The Capture Gap

EXECUTIVE SUMMARY

The argument in brief

Investment negotiations still open with the question that has governed economic development for seventy years: how many jobs? On the projects now defining the investment cycle, that question no longer produces a usable answer. This report explains why the failure is structural rather than statistical, and what should replace it.

01 · The question that stopped working

The direct-jobs number can no longer carry the deal — and the softer phrases that replace it are a symptom, not an answer.

Investment negotiations still begin with a question that has governed the practice of economic development for most of its modern life. A company arrives with a project. The government asks what the place will receive in return. The answer is given in jobs.

Lately, the answer has become harder to sustain.

Consider the project now landing on desks in every energy-rich jurisdiction: the data centre. Its capital value may exceed anything the region has previously attracted. Construction is substantial. Once operating, however, the facility may employ only dozens or a few hundred people while requiring extraordinary quantities of electricity, along with land, water, transmission capacity and, often, tax concessions. Data centres can produce wider employment and supplier effects, particularly when several hyperscale facilities form a genuine technology cluster, but the effects vary sharply by project type and are routinely overstated when pre-existing regional growth is ignored. Meanwhile the physical demands are not theoretical: the International Energy Agency estimates that global data-centre investment reached roughly half a trillion dollars in 2024 and that the sector’s electricity consumption could more than double by 2030.

Everyone around the negotiating table knows that the direct-jobs number cannot carry a deal like this. So the conversation moves into softer territory — ecosystem effects, signalling value, technological leadership, putting the region on the map. These may be real. They are also the phrases governments reach for when an old metric has stopped answering the question and no accepted replacement exists.

It would be reassuring to call this a measurement problem: find a more sophisticated multiplier, update the incentive model, carry on. The problem runs deeper. The jobs metric is weakening because the economic relationships beneath it are weakening. Modern economic development was built for an era in which productive capability was scarce, employment was one of its principal channels of diffusion, and wages were the main route by which production became broadly distributed purchasing power. Artificial intelligence does not abolish any of these relationships. It makes each of them less reliable — and it does so simultaneously.

The argument here is not that learning has become irrelevant, that expertise is now universal, or that machines are about to perform all economically useful work. It is that economic development can no longer assume that accumulating capability will automatically produce employment, or that hosting production will automatically produce locally retained prosperity. Capability remains necessary. It is no longer sufficient.

The next theory of economic development must place a second function beside learning: the deliberate construction of claims on the value an economy helps create.

The distance between value produced in a place and value retained there is the capture gap. Closing it is becoming the profession’s central task.

EXHIBIT 01 · THE PROJECT OF THE MOMENT

The Sturgeon Data Centre — Sturgeon County, Alberta

Announced July 8, 2026 · Meta’s first data centre in Canada, and one of the largest private-sector investments in Canadian history

C$13B+

total capital investment in a one-gigawatt, AI-optimized campus on 1,750 acres in Alberta’s Industrial Heartland

≈ 3,000

construction workers on site at the peak of the build

300+

permanent jobs once the facility is operational

≈ C$43M

capital investment per permanent operating job

C$250M/yr

projected annual provincial benefit from taxes, royalties, levies and fees, per the Government of Alberta

C$60M

Meta’s committed spending on local roads and water infrastructure

The campus will be powered by new gas-fired generation planned alongside the project, including the C$4.6-billion, 970 MW Greenlight facility. Meta states that it pays the full cost of its energy use and is funding new generation and grid infrastructure so that other consumers are not negatively affected. Sources: Meta; Government of Alberta; CBC News; The Globe and Mail (see Sources and method).

02 · The theory the profession inherited

Postwar economics went looking for the scarce factor behind growth and found it in knowhow. From that finding came the profession’s entire toolkit — including the jobs metric itself.

Economic development rarely announces its intellectual debts, yet nearly every instrument in its standard toolkit rests on an argument about why places become prosperous — and the argument can be traced.

Much of postwar development economics was a long search for the scarce factor behind growth. Early theories emphasized capital: poor countries lacked machinery, infrastructure and finance, and investment would set accumulation in motion. Robert Solow then showed that capital deepening accounted for only part of long-run growth; the large unexplained residual was labelled technological progress and left, for a time, as a black box. Paul Romer moved ideas inside the model, where knowledge could accumulate, spread and generate increasing returns. Institutional economists showed why the same technology produced different results under different legal and political systems. And the economic-complexity tradition supplied the field’s most elegant finding: rich economies did not merely own more machines or resources — they carried a greater volume of productive knowhow, distributed across workers, firms and institutions.

