The Great AI Fork: What the IMF’s New Model Says About Asia’s Next Two Decades

A fresh IMF working paper puts a number on something economists have quietly worried about for a while — that AI might not lift Asia together so much as split it. Here’s what the model actually shows, and what it means for anyone sitting on an AI investment decision right now.

TL;DR

• The IMF’s new working paper, AI and Economic Divergence in Asia (Che, Xin & Yoshida, August 2026), treats AI adoption as a capital-deepening choice rather than a software rollout — and finds advanced Asia adopts almost immediately while many emerging economies get pushed back a decade or two, longer still if AI progresses fast.
• Late adopters don’t just miss out. They get actively squeezed: early adopters’ investment pushes up global interest rates, and that raises the cost of capital for everyone still trying to catch up.
• Structural reform is the biggest lever laggards actually control — a decade of serious reform to skills and productivity can pull adoption forward by ten to twenty years.
• Inside every country that adopts, AI reshuffles who wins: high-skilled workers over low-skilled ones, and older asset-holders over younger wage-earners.
• Redistribution helps, but there’s no free lunch — targeted transfers cut inequality harder, universal transfers cost less growth. Which one makes sense depends entirely on where a country starts.
• For healthcare, MedTech and life sciences specifically, the paper’s mechanisms translate into uneven margin gains — concentrated in specialist-heavy, capital-intensive care and precision-manufacturing hubs, while EMDE diagnostic chains and device manufacturing face a rising cost of capital right when they need it most.
• India’s a harder fit for the model than most of the sample — a managed capital account, a huge informal workforce, and a services-export engine that supplies AI labor to the world all sit outside what the framework can represent, so India’s chart position should be read as directional, not calibrated.

1. AI as Capital, Not Code

Most corporate conversations about AI still treat it as a productivity add-on — a copilot here, an automated workflow there. The IMF authors model something structurally different: what they call AI adoption “at scale,” meaning the point where AI stops being a marginal tool and starts reshaping the whole production structure of an economy, because firms and governments are pouring money into the physical scaffolding underneath it — data centers, chips, cloud capacity — not just subscriptions.

That’s a deliberate line to draw. A team using a language model doesn’t move a country’s GDP numbers. What the model actually tracks is the moment a country’s capital-to-labor ratio gets favorable enough that switching the entire economy to a more capital-intensive, skill-biased way of producing things becomes the obviously profitable move.

The technology doesn’t just create a one-time gap in income levels — it can actively work against late adopters during the transition, because one country’s AI investment boom becomes another country’s higher borrowing costs.

The model itself is a small open economy, built around overlapping generations — households living to 100, three skill tiers, capital that moves freely across borders while workers stay put — calibrated separately for fifteen Asia-Pacific economies. It’s adapted from an earlier NBER framework by Benzell and coauthors, and it’s designed to answer two questions at once: who gets to AI first, and who gets hurt once they do.

2. Why Early Movers Keep Pulling Further Ahead

The mechanism here is almost uncomfortably simple, which is part of what makes it convincing.

Three things decide who adopts early: how capital-intensive a country already is, how skilled its workforce is, and its underlying productivity. Countries already running capital shares near 40–50% of output — Japan, Korea, Singapore — are mechanically closer to the threshold. Throw in an aging, shrinking workforce, and capital-heavy technology becomes even more attractive as a way to keep output up without more workers. Japan and Korea, for instance, are staring down a 25–30% decline in working-age population by 2050.

That sets off a loop. Capital-rich, skill-rich, aging economies adopt first. They invest heavily to deploy the new technology. That investment surge pushes up global interest rates. Countries still trying to accumulate enough capital to cross their own threshold now face a higher cost of capital than before AI existed at all — so their accumulation slows, and their adoption date gets pushed further out.

The paper puts real numbers on this. Under a moderate scenario, emerging and developing economies see growth run half a point to six-tenths of a point below what it would have been without AI, every year, during the pre-adoption period — not because they’re standing still, but because they’re fighting a headwind that wouldn’t exist otherwise. Investment-to-GDP ratios come in 5–6 points lower over the same stretch.

The uncomfortable twist: faster global AI progress doesn’t just widen the gap, it can stretch it out longer. The quicker rich economies deepen their capital stock, the sharper and more sustained the interest-rate spike, and the harder it becomes for everyone else to catch the elevator before the doors close.

