Anthropic’s new economic scenario model says AI shifts income from workers to capital owners by 2030 — up to 15 points in the extreme case. In Singapore that framing is wrong, because most households own both sides of the split. The real question is which households have enough capital stock to win.
This week Anthropic’s economics team published a scenario model of AI’s economic future (Korinek et al., 2026), built on a survey of more than 10,000 Americans[5]. An interactive explorer also lets readers plug in their own predictions about capabilities and adoption. The structure is worth stealing even if the numbers are US-specific[5]: it treats the economy as bundles of tasks — some augmented by AI, some automated, some newly created — and maps three scenarios from “internet-like” to “unprecedented.”
| Scenario | GDP by 2030 (vs no-AI path) | Labour share of GDP |
|---|---|---|
| Modest — impact roughly like the internet’s; gains within historical norms, arriving gradually | +1.6% ($34.1T) | 59.4% (down 0.6 pts from ~60%) |
| Substantial — AI capable of half of all knowledge work by 2030, mostly autonomous; economy grows at twice the normal rate | +8.3% ($36.3T) | 56.1% (down 3.9 pts) |
| Extreme — AI more productive than humans at nearly all knowledge work, likely requiring recursive self-improvement; GDP grows ~15%/yr | +32.4% ($44.4T) | 45.2% (down 14.8 pts) |
The survey’s typical respondent lands near the substantial scenario: GDP about 10% higher by 2030, unemployment around 5%. Roughly one in ten respondents holds views consistent with the extreme case. The model’s fourth finding is the one that matters for this memo: in every scenario the pie grows, but a larger share of each dollar goes to capital. Even when average wages rise, knowledge workers’ wages stagnate or fall while non-knowledge work pays more.
Source: Anthropic — “Economic scenarios” — Korinek et al. (2026), based on a survey of 10,980 US adults. anthropic.com
The headline framing is “labor vs capital.” That framing assumes two separate groups: people who sell time, and people who own the machines that replace it. In Singapore, most households are both at once — which changes what the model actually predicts for this economy.
Three structural facts make a direct port of Anthropic’s framework misleading here. Each one is verifiable in the national accounts, and each one changes who “wins” when capital’s share rises.
1. The labour-vs-capital split already happened — before AI arrived. Singapore’s labour share of GDP has fallen from 49.4% in 2014 to about 43% by 2024 (ILO).[1] That is a seven-point shift over one decade, driven by capital deepening and the structure of a small open economy — not by AI. The US model starts at ~60¢ per dollar going to workers; Singapore starts lower. So when Anthropic’s scenarios push the labour share down further, they are describing an acceleration of a trend this economy has already been running for ten years, not the start of one.
2. Most workers here are capital owners — through CPF and property. The Central Provident Fund forces roughly 20%+ of wages into savings that buy housing and financial assets[6]. Property dominates the national balance sheet (the wealthiest fifth of resident households held average net wealth of S$5.3 million in 2023, driven chiefly by property). When AI shifts income from labour to capital, a large slice of that “capital” lands back inside the same household — as CPF balances, rental income, and house prices. The conflict is not between workers and owners. It is between households with enough capital stock to capture the shift, and those without.
3. Adoption here isn’t a choice — it’s demographic arithmetic. The resident total fertility rate hit 0.87 in 2025, down from 0.97 the year before[2]; one in five citizens is now aged 65 or older (SINGSTAT).[2] With the labour supply shrinking structurally, “modest” AI adoption is closer to a floor than a baseline: productivity per worker has to keep rising just to hold GDP per capita flat. The state — EDB, SkillsFuture, the National AI Council — actively pushes adoption in ways no US federal agency does for private-sector software. That shifts scenario probabilities toward substantial relative to what the American survey implies.
The US model has two parties. Singapore’s economy has three — and the third one is where AI changes the policy levers most directly.