That knowhow was difficult to copy because much of it was tacit. A blueprint could cross a border; the organizational competence required to turn it into a reliable factory usually could not. Capability lived in habits, supplier relationships, engineering judgment and thousands of people who each knew a different part of a system.

From this proposition came the operating logic of the profession. If prosperity rested on accumulated capability, development was fundamentally a learning process: attract firms that possess knowhow, connect them to domestic suppliers, train workers, absorb techniques, move into adjacent products, export, repeat. The special economic zone, the supplier-development programme, the technical institute, the innovation district and the investment-promotion agency all belong to this tradition. So does the jobs metric itself. Employment mattered as income, but it mattered equally as transmission — the worker learned to run the equipment, meet international standards, manage a line, and eventually carried that knowledge elsewhere.

A payroll was a curriculum.

The framework also carried a moral promise that became the profession’s founding faith: geography was not destiny. A place without minerals, oil or a large domestic market could still learn its way into prosperity. Korea had little energy; Singapore had no hinterland; it did not matter, because the decisive resource was capability, and capability could be accumulated.

That remains one of the most important truths in development history, and it is not being repealed outright. It is being qualified — by a technology that changes which kinds of capability are scarce, how they move, and who owns the systems through which they are now supplied.

03 · What AI changes, and what it does not

The claim is not that machines now hold what practitioners know. It is that enough cognitive work is being commoditized to move the returns elsewhere.

Every major technology embeds knowledge. A machine tool contains metallurgy; enterprise software contains accounting and process design; the internet made recorded information available almost anywhere. None of these eliminated the value of local competence, because owning the machine did not confer the ability to maintain it, and reading the instructions did not reproduce the judgment of an experienced operator.

Artificial intelligence goes further, because its product is a portion of the cognitive work itself. It can draft, translate, classify, summarize, diagnose patterns, generate code and assist with design. It does not possess human tacit knowledge in any complete sense — much of its training material consists of recorded outputs that reflect expertise rather than the situated experience behind it — and it remains unreliable in unfamiliar, ambiguous and high-consequence settings. Responsibility, physical practice, institutional memory and judgment under uncertainty do not disappear because a model produces a fluent answer.

But AI does not need to reproduce every dimension of expertise to alter its economics. It needs only to make a material share of useful cognitive production cheaper, faster and more widely available — and that threshold has already been crossed. The capability that the complexity literature treated as the stickiest asset in economics is, in growing part, being priced, packaged and sold as a subscription, by roughly five firms.

The evidence so far shows both sides of the process. AI tools can raise productivity substantially in bounded tasks, often helping less-experienced workers most; in complex settings, the gains depend on domain expertise, data quality, management and oversight. Some firms are cutting junior roles; others are hiring and expanding output. A recent World Bank study of digitally delivered services found firms in the Philippines using AI for training, coding and process automation — only a minority reported job losses, and some reported gains — while entry-level roles grew scarcer and postings in exposed IT and business-process occupations fell sharply from late-2022 levels. This is not a simple replacement story. It is a story of task recomposition, productivity gains, and a narrowing route into professional work.

The trajectory is best separated into three horizons. In the present one, AI commoditizes portions of codified, verifiable cognitive work — drafting, routine analysis, translation, software production, customer interaction. In the next, more reliable agents may coordinate longer sequences of work across enterprise systems; the speed is uncertain, the direction less so. A third horizon — systems matching human performance across most domains — remains speculative, and no development strategy should depend on its arrival.

It does not need to. Even partial commoditization is enough to change the field, because if a capability that once took years to build can increasingly be rented, its mere possession becomes less distinctive, and the returns migrate toward the assets needed to apply, govern and combine it. Access, after all, is not adoption: a firm in Lagos and a firm in Boston may subscribe to the same model, but they do not thereby acquire the same electricity system, data, financing, institutions or customers. The IMF’s AI Preparedness Index is built around exactly these differences, and wealthier economies currently score better on most of them.

Knowhow is not becoming universal. Generic cognition is becoming cheap while the complements around it become more valuable.

The learning problem has not vanished — its object has changed.

04 · Scarcity after cognition

Electricity, context, jurisdiction, organizational capability, capital, clusters — the inventory of advantage once cognition is cheap.

When one input becomes cheap, advantage does not disappear; it migrates toward whatever remains scarce, or whatever makes the cheap input productive. In the intelligence economy, that list is broader than the traditional inventory of skills and technology.