3. Growth Winners, Laggards, and the Eventual Catch-Up

The paper walks through three country sketches — Korea as the early mover, Vietnam somewhere in the middle, Bangladesh bringing up the rear — and the same shape repeats across the region. Early adopters see growth jump the moment they cross the threshold, and the jump scales with how fast AI itself is progressing. Late adopters see the opposite at first: growth actually dips below baseline before they adopt, only recovering — and then overshooting — once they finally cross their own line.

That overshoot matters. Once a late adopter does cross, growth doesn’t just return to normal, it runs hot for a while, as the economy catches up fast under the new, more productive technology. Averaged across the region, the authors put the long-run growth boost at roughly half a point to a full point annually for advanced Asia, and two-tenths to eight-tenths of a point for emerging economies — a wide range, and deliberately so, since it depends heavily on how quickly the underlying technology itself moves.

Worth saying plainly: the authors are upfront that these are illustrations of a mechanism, not forecasts. The fastest scenario they run — capital share rising ten times the historical US pace — is meant to show direction, not to predict what will actually happen to any one country’s investment numbers.

4. Where Governments Actually Have Leverage

This is the section that matters most for anyone advising a government, a sovereign fund, or a company with real EMDE exposure. The authors simulate two reform packages — moderate and ambitious — run over ten years, hitting three levers at once: raising the share of high-skilled workers, lifting labor productivity across every skill tier, and improving how efficiently capital gets allocated.

The payoff is bigger than you’d expect. Countries otherwise projected to adopt AI in the 2050s or 2060s could pull that forward by one to two decades through sustained, serious reform. And the gains compound — earlier adoption means capturing AI’s upside sooner and spending fewer years exposed to the interest-rate headwind other countries’ adoption creates.

The authors are honest that the reform magnitudes they model are large — on the order of Korea’s transformation in the 1970s, or China’s in the 2000s. They’re not claiming this is easy or typical. They’re showing what’s possible if a country actually commits to it.

5. Inequality Doesn’t Wait for the Growth Dividend

Two fault lines open up inside every country that adopts, and neither is subtle.

Skill first. High-skilled workers gain the most in every scenario, because their wages are least affected by the declining labor share and most boosted by their complementarity with all that new capital. Low-skilled workers have a rougher ride — under slower AI progress, their wages can actually fall below what they’d have been without AI at all, before eventually recovering once capital deepening overtakes the drag from a shrinking labor share. Under faster AI progress the picture flips somewhat — enough capital gets built fast enough that even low-skilled wages rise early — but the gains stay much smaller than for everyone else.

Then generation. This is the finding I find more striking, honestly. Older cohorts see gains earlier and bigger than younger ones — not because they work more, but because they hold assets. As AI-driven investment pushes interest rates up, people who’ve spent decades saving earn a lot more on what they’ve already put aside, while younger and mid-career workers see their gains build slowly, through wages, one paycheck at a time. In some scenarios mid-career income even dips below baseline for a while — partly because low-skilled workers choose to work less, partly because higher interest rates make it more expensive to borrow against future earnings. The paper is careful to note this isn’t a welfare loss in the strict sense — people are optimizing over their whole lives — but it’s still a real, visible widening of the income gap between generations.

The shift of national income toward capital favors asset-rich households — while younger workers, who depend on labor income and borrow against future earnings, capture a smaller share of the gains during the transition.

6. The Redistribution Trade-Off Nobody Gets to Skip

The authors test two ways to redistribute the gains, each costing 5% of GDP: a targeted transfer aimed at low- and medium-skilled workers, and a universal transfer paid to everyone. Both reduce inequality. Targeted transfers do it more effectively.

But targeted transfers cost more growth. Taxing high-skilled workers to fund a transfer they never see creates a sharper disincentive to work, and the low- and medium-skilled recipients also tend to work a bit less once the extra income arrives. Universal transfers spread the tax burden and the benefit across everyone, which softens the distortion for any single group — at the cost of redistributing less precisely. And the output cost of either approach scales with how high a country’s tax rates already are: Japan pays more for the same redistributive punch than Singapore does, purely because it’s starting from a higher baseline.

There’s no universally right answer here. It genuinely depends on a country’s fiscal starting point, its skill mix, and its demographics.

