AI Labor Savings Do Not Flow Straight to Shareholders

Anthropic's economic scenarios and occupational exposure data point to different opportunities in memory, power equipment and enterprise AI. The investment test is pricing, recurring margins, cash conversion and the valuation already paid.

In Anthropic’s extreme scenario, the economy is 32.4% larger at the start of 2030 than on a no-AI path, while cognitive employment is 21.5% lower than in mid-2026. Neither number is a forecast for semiconductor revenue or a probability of job loss. The comparators are different, and the scenario has no assigned probability.

For Korean equities, the useful question is who retains the savings when office work becomes cheaper. A bank may reduce processing costs while its outsourcing supplier loses billable hours. A software vendor may sell more AI services while paying still more for infrastructure. A memory supplier may benefit from demand, but expansion, depreciation and the purchase valuation determine shareholder returns.

The research supports a conditional screening framework, not indiscriminate purchases of AI-related stocks. Separate constrained infrastructure, enterprise implementation, productivity beneficiaries and businesses whose revenue unit is disappearing. Company mappings and investment judgments below are the author’s analysis, not Anthropic’s recommendations. Current share prices and valuation multiples have not been verified for this article; no target prices or immediate buy calls are offered.

The macro model is not an occupation-by-occupation forecast

The September economic paper models two labor groups. The detailed exposure readings come from a separate March study using observed Claude activity and task characteristics. Exposure is not displacement, and US occupational aggregates cannot simply be transplanted to Korea.

Published Table 3: start of 2030ModestSubstantialExtreme
GDP level versus no-AI path+1.6%+8.3%+32.4%
Average wage versus no-AI path+0.7%+2.1%+9.7%
Cognitive wage versus no-AI path+0.4%−0.3%−11.5%
Cognitive employment versus mid-2026−0.5%−3.9%−21.5%
Aggregate unemployment rate3.9%4.6%11.9%
Capital stock versus no-AI path+2.3%+13.8%+56.3%
Capital share of income40.6%43.9%54.8%
Capital income versus no-AI path+3.1%+18.9%+81.4%

These are published model outputs, not independently re-solved estimates. GDP gaps are level differences, not compound annual growth rates. Higher average pay among employed workers also says little by itself about displaced workers’ income.

The investor’s inference is that aggregate prosperity and the profitability of an individual business can diverge. The owner of a needed input and a vendor paid for a shrinking activity can face opposite outcomes in the same economy.

Capability, deployment and automation have different denominators

Task mass in the paper is weighted by the pre-AI wage bill. It is not headcount. With capability mass m, deployment d and automation fraction ψ, the mechanical quantities are:

2030 assumption or calculationModestSubstantialExtreme
Capability mass m20%30%50%
Deployment within affected tasks d20%40%60%
Automation fraction ψ50%75%90%
AI-performed mass m×d4%12%30%
Automated mass m×d×ψ2%9%27%
New-task reinstatement ratio ρ50%25%0%
Net labor-task mass removed (1−ρ)×m×d×ψ1%6.75%27%

The final row is not an employment forecast. Another unit trap is the productivity parameter a: 0.30, 0.45 and 0.80 are log gains, corresponding to exp(a)−1 of about 35.0%, 56.8% and 122.6%, not 30%, 45% and 80%.

For equities, do not multiply the number of office workers by a developer’s token consumption. A usable demand model needs deployed workers or workflows, frequency, task duration, successful completion, retries and the computation required per completion. It then needs product mix and prices. A GDP uplift is not an HBM shipment forecast.

Occupational change identifies the customer and the endangered revenue unit

The exposure percentages below are Figure 3 values from the March research, not September adoption rates. The business implications are analytical hypotheses.

OccupationObserved exposurePotential spending shiftRevenue at risk
Computer programmers74.5%Deployment, testing and infrastructureLabor-hour-based development
Customer service representatives70.1%Integrated AI contact centersSeat-based outsourcing
Data entry keyers67.1%Documents, ERP and exception handlingRoutine input services
Medical record specialists66.7%Coding, claims and records integrationLabor-priced administrative services
Market research/marketing specialists64.8%Proprietary data and measurable experimentsGeneric reports and content
Wholesale/manufacturing sales representatives*62.8%CRM, quotations and orderingStandardized outreach
Financial/investment analysts57.2%Licensed data, verification and audit trailsUndifferentiated information summaries
Software QA analysts/testers51.9%Release assurance and regression testingManually billed test execution
Information security analysts48.6%Authorized response and identity controlsUndifferentiated alert triage
Computer user support specialists46.8%Diagnosis, access approval and recoveryRoutine support labor

*Excludes technical and scientific products, following the source category. These exposures should not be generalized to every salesperson or developer.