4.1

Physical substrate

Intelligence, it turns out, is made of electricity. It requires generation, grids, fibre, cooling, land, water, semiconductors and minerals — none of them weightless, many of them slow to build. The significance is often local rather than global: data centres consumed only about 1.5 percent of world electricity in 2024, but capacity concentrates in a small number of regional clusters, so a facility that is trivial in the global balance can be transformative — or destabilizing — within one grid. Transmission takes years to permit; transformers and turbines face long order books. The development question is never simply whether a project consumes power. It is what that power could otherwise have enabled, who pays to expand the system, and whether the project widens or narrows the region’s future options.

4.2

Context

General models are built from broad patterns; economic value is usually created in particulars. A mining district carries decades of geological records and operating knowledge. A health system holds longitudinal records and clinical pathways. A regulator contains years of precedent never assembled in usable form. This place-bound information is the one form of knowhow the commoditization cannot reach, because it was never written down anywhere the models could read. It is not automatically valuable — most of it is fragmented, legally restricted, or carried by employees approaching retirement — and converting it into an asset requires unglamorous work: digitization, standards, security, consent, access rules. That work may prove more valuable than another generic AI centre. A model can be purchased. A well-governed body of local industrial or public-service knowledge cannot be recreated quickly by anyone.

4.3

Jurisdiction

Governments possess something technology firms cannot manufacture: the authority to permit, certify, inspect, tax and legitimate, and access to regulated environments where products can be tested on real grids, roads, farms and patients. Exercised badly, this authority destroys value. Exercised competently, it creates value by reducing uncertainty and letting technologies move from demonstration to deployment. Jurisdiction is not merely a restriction on production. It is an input into it.

4.4

Organizational capability

Cheap intelligence does not implement itself. Firms must redesign workflows, clean data, integrate systems, train workers and verify outputs, and the World Bank and IMF keep finding large differences in returns arising from management, data and institutions. This is where the strongest version of the old theory survives: capability now includes knowing which tasks to automate, how to reorganize production around new tools, and how to retain what is learned during implementation.

4.5

Capital and risk-bearing

AI lowers the cost of some cognitive inputs while raising the scale of capital required elsewhere — compute, energy, advanced manufacturing, biotechnology. Places able to finance long-duration investment, absorb failure and maintain policy continuity hold an advantage no model subscription equalizes.

4.6

Clusters and market access

Production is rarely the sum of isolated inputs. A data centre built alone may produce modest spillovers; several hyperscale facilities connected to fibre, energy investment, universities and domestic firms may create something different. Ecosystem effects exist — but they cannot be presumed from the project category. They must be designed, contracted and measured. Trusted brands, procurement relationships, language and access to paying markets remain advantages of the same kind.

The emerging geography is therefore not a return to crude resource determinism. It is a geography of combinations: power plus institutions, data plus governance, capital plus execution, jurisdiction plus trust. Places cannot choose their geology, but they decide whether an endowment becomes a raw input sold cheaply, the base of a domestic industry, or a permanently owned financial asset.

Geography matters again. Institutions decide what it means.

05 · Two ladders under pressure

The consequences are most serious for economies that relied on the established routes into world markets.

The classic ladder began with labour-intensive production: workers moved from agriculture into garments, assembly and basic manufacturing; firms gained experience; suppliers emerged; countries climbed into more sophisticated products. Automation has been eroding the bottom rungs unevenly — robots do not sew every garment or harvest every crop, and low wages still matter in many industries — but the labour required per unit of output has fallen across much of manufacturing, and new factories often arrive with fewer entry-level jobs than their predecessors.

A second ladder emerged through tradable services. Call centres, software, back-office administration and business-process outsourcing let countries export cognitive labour before they possessed deep industrial supply chains. India and the Philippines were the most visible climbers, but the model spread widely. AI bears directly on the standardized, remotely delivered, verifiable tasks on which that ladder’s lower rungs depend — which is to say, on exactly the work the systems perform best.

The effect will not be uniform, and the Philippines illustrates both possibilities at once. Its digitally delivered services sector employs roughly 1.8 million people and has built real capability over decades; it is not an industry that vanishes when a new model ships. A contact centre may use AI to raise agent productivity and win contracts; a software firm may expand because its engineers produce more. Yet the World Bank’s research also finds that foreign ownership limits domestic control and reinvestment, and that AI tools can shift a growing share of revenue to software providers outside the country even while Philippine firms remain competitive.

