What Comes Next

The investigation has reached a useful transition point. The next research should not simply add more material about Musk, Bezos, Altman, Zuckerberg, Huang, agents, or data centers. Those subjects are now sufficiently mapped. The highest-value branches are the ones capable of testing the assumptions underneath the final thesis.

The central unresolved proposition is now:

If increasingly capable machine intelligence becomes abundant and inexpensive while its physical production remains capital-intensive, where does scarcity—and therefore durable economic and institutional power—actually migrate?

That question drives the ranking below.

1. The Economics of Abundant Intelligence: What Becomes Scarce?

Why This Matters

This is the most important discovery produced by the investigation, but also one of the least tested.

We began by treating advanced intelligence as the scarce resource. By the end, evidence of falling AI costs and expanding capability suggested something different: intelligence itself may become progressively less scarce.

Economic history suggests that when the price of an important input collapses, value often migrates toward complementary scarce resources. But we have not established what those complements would be for machine intelligence.

Possible candidates include:

• judgment and decision authority;
• trustworthy information and verification;
• human attention;
• energy and physical resources;
• identity and reputation;
• proprietary context and data;
• relationships and authenticity;
• capital;
• access to real-world action.

This branch could reorganize the entire investigation.

Key Questions

What actually happens economically when cognitive services become dramatically cheaper?

Which human cognitive capabilities are genuinely scarce today, and which merely appear scarce because expert labor is expensive?

When analysis, writing, coding, translation, design, research, and advice become abundant, which complementary inputs appreciate?

Does judgment become more valuable—or can judgment itself increasingly be automated?

Does abundant generation make verification scarce?

Does abundant synthetic content create an economic premium for authenticated human provenance?

Does AI reduce the importance of expertise while increasing the importance of goals, taste, reputation, relationships, and responsibility?

What happened historically when other scarce capabilities became abundant—for example physical power, calculation, communication, copying, storage, navigation, or information distribution?

Most importantly:

Is intelligence actually the binding constraint on economic and institutional performance today?

If it is not, AI could be far less transformative than capability benchmarks suggest.

Evidence to Seek

The strongest investigation would combine several evidence streams.

Economic history should examine falling costs of steam power, electricity, computing, telecommunications, storage, copying, and internet distribution and identify which complementary assets became more valuable afterward.

Current AI evidence should track price per useful task rather than benchmark performance alone.

Firm-level experiments should test whether providing abundant AI capability improves decisions, output, innovation, or merely increases the amount of analysis produced.

Labor data should examine changing wage premiums for expertise, judgment, management, interpersonal work, reputation, and accountability.

Markets for authentication, digital identity, provenance, cybersecurity, insurance, human-created content, and verification should be examined for early evidence of scarcity migration.

Household and small-business evidence would be particularly valuable because cheap intelligence may create its greatest welfare effects outside large corporations.

The investigation should also actively seek negative cases where intelligence was supplied but outcomes did not improve.

Those cases may reveal the actual bottleneck.

Possible Outcomes

Outcome A: Intelligence was the bottleneck. Cheap AI produces large productivity and welfare gains. The present investigation may actually underestimate the transformation.

Outcome B: Judgment becomes the bottleneck. AI generates abundant possibilities while humans and institutions remain responsible for selecting goals and evaluating tradeoffs. Human comparative advantage moves toward judgment.

Outcome C: Trust becomes the bottleneck. Cheap generation creates expensive verification. Identity, provenance, reputation, cybersecurity, authorization, and institutional trust become core infrastructure.

Outcome D: Physical reality becomes the bottleneck. Intelligence becomes cheap while energy, housing, manufacturing, minerals, healthcare capacity, construction, land, and real-world execution remain constrained. AI proves less economically revolutionary than digital capability suggests.

Outcome E: Intelligence was never the primary bottleneck. Motivation, coordination, politics, organizational inertia, capital, relationships, and implementation continue dominating outcomes. This would materially weaken the strongest transformative interpretation of AI.

Priority: #1 — Highest informational value.

2. Does AI Actually Produce Enough Economic Value to Justify the Infrastructure Boom?

Why This Matters

This is the most important empirical unresolved problem from the original investigation.

We established that AI infrastructure investment is enormous.

We established that AI capabilities are improving.

We established task-level productivity gains.

But we have not established that the economic value generated by AI is sufficient to justify the amount of capital being committed to producing it.

Those propositions are routinely treated as equivalent.

They are not.

This branch could distinguish a genuine general-purpose infrastructure revolution from a technologically important but financially overbuilt investment cycle.

Key Questions

How much attributable revenue is AI actually generating?

What are inference and training economics after accounting for depreciation and energy?

How quickly do GPUs become economically obsolete?

What percentage of announced data-center capacity becomes financed, permitted, constructed, energized, occupied, and economically utilized?

What utilization rates are required to justify current investments?

