival in winter). Construction, transport and information were indirectly supplied and then not in any scale.

In the era of coal and steam, the regime was entirely optimized for manipulating physical matter. Construction exploded. Steam-powered pumps allowed for deep-core mining, and coal-fired blast furnaces allowed for the mass production of cheap iron and structural steel. Steam cranes, steam shovels, and mechanized factories completely transformed the built environment. Steam revolutionized mobility. For the first time, overland transport was decoupled from muscle. The steam railway and steamship compressed global geography, allowing food, raw materials, and armies to move across continents and oceans at unprecedented speeds.

The hydrocarbon regime was about fluidity, speed, and synthetic materials. Oil completely conquered transport. The internal combustion engine powered the automobile, the diesel truck, the container ship, and the airplane. This regime created the modern, hyper-mobile, globalized “just-in-time” supply chain. This era built the concrete-and-asphalt world. Oil powered heavy earth-moving equipment, while natural gas became the essential feedstock for the chemical synthesis of plastics, polymers, and modern cement production. Natural gas and fuel oil became the dominant, automated baseline for residential and industrial HVAC.

This fourth transition, the age of electricity is about the total electrification of material life and information. And we have officially crossed a significant milestone. In the global electricity mix, the combined generation from low-emissions sources (renewables and nuclear) has climbed to 43%. For perspective, in advanced economies and major hubs like the EU, wind and solar have structurally surpassed the total share of electricity generated by fossil fuels. The electrification of the industrial economy is one thing, but the expansion of cloud computing, blockchain networks, and AI data centers requires continuous, high-density, uninterrupted baseload power. We are literally harvesting the sun, wind, and atom to compute abstract data.

Historically, energy transitions are not civilizational upgrades; they are civilizational rewrites. When agriculture replaced foraging, it didn’t just change the diet; it invented the state, private property, and institutionalized hierarchy. When coal-fired steam replaced wood and water, it didn’t just speed up weaving; it created the urban proletariat, global empires, and linear clock-time.

For the first time in human history, civilizational energy demand has decoupled from a linear relationship with the material world. We are transitioning from an era where energy was consumed to manipulate atoms to an era where an accelerating, non-linear share of energy is consumed to organize bits and bytes.

Historically, industrial energy demand grew in a largely linear fashion, bound tightly to demographics and physical limits. To transport twice as much cargo overland via steam railway or diesel truck, you needed roughly twice as many tracks, vehicles, and fuel. To heat a city, your energy consumption scaled proportionally with the number of households and ambient winter temperatures.

Even the current electrification of the material economy, converting ICE vehicles to EVs, or domestic gas furnaces to electric heat pumps, follows this predictable, linear trajectory. There is a firm physical ceiling to how many kilometres a population can drive, how many tons of steel a society needs to smelt, and how many calories a human body must consume.

Informational energy demand, by contrast, operates under no such physical caps. It is driven by the insatiable, self-reinforcing loops of hyperscale computing, LLMs, GPTs and RSIs. AI models are utilized to generate synthetic data, which is then used to train larger, more resource-intensive next-generation models. The process is entirely software-driven, meaning it can scale at a velocity that completely detaches from human demographic growth. The better our models, the more efficient they are, the more we demand of them and their energy appetites increase. Compute demand expands exponentially to fill all available capacity.

The scale and speed of this transition makes extrapolation and inference challenging? Who knows what the future will look like, if trajectories are non-linear but exponentially and dimensionally generative. But let us focus on one dimension: cost.

Through the transitions, the matter of costs is of interest. In the hunter gatherer regime, part of the cost was the calories invested to obtain more calories, In the Agrarian it was calories invested to calories harvested. In the coal and steam era it was the cost of extraction. The same applies to the era of hydrocarbons. Over these transitions, up front or coincident costs have been easy to identify if not to measure or quantify. However, deferred costs have been less easy to identify or quantify, especially if these costs were indirect and non-monetary, I refer of course to the environmental costs which we can proxy with carbon dioxide and other GHGs. What other costs might there have been in the early transitions? Did sedentary comforts make us weaker or more sickly or less resilient? Will the age of AI make us dumber? What other deferred costs might we have missed and when and how might they come to haunt us?