That is the development problem in miniature. Output rises. Exports hold. Productivity improves. And the share of value retained domestically can still fall. The question is no longer only whether an economy can participate in a growing industry. It is where, within the industry’s changing chain of value, its claims reside.

06 · The circuit that was never automatic

The wage circuit depended on institutions, and it was fraying before AI arrived. What is new is that labour costs are now a target in cognitive production too.

Industrial capitalism created a powerful transmission mechanism: production required workers, workers received wages, wages supported the demand that justified the production. Mass production and mass consumption reinforced each other.

The mechanism was never universal or self-executing. It depended on institutions — collective bargaining, public education, labour standards, taxation, social insurance, competition policy — and many economies industrialized under severe inequality, with export demand, state spending or household debt substituting for broadly shared wages. Still, labour’s indispensability gave workers and governments leverage, and the wage bill remained one of the main routes through which output entered household income.

That route has already grown less dependable. OECD evidence shows productivity and real labour income diverging in a significant group of countries, with median pay lagging average pay. The causes are contested and extend well beyond technology — weakened bargaining institutions, globalization, market concentration, cheaper capital goods. AI arrives inside this existing fracture rather than creating it. Its significance is that it targets labour costs in cognitive as well as physical production, which means a region can now host rising investment, output and productivity without receiving an equivalent rise in household income. The value is still created. It simply reaches residents through a narrower wage channel, or not at all.

Economic-development practice has treated local value capture as a consequence of successful production: attract the project, create the jobs, and prosperity follows. The intelligence era requires capture to be designed as an outcome in its own right.

07 · From capability to capture

Two functions, distinct and interdependent: capability formation and value capture. Either one alone fails — in opposite directions.

The new framework does not replace learning with ownership. It joins them. Economic development now has two distinct, interdependent functions.

The first is capability formation : enabling people, firms and institutions to adopt technology, improve production, create new businesses, and bargain from a position of competence.

The second is value capture : ensuring that a meaningful share of the gains generated through local assets, institutions, labour and public support remains with the population over time.

Capability without capture produces an efficient enclave — a jurisdiction that supplies the land, power, workers, data and permission while profits, intellectual property and strategic control accumulate elsewhere. Capture without capability produces a rentier economy — a government that collects revenue but fails to build productive firms or adaptable institutions. The task is to combine them, and the combination requires a sharper definition of what it means for an investment to contribute to development at all. Gross capital expenditure is not enough; neither are direct jobs or corporate announcements. The relevant question is how much durable value the project leaves behind after its full public and economic costs are counted.

THE CENTRAL CONCEPT

That is the capture gap — the distance between the value an activity produces in a place and the value that place retains.

08 · Measuring the capture gap

Not a single immaculate number — a four-column ledger that forces the negotiation to include everything the jobs count excludes.

AT ITS SIMPLEST

Capture gap = total value attributable to an activity in a place − total value retained by that place Local capture rate = locally retained value ÷ total attributable value

Neither number will be perfectly objective. Value crosses borders through supply chains, transfer pricing and intellectual property; opportunity costs depend on counterfactuals; some benefits — resilience, reputation, option value — resist pricing. But jobs estimates suffer from most of the same problems, and governments nevertheless spent decades building conventions for employment multipliers and fiscal impact. Capture accounting deserves the same seriousness: common definitions, stated assumptions, independent review, and reporting after the project is operating rather than only at announcement.

A practical capture ledger has four columns.

THE LEDGER WHAT IT COUNTS

Value retained Wages and benefits received by residents; profits accruing to locally owned firms and equity holders; taxes, royalties and revenue participation; margins earned by domestic suppliers; infrastructure transferred to public or shared use; intellectual property and data rights retained locally; funded research, training and implementation capability; assets accumulated in public, pension, Indigenous or community funds; and strategic options created for future industries.

Public and economic costs Grants, exemptions and below-market financing; publicly funded land, roads and grid upgrades; the opportunity cost of electricity, water, labour and industrial land; system costs imposed on other users; environmental liabilities; guarantees and contingent fiscal risks; displaced investment; and the cost of technological or commercial lock-in.

Distribution Traces retained value to its recipients. A project may generate a strong national fiscal return while leaving the host community worse off; domestic ownership may concentrate in a small

group; supplier contracts may flow to firms elsewhere in the country. A capture rate without a distributional account can disguise the very problem it was built to expose.

Duration Assesses retained value over time. Construction wages are immediate; grid commitments last decades; tax holidays expire; public equity can pay for generations. A project attractive over five years may be poor over twenty, or the reverse.