The same change can benefit a customer and hurt its supplier. A reduction in processing time does not tell us which company keeps the economic surplus. That depends on contracts, competition and whether billing is tied to time, seats, transactions or outcomes.

Memory needs deployment evidence, not a labor-market multiplier

Samsung Electronics’ July results attributed memory strength to server-focused demand and higher prices. Management also described server DRAM, enterprise SSDs and HBM as areas of expected second-half demand growth. Those statements are company evidence and guidance, not independent proof of future demand.

Samsung Electronics and SK hynix therefore belong on a memory watchlist, but the test is application-level demand, product qualification, shipment mix, pricing and cash returns after expansion. This article does not estimate SK hynix earnings or imply an unverified valuation discount.

A reasonable upside hypothesis is that recurring office-agent execution broadens inference demand. A reasonable downside hypothesis is that more efficient inference, lower prices, weak customer monetization or excessive expansion reduce the profit captured per unit of usage. Both can be true at once: token volume may grow while hardware profits disappoint.

Within one diversified company, effects can also offset. Samsung reported an operating loss in its mobile and networks businesses alongside memory strength, citing component cost pressure in its July results. Investors should not apply a memory pricing benefit to the entire group without a segment bridge.

Power equipment benefits from scarce delivery capacity, not unlimited scarcity

HD Hyundai Electric told Reuters that its backlog reached USD 8.5 billion at end-June, 23% above year-end, with more than three years of orders and discussions extending to 2030 deliveries. This is evidence of contracted demand and long lead times, not certainty about the margin of every future order.

An investor should connect orders to delivery schedules, advances, cancellation terms, cost pass-through, factory expansion and cash collection. A long backlog can protect utilization; it can also contain low-priced contracts signed before costs rose. Backlog value alone cannot resolve that distinction.

The macro paper offers a useful stress test. Holding the extreme technology scenario but changing capital supply elasticity gives:

Capital-supply elasticity εGDP gapNet return on capitalCapital stock gap
1+21.3%10.3%+33.5%
3+32.4%8.3%+56.3%
6+37.2%7.5%+67.1%
Unlimited+43.3%6.5%+82.2%

These are Table 5 outputs; the no-AI net return is 6.5%. Capital elasticity is not a company’s cost of capital. The investment inference is narrower: demand can support more capital without sustaining a permanently higher return on each unit. Successful capacity additions can eventually erode the scarcity that made a supplier attractive.

Samsung SDS and LG CNS require a bridge from projects to recurring profit

Samsung SDS reported Q2 revenue of KRW 3.7178 trillion and operating profit of KRW 231.8 billion, up 5.9% and 0.7% year on year. Cloud revenue grew 17%. Logistics revenue was KRW 1.9553 trillion, about 52.6% of total revenue by division.

That mix is why company-wide revenue cannot be labeled AI revenue. The analytical question is whether new deployments become reusable products and recurring operating contracts, or remain labor-intensive implementation projects. Faster code generation may reduce delivery costs while customers demand lower project prices. A growing AI order book is not enough to determine the net effect.

LG CNS describes contact-center offerings spanning integration, generative AI assistance, subscriptions and usage-based services. Its own product material discusses implementation, maintenance and security challenges. This establishes commercial exposure, not a verified AI profit contribution.

For both companies, request the share of repeatable products, external-customer renewals, deployment costs, ongoing human support, infrastructure depreciation and operating cash flow. Infrastructure ownership and software distribution should be valued separately when their capital requirements differ.

Banks and insurers may gain productivity and still lose elsewhere

Financial analysts, support staff and administrative workflows suggest a cost-saving opportunity for financial institutions. But savings may be passed to customers through fees or prices. Integration expenses, compliance, error correction and labor adjustment reduce what remains.

There is also a second-order risk. A technology transition that damages some borrowers’ employment can worsen credit costs at the same time it improves back-office efficiency. This is a scenario to test, not a forecast for Korean defaults. Insurers similarly need claims-cost and pricing analysis rather than a blanket labor-savings uplift.

Education, healthcare and field services need the same caution. Stronger demand for human services may raise both sales and wages. Without pricing power, the operator may not be the beneficiary. The paper excludes advanced robotics, and its cognitive category includes healthcare practitioners and technical occupations; the other-group wage gain cannot be assigned directly to nurses.