7. An India Lens: Where the Model’s Assumptions Strain

India appears throughout the paper as a supporting example — young, growing, labor-abundant, a later adopter. That’s fair as far as it goes. But several of the model’s simplifying assumptions bend harder against India than against almost any other economy in the sample, and it’s worth naming them before treating the adoption-timing charts as a read on India’s actual trajectory.

The capital-mobility assumption doesn’t fit India particularly well. The whole interest-rate-headwind mechanism — early adopters’ investment pushing up global rates, late adopters getting squeezed — depends on treating each economy as a price-taker in a fully open global capital market. India isn’t one. The RBI runs a managed capital account, with quantitative limits on external commercial borrowing, FPI debt ceilings, and active intervention in the currency market. The paper flags this caveat itself, for countries with “closer capital accounts and limited pass-through from global to domestic interest rates” — and India is arguably the largest economy in the sample where that caveat actually bites. The model may be overstating how much of the global financing headwind India actually imports.

A single national adoption date hides more divergence than the paper’s cross-country charts show. The model treats each country as one calibrated economy crossing one threshold. For a federal, continent-scale economy like India, that’s a significant compression. Bengaluru’s GCC and IT-services corridor is already operating close to frontier AI-adoption behavior — capital-intensive, skill-rich, plugged into global AI supply chains — while large parts of India’s labor market remain informal, low-skilled, and effectively outside the model’s formal wage-and-hours framework altogether. The within-country dispersion inside India may rival or exceed the cross-country dispersion the paper spends most of its pages on.

India isn’t purely an AI-taker, and the paper’s own caveat about “missing rents” cuts both ways here. The authors are explicit that they don’t model the rents flowing to countries that produce AI rather than adopt it — chips, IP, platform dominance, frontier compute. Applied naively, that reads as a warning that EMDEs like India are worse off than the model shows. But India’s IT-BPM and GCC sector is one of the more significant global suppliers of the labor that builds and deploys frontier AI systems for the rest of the world — a services-export channel the model’s single-good, single-sector structure has no way to represent. That’s a real omission, but not obviously one that cuts against India the way it does against, say, Bangladesh or Nepal.

The informal-sector gap is the biggest structural miss. The household side of the model assumes workers optimally choose hours against a market wage, with labor-income taxes and transfers flowing through a formal tax-and-benefit system. A large majority of India’s workforce sits outside that system entirely — informal, cash-based, without access to the kind of redistributive transfer mechanisms the paper’s targeted-versus-universal analysis assumes governments can cleanly deploy. The equity-efficiency trade-off the paper models so carefully in Section 6.2.3 may simply not be executable at the fiscal and administrative scale India would need, regardless of which scheme looks better on paper.

Net read: the paper’s broad direction for India — young demographics and labor abundance delaying economy-wide adoption, comprehensive reform pulling that timeline forward — is probably right as a first-order story. But the specific mechanisms doing the work in the model (perfect capital mobility, a single national threshold, a formal labor market, an economy that only consumes AI rather than also exporting the labor that builds it) are each weaker fits for India than for most of the other fourteen economies in the sample. Read India’s chart position as a rough signal, not a calibrated forecast.

8. Industry Focus Note: The Clinical Capital Divide — Healthcare, MedTech & Life Sciences

The paper never mentions healthcare by name — but its three core mechanisms map onto the sector unusually cleanly, and the IMF’s own action plan for policymakers reads almost like a blueprint for what healthcare operators need to do too.

The IMF’s prescription for countries facing this divide is essentially threefold: run comprehensive structural reforms to strengthen human capital and productivity, deploy redistributive policies to counter widening wage and generational gaps, and accept that targeted transfers buy more equity while universal transfers buy more efficiency. Swap “country” for “hospital system” or “MedTech manufacturer” and the same three imperatives — reform your fundamentals, protect your workforce deliberately, and choose your trade-offs with eyes open — turn out to translate directly into a Clinical Capital Divide across Asia’s healthcare economy.

1. The margin squeeze on EMDE diagnostic chains. The model’s central mechanism — early adopters’ AI investment pushing up global interest rates — lands hardest on healthcare systems that lean on debt to fund physical expansion. Building tertiary hospitals, scaling diagnostic-chain footprints, standing up new imaging or pathology capacity: all of it is capital-intensive, and in EMDE markets, mostly debt-financed. The same window in which these systems need to invest in AI-enabled diagnostics and predictive workflows is the window in which their borrowing just got more expensive, purely because Korea, Japan, and Singapore are drawing down the same global capital pool faster.