In the Agrarian regime, the deferred costs included skeletal and structural weakening: Upper Paleolithic hunter-gatherer skeletons present bone densities matching modern ultra-marathon runners. Their long bones were thick, straight, and virtually free of degenerative disease. Agrarian skeletons show widespread cortical thinning (bone loss), joint degeneration, and shorter statures. Sedentism meant living in permanent proximity to human feaces, contaminated water, and domesticated livestock. This created the First Epidemiological Transition: a massive, deferred explosion of zoonotic diseases (smallpox, influenza, measles) that systematically weakened the human immune baseline. The Nutritional Single-Point Failure: Foragers ate hundreds of varied species, making them hyper-resilient to localized ecological shocks. If one plant died, they moved to another. Farmers hitched their entire survival to a single monocrop (wheat, rice, or maize). When the weather failed, the deferred cost was catastrophic, system-wide famine. The Forfeiture of Autonomy: Foragers maintained highly egalitarian social structures. If a leader became tyrannical, the group simply walked away. Agriculture anchored people to fixed assets (cleared fields, irrigation canals). You could no longer leave. This spatial trap forced humanity to accept the deferred social cost of permanent hierarchy, institutionalized slavery, and the surrender of personal autonomy to the early bureaucratic state.

The hydrocarbon regime brought is own costs. The Invention of Chronic Stress: Foragers and early farmers worked intensely but intermittently, punctuated by long periods of rest, storytelling, and social bonding. The industrial engine required continuous, unyielding output. The deferred cost was the enforcement of clock-time, disciplining the human nervous system to operate like a mechanical piston. This severed our connection to natural circadian rhythms, inventing modern sleep disorders, chronic anxiety, and industrial fatigue. While greenhouse gases are the headline environmental proxy, the immediate deferred costs were local and visceral: the lead poisoning of urban soils, the destruction of aquatic ecosystems via industrial runoff, and the inhalation of particulate matter that permanently altered human respiratory health centuries before global warming was quantified.

What about this Fourth Transition? I cannot see, I wonder who can. Will AI make us dumber? From an evolutionary perspective, the risk of cognitive atrophy is immense. Human intelligence is a “use-it-or-lose-it” system. When we outsource navigation (GPS), memory (search engines), and synthesis, analysis, and critical writing (generative AI), we are systematically stripping the brain of the cognitive resistance training required to build deep neural pathways. If AI handles the heavy lifting of processing information, humans risk becoming mere “attention managers” or syntax-checkers. We lose the capacity for deep, sustained focus and the serendipitous cross-pollination of ideas that occurs when a human brain struggles with complex problems.

Then there is a more fundamental trade off. Efficiency X Robustness = a Constant. Here we think of society and economy as systems for processing information and turning them into decisions which become information for other and subsequent agents in the system. An efficient system minimises the cost of decisions while a robust system minimises the cost of errors and suboptimal outcomes.

Large language models do not think; they calculate probabilities. By definition, AI output is an exercise in convergence. We can see two consequences. The Loss of Cognitive Mutagens: Human progress relies on eccentricities, weird leaps of logic, grammatical errors, localized slang, and radical, low-probability insights, the “noise” or entropy of human consciousness. AI is basically a mathematical filter sorting out noise from signal, returning one and discarding the other. A Syntactic Monoculture: We are rapidly building a world which looks and feels exactly the same. We are replacing a high-entropy, diverse global tapestry of human expression with a hyper-polished, low-entropy corporate monoculture.