The purpose is not a single immaculate number. It is to force the negotiation to include everything the jobs count excludes.

09 · What the framework changes in a data-centre deal

Same capital value, similar operator headcount, opposite development outcomes — and three live projects that show the pattern is no longer hypothetical.

Return to the opening negotiation. Two proposed data centres carry the same capital value and employ a similar number of operators. Under the old framework they are nearly identical. They may be entirely different development propositions.

DATA CENTRE A

The hosted project

DATA CENTRE B

The negotiated project

Receives a long property-tax exemption. Pays the full incremental cost of generation and connection.

The utility finances new transmission and spreads the cost across ratepayers. Its power agreement protects existing customers and permits demand flexibility during system stress.

Secures a large block of firm power with no curtailment obligations. A domestic pension fund or public investment vehicle holds a minority equity stake on commercial terms.

Equipment and specialist services are imported; waste heat is vented. Heat supplies a nearby industrial or district-energy user; local universities and firms receive a defined compute allocation.

No local compute access; no enforceable commitment to research, suppliers or shared infrastructure. Local suppliers receive support to qualify for contracts, and the obligations are enforceable rather than aspirational.

The region hosts the project — construction activity, a modest operating payroll, whatever taxes survive the exemption — and surrenders scarce grid capacity while assuming part of the infrastructure risk. The region holds claims on the project — revenue, equity, compute, heat, supplier margins — that outlast the announcement cycle.

The direct-jobs count barely differs. The capture rate is not close.

◆ The pattern in the wild

None of this is hypothetical. In July 2026, Alberta landed one of the largest private-sector investments in Canadian history: Meta’s C$13-billion, one-gigawatt Sturgeon Data Centre, profiled in Exhibit 01. Under the jobs framework the evaluation is quick and unflattering — roughly C$43 million of capital per permanent position. The announced terms are more interesting than the jobs count. Meta states that it pays the full cost of its energy use and is funding the new generation and grid infrastructure its campus requires; the provincial government projects at least C$250 million a year in taxes, royalties, levies and fees; the company is spending C$60 million on local roads and water. Several of the negotiated project’s conditions appear, at least on paper. What the public record does not yet show is the rest of the ledger: how that annual figure was modelled and what makes it enforceable; whether any local institution gains compute access, data rights or supplier positions it could build on; and what a gigawatt of firm demand forecloses in a grid that other industries also want to enter. Under the jobs framework, the deal is finished being evaluated. Under a capture framework, the evaluation has just begun.

In Texas, the flagship Stargate campus in Abilene — the most-watched AI infrastructure project in the United States — had more than 6,000 construction workers on site at its peak. Its permanent-employment figure depends on who is counting: published estimates run from roughly 300 jobs to nearly 1,700. When the central metric of a negotiation can vary by a factor of five across accounts of the same campus, the problem is not sloppy journalism. The metric itself has stopped doing the work.

The contrast case sits in Windsor, Ontario. The NextStar Energy battery plant represents more than C$5 billion of investment and a target of 2,500 permanent jobs at full production — roughly C$2 million of capital per job, some twenty times the employment density of the Alberta campus. It is also a caution against treating employment density as the whole answer. In February 2026, Stellantis sold its 49 percent stake in the venture to LG Energy Solution, and the plant’s output was redirected toward grid-storage batteries. The jobs stayed; the ownership, direction and upside moved. Even the most jobs-rich project on this page ends up arguing for a ledger that counts more than jobs.

EXHIBIT 02 · CAPITAL PER PERMANENT JOB

Same headline number, different development propositions

Announced capital investment divided by announced permanent jobs, two Canadian megaprojects

Meta Sturgeon Data Centre, Alberta · C$13B · 300+ permanent jobs

NextStar Energy battery plant, Ontario · C$5B+ · 2,500 jobs at full production

A twenty-fold gap in employment density between two flagship projects — and neither ratio, on its own, settles whether the deal was good. That is the ledger’s job. Sources: Meta; Government of Alberta; NextStar Energy; LG Energy Solution (see Sources and method).

This does not mean every jurisdiction should impose every condition — a weak market cannot negotiate as though it were the only viable site, and excessive or unpredictable demands deter investment. Nor are all data centres sterile: recent research finds that hyperscale clusters can support meaningful employment in construction, information services, fibre and related firms, while stand-alone colocation projects produce much weaker spillovers — and that siting decisions are driven more by power, land and fibre than by tax incentives, which is itself an argument for giving fewer of them. That is exactly why the ledger matters. It replaces categorical arguments — data centres are transformative, data centres are useless — with a project-specific account of what the jurisdiction contributes, what it receives, and what it gives up.