The foreign workforce stands at roughly 1.23 million, about a third of total employment (MOM). In 2025, four out of every five newly created jobs went to non-resident workers while resident employment grew by only ~12,000. That mix is managed with unusually direct instruments — sectoral quotas, levies, and the Progressive Wage Model’s floor for lower-wage roles.
AI changes which tasks need humans at all. As knowledge work gets automated, demand shifts toward the tasks that can’t be: care, trades, physical services. The foreign workforce is concentrated in exactly those categories — construction, manufacturing, services. So the scenario model’s “knowledge workers displaced, non-knowledge wages rise” finding lands here as a policy question with no US analog: does Singapore tighten or loosen the foreign mix as knowledge work shrinks? Quotas and levies are dials that can be turned within a budget cycle; in the US, the equivalent adjustment happens through migration politics that take a decade.
The wage data is already showing the split the model predicts. Median nominal wages rose about 5% to S$5,775 in 2025[3]. The growth is concentrated: PMET (professional, managerial, executive, technician) roles are growing at roughly four times the pace of non-PMETs, whose median income has crept along at about 0.5% a year. The divergence between the two groups is the early signature of task-level AI effects, visible before any large-scale displacement.
The same three scenarios, translated into what they mean for this economy’s capital-labour relationship:
Modest — the internet path. AI adds real but gradual productivity; GDP growth stays within historical norms. The labour share drifts down a point or two more over five years, consistent with the 2014–2024 trend rather than breaking it. PMET wage growth continues to outpace non-PMETs; the foreign mix is managed as usual. For households: the capital side of the balance sheet quietly compounds while wages grow at inflation-plus-a-bit. The split between capital-rich and capital-poor households widens slowly, invisibly.
Substantial — the revolution in knowledge work. AI does half of all knowledge work by 2030; the economy grows at roughly twice its normal rate (Singapore’s own growth forecast is already being upgraded to 4.5–5.5% for 2026, before this takes hold[4]). Knowledge workers’ wages stagnate — the PMET premium flattens or inverts as the tasks that earned it get automated; non-knowledge and care-sector wages rise on demand. The labour share falls several points further. For households: this is where the CPF channel matters most — capital income (property, financial assets) keeps rising even as wage growth for knowledge workers stalls, so household wealth can keep growing while paychecks don’t.
Extreme — recursive self-improvement. GDP grows ~15% a year; the labour share falls to roughly 45¢ per dollar in the US model. In Singapore terms: total labour income barely changes even as the economy more than doubles, and the gains concentrate in capital assets. The foreign workforce question becomes acute — with knowledge work largely automated, the managed mix of non-resident workers shifts decisively toward physical and care tasks. For households without meaningful capital stock, this scenario is genuinely bad; for those with property and financial assets, it is the best economic environment in a generation.
The common thread across all three: capital’s share rises. The only variable is by how much — 0.6 points, 3.9, or 14.8 in the US model’s own arithmetic.
If capital’s share of output is rising across every plausible scenario, the allocation implication is not subtle: wage growth will understate asset appreciation, and a portfolio sized to salary alone is structurally underweight the side of the economy that wins.
1. Hold enough capital assets — the split rewards balance sheets, not paychecks. In every scenario above, the households that do well are those whose income includes property and financial-asset returns, because that is where the AI dividend lands in this economy. The practical read: keep the equity and property sleeves sized for a world where labour share keeps falling — not sized to current wage growth, which is the lagging indicator.
2. Property is Singapore’s specific vehicle for the AI dividend. Productivity gains flow through house prices even when wages are flat: higher output per worker supports household wealth via the balance sheet rather than the paycheck. That channel is why the “workers lose” narrative from the US model understates what happens here. A household whose CPF and HDB/condominium position appreciates with productivity has already captured part of the capital share without buying a single AI stock.
3. The SGX problem: much of the capital share accrues to foreign owners, not local households. This is the uncomfortable corollary. The global AI trade — compute, frontier models, the capex cycle — is being captured by US and a handful of Asian names; Singapore-listed companies have largely lost out (the briefs this month flagged exactly that). EIMI/CNYA capture more of the trade than SGX does. So “own capital” in practice means owning the global AI complex through broad EM/Asia funds — not assuming local equities will deliver it.