By standardizing our global cognitive network under a unified algorithmic regime, we risk creating a civilizational single-point-of-failure. If the core models have a blind spot, or if their optimized logic fails to account for a sudden real-world mutation, the entire globalized matrix risks simultaneous failure. We will have traded our diverse, messy, chaotic resilience for a hyper-efficient, fragile digital monoculture that lacks the evolutionary toolkit to save itself when the environment shifts.

In the next article, we will look at some practicalities.

The standard model of economic behaviour starts with a simple assumption: people act in their own interest. From this one seed, an entire architecture of markets, prices, and efficiency grows. Adam Smith’s great insight was that private selfishness, channelled through competition, could produce public good. The butcher and the baker serve us not out of kindness but self-regard, and yet we are fed.

But taken too literally, this framing leads somewhere sterile. It describes a world of strangers who transact once and walk away. A world of single-shot games.

In a single-shot game, your counterparty’s interests are irrelevant to your strategy. In a constant-sum game, they are worse: they are opposed. What you gain, they lose.

Life, however, is not a single-shot game. Business relationships recur. Reputations compound. Communities persist across decades. The moment you introduce repetition, the whole calculus shifts. Life is symbiosis.

In repeated games, your counterparty’s interests start to matter, not as some moral concession but as a logically strategic one. Cooperation that would be irrational in a single encounter becomes entirely rational across a long enough horizon. This is what the Folk Theorem tells us: given sufficient patience, cooperative outcomes can be sustained. Go look up Folk Theorems and Prisoner’s Dilemma.

There is something beyond strategic patience at work in human cooperation. Game theory in its classical form treats each player as an opaque utility maximiser. Real human behaviour adds the possibility that we are not opaque to one another, and that our interests are not as independent as the model assumes. Go look up the game of dividing the dollar.

Adam Smith understood this before the formal mathematics existed to describe it. The Theory of Moral Sentiments, which he wrote before the Wealth of Nations and which forms its basis, is an investigation into exactly this: how social creatures come to share in each other’s experience. Two things matter here.

The first is empathy, which in its cognitive sense is the capacity to genuinely model another person’s state of mind, beliefs, and situation. The empathic player does not merely observe the other; they inhabit the other’s perspective well enough to anticipate it. This is not sentimental but logical, as it aids convergence to agreement and transaction.

The second is sympathy, which runs deeper: the positive interdependence of what we actually care about. The sympathetic person does not just model another’s welfare; they share in it. Their own satisfaction is partly made up of the flourishing of those around them.

Together, these two capacities explain how the behavioural foundations of the Wealth of Nations are actually met. Markets work not because people are purely self-interested, but because self-interest is embedded in a web of social feeling that makes the relevant games long and non-zero-sum.

So what does any of this mean for investing?

It means the conventional framing, which treats commercial investing and impact investing as philosophically distinct activities, is a bit weak. The narrowly conceived commercial investor is playing a single-shot game in a constant-sum world. The impact investor is someone who has understood that the real game is repeated, that it is not constant-sum, and that the returns worth caring about extend well beyond the financial.

This is not altruism. It is enlightened self-interest.

The person who subordinates narrow personal interest to the interests of the collective is not abandoning self-interest. They are recognising that their interests, properly understood, are bound to the system they inhabit. An individual that depletes the social and environmental commons from which it draws its wealth is engaged in a slow act of self-destruction, however healthy the financial returns look in the meantime.

When impact investments yield below market returns, something real has happened. The shortfall is not a loss to be explained away. It is the price paid for other return streams: environmental stability, social cohesion, the health of the communities and ecosystems the portfolio depends on. These are returns. They are less legible than an IRR, but they are no less real, and across a long horizon, they may matter more than anything else.

Multiple return streams, are what serious long-term investing obtains. Financial returns are one stream. Environmental returns are another. Social harmony, institutional trust, and the preservation of conditions in which future generations can do well are others still. The work is to hold all of these in view at once, to make trade-offs consciously, and to resist the temptation to collapse a rich problem into a single number.