≈ C$43M / job

≈ C$2M / job

10 · The architecture of claims

Four families of instruments — fiscal, ownership, contractual, institutional — broader than royalties, and older than they look.

A claim is an enforceable entitlement to a share of economic value, one that need not depend on a person supplying labour at the moment the value is produced. The available instruments are broader than conventional royalties, and they fall into four families.

Fiscal claims

Taxes, royalties, congestion charges, land-value capture, windfall mechanisms, payments for scarce public inputs. The objective is not to maximize the rate but to distinguish ordinary returns, which reward risk and capital, from rents that arise because a company has obtained privileged access to a scarce local asset. A megawatt in a constrained grid, a licence in a protected market or exclusive use of a public dataset can carry economic rent even when no mineral is extracted.

Ownership claims

Public equity, community and Indigenous participation, pension-fund co-investment, sovereign and regional wealth funds. Ownership is not appropriate for every project, and governments are not automatically skilled investors; political interference and concentrated risk can destroy value. Properly structured, however, ownership converts a one-time transaction into a continuing balance-sheet asset. Norway built its Government Pension Fund Global to turn petroleum revenue into permanent financial wealth; Alaska converted part of its non-renewable endowment into a fund whose earnings pay resident dividends. The lesson is not that every place should copy either model. It is that temporary endowments can become durable, broadly held claims — when someone designs the vehicle.

Contractual claims

Negotiated directly into project agreements: revenue participation, infrastructure contributions, compute allocations, research funding, supplier-development and training obligations, performance guarantees, and rights that activate when public costs exceed agreed thresholds. The governing principle is enforceability. “Community partnership” and “ecosystem development” are not claims. A defined obligation with a timetable, a reporting requirement and a remedy for non-performance is.

Institutional claims

The rules under which value can be generated and transferred: rights over data, standards, certification, intellectual property, testing environments, interoperability, and the conditions attached to procurement. A government that digitizes public information and grants unrestricted access may enable innovation while surrendering an asset; one that forbids all use preserves control while producing nothing. In some sectors these claims will matter more than tax revenue — the right to a domestic standard, to an adapted model, or to data improvements flowing back into the public system can shape an industry long after any incentive expires.

11 · Two objections

Isn’t this the resource curse in new clothes? And won’t capital simply go somewhere cheaper? Straight answers to both.

◆ Isn’t this just the resource curse with a broader vocabulary?

Rentier states are development economics’ most familiar failures: resource income weakens taxation, hollows institutions, feeds patronage and leaves the economy exposed to a single source of wealth. If governments start treating electricity, data and market access as rents to extract, why expect anything different?

Because claims are not a substitute for production but a means of ensuring production leaves a durable local stake — an economy cannot negotiate its way to prosperity without also forming firms, workers and institutions capable of generating new value. Because the relevant comparison has changed: the resource-curse literature contrasted rent dependence with industrialization through employment, and where labour-intensive ladders remain climbable they should be climbed — but when the wage channel weakens, the real choice is not between rents and an undisturbed learning economy. It is between transparent public claims and value that leaves without them. And because the difference between a claims state and a rentier state is, in the end, governance: rents corrode when they flow through opaque channels to narrow coalitions, and rules-based fiscal systems, independently governed funds, published contracts and broad ownership change that political economy without abolishing its risks. One institution converts scarce assets into diversified public wealth while continuing to build capability. The other consumes proceeds and protects incumbents. They are not the same institution wearing different clothes.

◆ Won’t capital simply go somewhere cheaper?

The second objection is practical: capital is mobile, and a jurisdiction that demands too much loses the project to a competitor that demands less. True — and any serious framework starts there. Not every place has bargaining power; some projects have dozens of viable sites, others require a particular grid, geology or regulatory environment. Claims must be priced, not proclaimed. That means identifying the investor’s genuine alternatives, valuing the location-specific inputs, determining the minimum public return that justifies committing them, and structuring obligations so the project remains commercially viable. Sometimes the right answer is to offer support, because the project creates capability the market would underprovide. Sometimes it is to approve without incentives. Sometimes it is to demand more. Sometimes it is to decline — a project that fails one jurisdiction’s required return may still be an excellent project for somewhere else. The framework is not anti-investment. It is opposed to the idea that investment volume is itself the objective.

12 · Capability as the condition of capture

A government that cannot model twenty years of electricity demand cannot price a data-centre connection. Capability is now the infrastructure of bargaining.