4. The wage split is a hiring and compensation signal, not just a macro one. PMET growth at ~4x non-PMET pace means the premium for human judgment, client relationships, and accountability is rising even as routine knowledge work deflates. For anyone pricing services or building teams: charge for the tasks AI can’t do yet, and stop paying market rates for the ones it can.
The bear case against “capital wins” has real substance in Singapore, and it starts with the same CPF that makes the thesis work.
1. The forced-savings channel means household wealth can rise even as labour share falls. If ~20% of every wage is compulsorily converted into capital assets, then a falling labour share does not translate one-for-one into falling household income — part of the “labor” side is already parked on the “capital” side. The US model’s clean split between workers and owners simply doesn’t exist here at the margin.
2. Property ownership blunts the displacement story. In a substantial or extreme scenario, knowledge-worker wage stagnation lands on households that also hold appreciating property — the two effects partially offset within the same balance sheet. The “workers lose” framing assumes workers don’t own the factory; in Singapore most of them do, through housing and CPF-linked assets.
3. The state has levers no US government does. Land value capture (the state already takes a share of property appreciation via LTV rules, ABSD, and land sales), the ability to redirect CPF policy, and direct control over the foreign workforce mix. Together, these mean Singapore can tax or redistribute the AI dividend in ways that are structurally unavailable in the US. A future NDR could plausibly introduce an explicit AI-productivity levy or expand family-policy transfers funded by it — the 2026 NDR’s front-loading of marriage and parenthood policy is a preview of that direction.
4. The model itself admits its limits. Anthropic’s reviewers (Acemoglu, Autor, Klenow, Nakamura among them) flagged that the scenarios omit aggregate-demand effects from data-center buildout, don’t follow individual workers through displacement, and may understate how much AI accelerates technological progress itself. The extreme scenario is closer to a thought experiment than a forecast — which is fine for framing, but it means the 14.8-point labour-share drop should be read as an upper bound on a range, not a target.
Five signals will tell you which scenario is playing out — and whether the state starts redirecting the dividend:
1. The PMET vs non-PMET wage gap, quarter by quarter. It is already at ~4x growth divergence (SINGSTAT/MOM). If the gap widens further while total employment holds, that is task-level AI effects showing up in pay data — the earliest measurable signature of the substantial scenario. A narrowing would suggest augmentation rather than automation.
2. CPF reform signals from the next NDR cycle. Any change to contribution rates, investment options (the expansion of CPF Investment Scheme), or property-linked rules is the state repositioning households on the capital side before the shift completes. It is a strong tell that policymakers see the same model readers do.
3. Foreign workforce quota and levy changes as knowledge work shrinks. MOM manages the mix with sectoral quotas; if those tighten in construction and services while PMET foreign hiring loosens, the state is pre-positioning for a task-mix shift. The 2026 NDR’s commitment to “continue managing” the foreign workforce size gives this dial a political owner.
4. SkillsFuture budget shifts toward reskilling knowledge workers specifically. A meaningful reallocation from general training funds into PMET-targeted programs would be the labour-side response to exactly the displacement channel Anthropic’s model describes — and its scale would calibrate how serious the state thinks the substantial scenario is.
5. Whether local companies start capturing AI value or keep ceding it abroad. The SGX underperformance in the global AI trade is the portfolio-level version of this memo’s thesis. If Singaporean firms remain buyers rather than owners of the capital share, the dividend flows out through dividends and FX rather than into local balance sheets. Watch for domestic compute buildout, local model deployment at scale, or M&A that puts AI ownership on SGX balance sheets.
The US model asks whether workers or capital win. In Singapore the question is narrower and more personal: which households have enough of the capital side to win — and does the state decide, before the shift completes, who counts as one of them?
This analysis is based on publicly available data as of 2026-09-11. For related coverage, see The Hiring Mix and The Back Office Restructuring.