Are enterprise customers achieving measurable returns after implementation costs?

Are AI companies subsidizing usage to establish market share?

Where are cancellations, restructurings, write-downs, defaults, and failed deployments occurring?

Could AI be technologically revolutionary while destroying enormous amounts of investor capital?

Evidence to Seek

This branch should prioritize harder evidence over announcements:

SEC filings;

audited financial statements;

cloud segment disclosures;

AI-related revenue disclosures;

capital-expenditure data;

depreciation schedules;

GPU useful-life assumptions;

data-center occupancy and utilization;

power-purchase agreements;

project cancellations;

construction delays;

credit spreads;

private-credit structures;

loan covenants;

bankruptcy records;

CoreWeave and comparable infrastructure financing;

hyperscaler financials;

utility interconnection queues;

actual versus announced gigawatts;

enterprise AI ROI studies;

AI startup revenue and failure data.

A dedicated failure ledger should be built.

That was one of the strongest methodological recommendations produced by the original investigation.

Possible Outcomes

If revenue and productivity increasingly justify investment, the general-purpose infrastructure interpretation strengthens considerably.

If AI use grows rapidly but infrastructure returns disappoint, the investigation would support a hybrid conclusion:

technological revolution + financial overinvestment.

If capacity remains underutilized and enterprise ROI remains weak, the infrastructure-bubble explanation becomes substantially stronger.

If efficiency dramatically reduces infrastructure requirements, today's build-out may resemble an expensive transitional phase rather than the permanent architecture of AI.

Priority: #2.

3. AI, Expertise, and the Disappearance of the Apprentice

Why This Matters

The labor investigation may have focused on the wrong observable.

Mass layoffs are obvious.

Skill formation is not.

Many professions historically train experts by assigning novices relatively routine work:

junior programmers write simpler code;

young lawyers review documents;

analysts build models;

medical trainees evaluate ordinary cases;

research assistants collect and summarize information;

junior accountants prepare routine work.

AI is particularly capable at many of these tasks.

That creates a potential paradox:

AI could increase the productivity of today's experts while reducing the process that creates tomorrow's experts.

This could become one of AI's most consequential effects without appearing initially as unemployment.

Key Questions

Are firms hiring fewer entry-level knowledge workers?

Which junior tasks are disappearing?

Are those tasks economically disposable but educationally essential?

Does AI accelerate novice learning or encourage cognitive outsourcing?

Do AI-assisted juniors reach expert performance faster?

Do experienced workers benefit more or less than novices?

What happens to professional credentialing when machines perform training-stage tasks?

Could firms eventually face shortages of senior expertise because they stopped developing juniors?

What happens when children learn research and writing with AI available from the beginning?

Does persistent AI use strengthen or weaken independent reasoning?

Evidence to Seek

Longitudinal employment data by age and career stage;

job postings;

graduate hiring;

internship data;

promotion rates;

professional-service headcounts;

GitHub/software-development data;

law-firm staffing;

accounting-firm staffing;

medical training;

consulting;

finance;

randomized educational studies;

longitudinal cognitive testing;

teacher observations;

AI-use telemetry where ethically available;

experiments comparing AI-assisted learning with AI-assisted task completion.

The crucial distinction is between:

performance while using AI

and

capability after AI is removed.

Possible Outcomes

If AI accelerates learning, fears of deskilling weaken and AI becomes an expertise multiplier.

If immediate productivity improves while unaided ability deteriorates, the investigation uncovers a significant hidden cost.

If entry-level hiring declines without corresponding training alternatives, AI could restructure professional labor through pipeline compression rather than mass unemployment.

If little effect appears, labor institutions may prove more resilient than expected.

Priority: #3.

4. Trust, Identity, and Authority in an Agentic Economy

Why This Matters

The original investigation began with model intelligence but repeatedly encountered another problem as systems became more capable:

Who is authorized to act?

A chatbot providing information creates one level of risk.

An agent that can:

send email,

purchase goods,

move money,

sign agreements,

modify software,

access corporate systems,

operate machines

creates an entirely different institutional problem.

Intelligence alone does not create authority.

The investigation therefore identified trust as a potential future infrastructure layer.

That hypothesis deserves direct testing.

Key Questions

How will humans and machines authenticate AI agents?

How does a service determine what an agent is authorized to do?

Who is liable when an agent makes an unauthorized or harmful decision?

Can authorization be delegated recursively from one agent to another?

How are permissions revoked?

How are actions audited?

How will humans distinguish legitimate agents from malicious impersonators?

Can open interoperability coexist with secure identity?

Will users own their AI identity and context, or will platforms?

Could insurers become de facto regulators by determining which autonomous systems are insurable?

Will agent ecosystems resemble the open web, banking networks, mobile operating systems, or something fundamentally different?