Nothing in this argument diminishes learning. It makes capability more demanding, because capability now performs three functions instead of one.

It remains a source of advantage where knowledge resists codification, where physical execution matters, and where local context changes the answer. It determines whether firms can use rented intelligence productively — a company that cannot redesign its processes will not be rescued by better software. And it determines whether a jurisdiction can negotiate and administer claims at all. A government that cannot model twenty years of electricity demand cannot price a data-centre connection. A health authority that cannot govern its data cannot set credible access terms. A development agency that cannot distinguish a hyperscale cluster from a stand-alone server shed cannot evaluate an ecosystem promise. Capability has not ceased to differentiate places. It has become the infrastructure of bargaining, diffusion and stewardship.

This also changes the preferred model of technology adoption. Training people on generic tools is useful but insufficient; firms need implementation support inside their own operations — process mapping, data preparation, integration, change management, measured results. The better analogy than the one-off AI seminar is the old institution of agricultural and industrial extension: specialists working alongside firms over time, with the lessons from one implementation becoming templates for the next, so that vendors and workers accumulate experience and the local cost of adoption falls.

The practical test is simple. Was the second implementation easier than the first? If not, the economy consumed a technology. It did not accumulate capability from it.

13 · The institution this requires

Five roles under one roof — attraction, valuation, negotiation, diffusion, stewardship — and an operating rhythm built around the ledger rather than the announcement.

The archetypal development institution of the learning era was promotional and pedagogical: it marketed the jurisdiction, recruited investment, trained workers, connected firms to programmes. Those functions remain. They are no longer enough.

The institution the claims era requires combines five roles:

01 Attraction — to bring productive firms, capital and market access into the economy.

02 Valuation — to determine what public and place-bound assets are being committed.

03 Negotiation — to convert those assets into enforceable fiscal, ownership, contractual or institutional claims.

04 Diffusion — to ensure domestic firms can use the technology and that each implementation leaves capability behind.

05 Stewardship — to manage the resulting assets and obligations over decades rather than announcement cycles.

This is not a sovereign wealth fund replacing an investment-promotion agency. It is a redesigned development system — staffed by promoters and relationship managers, but also by infrastructure economists, commercial negotiators, data-governance specialists, fiscal analysts and extension practitioners, with formal partnerships to utilities, finance ministries, regulators, universities and pension funds where the competence cannot sit in-house. Its operating rhythm changes too: an opportunity-cost assessment before incentives are discussed; a distinction between benefits the company would produce anyway and those created by the public agreement; a published capture ledger for major projects, revisited at five and ten years; performance tracked in claims secured, domestic assets created, supplier margins retained and options opened or closed — not in jobs announced.

This will be uncomfortable for organizations built around annual investment totals and ribbon-cuttings. A flagship project may score poorly; a smaller domestic expansion may create more lasting value; the best decision may be not to bid. The asymmetry it corrects is real and currently enormous: the typical public negotiation with a frontier technology firm today pits a company that models energy markets in-house against an agency staffed and evaluated for announcements. Compute, energy and talent will all be supplied by markets. Bargaining capacity will not. It has to be built deliberately, by whichever level of government owns the inputs.

Term sheets, in short, rather than tax holidays.

14 · The strategy after “becoming an AI hub”

An ambition the next region can repeat without changing a word is not a strategy. A real one starts with an exact inventory and ends with claims.

Every jurisdiction now wants some version of the same thing — an AI cluster, an innovation corridor, a place in the next technological economy. These ambitions are too general to guide a single decision, and they contain no standard for saying no. “Becoming an AI hub” is not a strategy, because the next region can say it without changing a word.

A real strategy begins with an exact inventory: the physical systems the place controls, the contextual information it has accumulated, the regulatory environments that could become trusted testing grounds, the firms and institutions that can actually implement the technology, and — the discipline most promotional documents skip — an honest judgment about which of these assets are truly scarce and which are merely described that way in the brochure.

Then comes the harder half, which is entirely about claims: who will own the resulting infrastructure, where the intellectual property will sit, who controls the data and the improvements built on it, which domestic organizations will be more capable after the project than before, what share of cash flow will have a local claimant, what costs other users will absorb, and what future activity the project enables or forecloses.

The governing question is not “how do we attract AI?” It is: what does this place contribute to the intelligence economy, and what enduring claim should that contribution command?

15 · The long game

The work of development must now be judged twice: did the economy become more capable, and did that capability create durable, broadly held claims on value?