Saved time must become cash, and cash must survive competition

Consider an illustrative workforce of 100 employees at KRW 80 million annual fully loaded cost each: KRW 8 billion total. Assume 40% eligible work, 50% deployment, 80% successful completion and 50% conversion of saved time into actual financial savings.

Annual gross savings = KRW 8 billion × 40% × 50% × 80% × 50% = KRW 640 million.

Assume KRW 160 million software fees, KRW 80 million direct inference/human-review costs not included in those fees, and KRW 40 million annualized integration/operating costs. Total costs are KRW 280 million; net customer benefit is KRW 360 million.

At only 20% cash realization, gross savings fall to KRW 256 million and net benefit becomes negative KRW 24 million. Break-even realization is 280 ÷ (8,000×0.4×0.5×0.8) = 21.875%. These are author assumptions, not observed company results or the paper’s own simulation.

A separate pricing example illustrates the vendor problem. A 20% seat reduction needs a 25% price increase merely to keep revenue unchanged: 0.8×1.25=1.0. If inference cost increases, unchanged revenue can still mean lower profit. Outcome or usage pricing may help, but only when the customer accepts the revised economics.

Wage adjustment makes headline unemployment highly assumption-sensitive

Table 6 changes wage rigidity under the extreme technology path:

Annual relative-wage rigidity ξCognitive wage gapCognitive unemployment
0, flexible−42.2%2.6%
0.5, central assumption−11.5%17.9%
0.9, highly rigid+2.8%24.0%

It would be a mistake to treat 17.9% as a precise forecast and then value consumer, recruitment or education stocks from that single number. Korean institutions, wage-setting, demographics and sector mix require separate evidence. A better portfolio stress test considers lower wage income, weaker hiring and slower consumption without claiming the model predicts their exact Korean magnitudes.

Earnings growth can be offset by the valuation paid

Suppose EPS grows 15% annually for three years, while the P/E moves from 35 to 25. The annualized price return, excluding dividends and taxes, is:

(1.15³×25÷35)^(1/3)−1 = approximately 2.8%.

The example uses no actual stock’s current multiple. It illustrates why a correct industrial thesis does not guarantee an attractive purchase. A price decision needs a verified share price, diluted share count, normalized earnings, net debt, maintenance and expansion spending, and a plausible exit valuation.

Free cash flow provides a second filter. Rising operating profit may be absorbed by equipment purchases and working capital. For cloud operators, utilization and equipment life matter; for memory suppliers, process investment and cycle-normalized returns matter; for power equipment, factory spending and collection schedules matter.

The watchlist is conditional and has explicit rejection tests

Candidate or business groupEvidence needed to strengthen the caseEvidence that weakens it
Samsung Electronics / SK hynixQualified high-value products, price/mix improvement and cash generationDemand growth without profits; excessive capacity or investment burden
HD Hyundai ElectricBacklog converts to profitable deliveries and cashDelays, cancellations, cost inflation or a faster end to scarcity
Samsung SDS / LG CNSReusable deployments, recurring contracts and margin improvementRevenue scales with staffing; AI profit contribution remains unverified
Banks / insurersRetained expense savings after implementation and risk costsFee competition or credit/claims costs offset savings
Seat/hour-priced servicesSuccessful transition to paid outcomesFalling billed units faster than achievable price increases

In a modest scenario, prefer already demonstrated returns and avoid paying for an unproven boom. In a substantial scenario, examine both infrastructure and the conversion of enterprise projects into recurring profit. In an extreme scenario, test credit, consumer demand, platform consolidation and supply expansion alongside the obvious volume upside. These are conditional analytical responses, not assigned probabilities.

The next purchase should wait for the missing company-level bridge: occupational change → paying customer → revenue unit → retained margin → free cash flow → price. The size of the labor market cannot substitute for any of those steps.

Source and spreadsheet limitations

The 57-page technical paper and the separate occupational study were reviewed. The original Excel was not obtained from the public page. The accompanying independent workbook transcribes tables and calculates mechanical checks and business illustrations; it does not validate original spreadsheet formulas or re-run the full monthly model.

One accounting check remains unresolved. Table 3’s extreme GDP gap of 32.4%, labor share of 45.2% and baseline share of 60% imply (1.324×0.452÷0.60−1)×100 = −0.259% labor-income change on a common-denominator identity, versus the reported +0.5%. The roughly 0.759-percentage-point difference cannot be explained by rounding those two inputs to one decimal alone. The published value is retained and the definition/output aggregation flagged for source verification, not declared a confirmed paper error. No investment conclusion relies on the precision of +0.5%.

Sources

Built with Hugo
Theme Stack designed by Jimmy