Evidence to Seek

NIST's AI Agent Standards Initiative;

identity and authorization standards;

Agent2Agent and comparable protocols;

OAuth and existing delegated-authorization systems;

cybersecurity research;

prompt-injection and agent-security experiments;

financial payment networks;

digital-signature systems;

enterprise identity management;

insurance underwriting;

contract and agency law;

consumer-protection cases;

autonomous-vehicle liability;

API permission systems;

fraud cases;

machine-to-machine commerce experiments.

Historical analogues should include banking authorization, corporate agency, electronic signatures, credit-card networks, and internet certificate infrastructure.

Possible Outcomes

If secure interoperable identity emerges, agent ecosystems could remain relatively open and portable.

If identity and context become platform-controlled, agent systems could create stronger lock-in than today's apps.

If liability and cybersecurity prevent broad delegation, the autonomous-agent economy could remain much smaller than current technological forecasts suggest.

If insurance and private standards govern deployment more effectively than legislation, an unexpected institutional layer could become highly influential.

Priority: #4.

5. AI Beyond the West: Demography, Scarce Expertise, and the Global Distribution of Gains

Why This Matters

The original investigation was inevitably biased toward the places building frontier AI.

That may not be where AI ultimately matters most.

The United States contains abundant lawyers, programmers, physicians, consultants, teachers, and financial professionals relative to much of the world.

For populations where professional expertise is scarce, inexpensive AI could have far greater marginal value.

At the same time, aging societies may experience automation differently from young societies.

This branch would test whether our entire labor and political-economy framework is too Western.

Key Questions

Does AI have larger productivity effects where expertise is scarce?

How does AI affect developing economies where informal work dominates?

Can AI substantially reduce language barriers?

Does AI improve access to education, healthcare information, legal assistance, agricultural knowledge, and software creation?

Do countries with aging populations adopt robotics differently from countries with young labor forces?

Could automation compensate for demographic decline?

Could AI undermine developing countries whose growth strategies depend on outsourced cognitive labor?

Will AI enable economic leapfrogging—or reinforce dependence on foreign cloud and model providers?

Who captures the resulting surplus?

Evidence to Seek

World Bank data;

ILO data;

UN demographic projections;

national statistical agencies;

AI adoption surveys outside North America and Europe;

India, Southeast Asia, Africa, Latin America, China, Japan and South Korea as comparative cases;

education trials;

agricultural-extension experiments;

healthcare-assistance studies;

translation performance across low-resource languages;

mobile AI adoption;

informal-economy surveys;

robot density;

age structure;

professional-worker density;

wage data;

cloud geography;

local model development;

digital-payment infrastructure.

The strongest design would compare countries differing in:

labor scarcity,

income,

professional-service availability,

language,

energy infrastructure,

digital connectivity,

demography.

Possible Outcomes

If AI's greatest gains occur where expertise is scarce, the current concentration narrative would need substantial balancing: centralized production might enable enormous global decentralization of expertise.

If poorer countries remain dependent on foreign infrastructure without developing complementary institutions, AI could instead reinforce international dependency.

If aging societies derive disproportionate value from robotics, automation may become partly a demographic adaptation rather than principally a labor-displacement technology.

If outsourcing economies are heavily disrupted, AI could redistribute economic activity internationally even without dramatic employment effects in wealthy countries.

Priority: #5.

Overall Priority Ranking

The five branches rank by their ability to change the central interpretation, not merely add interesting material:

1. The Economics of Abundant Intelligence — What Becomes Scarce?
2. Does AI Produce Enough Economic Value to Justify the Infrastructure Boom?
3. AI, Expertise, and the Disappearance of the Apprentice
4. Trust, Identity, and Authority in an Agentic Economy
5. AI Beyond the West — Demography, Scarce Expertise, and Global Distribution

There is an important distinction between #1 and #2.

#2 is the most important empirical test of the existing thesis.

But #1 has the greatest potential to produce an entirely new thesis.

That makes #1 the highest-value next investigation.

Why I Would Start With #1

The original investigation effectively moved through three questions.

It began with:

Who is building the future?

Then improved to:

Where will durable technological control settle?

And finally uncovered:

What happens if intelligence itself stops being scarce?

The third question contains the first two.

If intelligence remains expensive and scarce, much of our existing infrastructure analysis becomes even more important.

If intelligence becomes abundant, however, the entire locus of economic power may move somewhere else.

That somewhere else could be:

energy if physical computation remains expensive;

attention if content becomes infinite;

trust if synthetic information becomes ubiquitous;

judgment if options become abundant;

identity if autonomous agents proliferate;

relationships if human interaction becomes scarce;

authenticity if synthetic production becomes dominant;

physical execution if thinking becomes cheap but acting remains difficult.

Or we may discover something more disruptive:

intelligence was never the primary bottleneck in human institutions to begin with.

That result would force us to reinterpret much of the original 456-page investigation.

Which is exactly what a high-value follow-up should be capable of doing.