Every development framework rests on an assumption so deep that practitioners eventually stop seeing it. The modern field assumed that productive capability was scarce, sticky and carried largely by people; that firms requiring it would employ workers; that employment would diffuse knowledge; and that wages would transmit part of the resulting value into the wider economy. These assumptions were never universally true. For most of the industrial era, they were reliable enough to organize an entire profession around. They are now becoming unreliable together.

The response is not to abandon learning, and not to retreat into economic nationalism. It is to stop confusing participation with prosperity. A place can host production without owning much of it. It can raise productivity without raising household income proportionately. It can supply valuable inputs without ever pricing them. It can attract record investment while quietly narrowing its own future choices.

The work of development must therefore be judged twice. Did the economy become more capable? And did that capability create durable, broadly held claims on value? The first question belongs to the theory the profession inherited. The second belongs to the era now beginning.

For seventy years, capability, employment and ownership arrived bundled together closely enough that the profession could concentrate on the first and treat the others as consequences. The bundle is coming apart. The places that prosper will be those that learn quickly, bargain competently, and own a meaningful share of what their assets make possible — that build the administrative capacity to value power, data, permission and context before trading them away; that can tell an ecosystem from an announcement, an asset from a subsidy, and a claim from a promise; and that are willing, sometimes, to decide that the right project is the one they decline.

Prosperity has never come from intelligence alone. It comes from the institutions that turn intelligence, effort and endowment into income, assets and choices that endure. Economic development’s next assignment is not merely to help places become more productive. It is to ensure they remain beneficiaries of what they help produce.

APPENDIX

Sources and method

This report is an argument, and it stands or falls on its reasoning. The empirical claims that anchor it are drawn from the public sources below, current to July 2026. Where a figure describes a specific project, it comes from the project’s own announcement or from contemporaneous reporting, and is presented as announced — announced figures are promises, not audited outcomes, which is part of this report’s point.

ENERGY AND DATA-CENTRE ECONOMICS International Energy Agency, Energy and AI (2025) and World Energy Investment (2025): global data-centre investment of roughly half a trillion US dollars in 2024; data centres at approximately 1.5 percent of global electricity consumption in 2024, with consumption projected to more than double by 2030.

PROJECT FIGURES Meta, “Hello, Sturgeon County!” (July 8, 2026): C$13 billion-plus investment, 1 GW capacity, 1,750 acres, approximately 3,000 construction workers at peak, 300-plus operational jobs, C$60 million in local infrastructure, full-cost energy commitments. Government of Alberta announcement and reporting by CBC News, BNN Bloomberg and The Globe and Mail (July 2026): projected provincial benefit of at least C$250 million per year; the C$4.6-billion, 970 MW Greenlight generation project.

Stargate Abilene: reporting by CNBC (September 2025), Texas Standard, Spectrum News and Data Center Dynamics (2025–2026): 6,000-plus construction workers at peak; published permanent-employment estimates ranging from roughly 300 to nearly 1,700.

NextStar Energy (Stellantis–LG Energy Solution joint venture), company releases and reporting by CBC News and Electric Autonomy (2022–2026): investment of more than C$5 billion; target of 2,500 jobs at full production; LG Energy Solution’s acquisition of Stellantis’s 49 percent stake (February 2026); reorientation of output toward grid-storage batteries.

AI, WORK AND DEVELOPMENT World Bank, research on artificial intelligence and digitally delivered services in the Philippines (2025): sector employment of roughly 1.8 million; firm-level adoption patterns; declining entry-level postings in exposed occupations; effects of foreign ownership on domestic control and reinvestment.

IMF, AI Preparedness Index : cross-country differences in infrastructure, human capital, innovation and regulation that condition returns to AI adoption.

OECD, work on the decoupling of median wages from productivity growth across member economies.

THEORETICAL FOUNDATIONS Robert Solow on growth accounting and the residual; Paul Romer on endogenous technological change; Douglass North and successors on institutions; Ricardo Hausmann, César Hidalgo and colleagues on economic complexity and the atlas of productive knowhow; the resource-curse and rentier-state literatures; Norway’s Government Pension Fund Global and the Alaska Permanent Fund as institutional references for converting temporary endowments into durable public claims.

A NOTE ON THE EXAMPLES The projects profiled in this report are used as illustrations of a structural pattern, not as verdicts on particular deals. Several announced terms of the Alberta project, in particular, align with practices this report recommends; whether the full ledger bears them out is a question that can only be answered after operation begins — which is precisely the reporting discipline the report argues for.