2. The restructuring of hospital workforces. Because AI is explicitly skill-biased in this model — rewarding high-skilled labor while compressing wages for others during the transition — margin expansion inside a hospital system won’t distribute evenly. Expect it to concentrate around top-tier specialists, radiologists, and clinical-informatics leaders well before it reaches the much larger base of nursing and allied-health staff who actually run day-to-day care delivery. That’s not a reason to slow AI adoption; it’s a reason operators need to get ahead of workforce restructuring — upskilling nursing and allied-health talent toward AI-complementary tasks — before displacement pressure builds rather than after.

3. MedTech manufacturing’s “greenfield” crisis. Structurally-prepared economies adopt AI earlier and see immediate gains; late adopters face delay. For precision-manufacturing hubs — Korea and Japan’s device base, Singapore’s biomanufacturing cluster — that means crossing the threshold for automated production and AI-assisted quality assurance well ahead of EMDE-based contract manufacturers. Those manufacturers now face a double bind: rising capital costs just to upgrade facilities, and a widening defect-reduction and cycle-time gap against early-adopting competitors who are already deploying AI-driven quality control. Greenfield expansion decisions that assumed a stable cost of capital need a second look.

4. Life sciences R&D and the rents of the frontier. The model deliberately leaves out the economic rents — IP, platform dominance, frontier compute access — that accrue to the countries actually producing AI rather than merely adopting it. That omission matters most for life sciences, where drug discovery, genomic sequencing, and clinical-trial optimization increasingly run on frontier model capability rather than adoption-at-scale in the paper’s narrower macroeconomic sense. Markets without a domestic AI or compute ecosystem risk becoming permanent renters of upstream tools rather than capturers of the compounding value underneath them — even if their CROs and diagnostic chains digitize quickly on the surface.

8. The Healthcare CXO Action Plan

Translating the IMF’s country-level playbook into an operator’s playbook:

• Lock in capital now. AI adoption at scale demands real upfront investment in physical and digital infrastructure — and the model says that financing is only going to get more expensive as early adopters bid up the global cost of capital. Secure long-term financing for clinical-technology expansion before that happens, not after.
• Upskill the frontline, deliberately. Shift HR and operating budgets toward retraining nurses and allied-health professionals now, positioning them as complements to incoming AI diagnostic and workflow tools rather than as labor competing against them.
• Reevaluate the supply chain on the right metric. Audit MedTech and device-manufacturing partners on capital-intensity and AI-readiness — not on the historical labor-arbitrage metrics that used to determine where contract manufacturing made sense.

9. Five Mistakes Leaders Make Reading Forecasts Like This One

1. Treating “adoption” as a light switch. The model’s adoption dates are structural thresholds, not calendar predictions — small tweaks to productivity or skill assumptions can shift them by years.
2. Assuming cheaper hardware bails out late adopters. It doesn’t, at least not in this model. As long as the technology frontier keeps moving, the skills and infrastructure needed to use it well don’t get cheaper at the same rate chips do.
3. Reading “growth gain” as evenly shared. Aggregate GDP going up and ordinary households doing better are two different stories, and the paper shows them pulling apart for years at a stretch.
4. Forgetting the interest-rate channel. Late-adopter pain isn’t just about missing out on AI — it’s an active financing headwind imposed by everyone else’s AI investment, transmitted straight through global capital markets.
5. Choosing targeted vs. universal transfers on principle instead of arithmetic. The right answer is empirically country-specific, not a matter of ideology.

10. Board / Investor Checklist

• Map country exposure against structural readiness — capital intensity, skills, productivity — not against headline AI-adoption announcements.
• For EMDE-facing portfolios, model the financing-cost headwind during the pre-adoption window, not just the delayed-upside case.
• Where you have any policy influence, push for reforms that move skills, labor productivity, and capital productivity together — piecemeal reform delivers far smaller gains in this model than a coordinated push.
• Stress-test workforce and compensation plans against a generational income divide, not just the skill divide most plans already account for.
• Treat every number in this paper as directional. The authors say so themselves, and any strategy built on it should say the same.
• For healthcare and life sciences portfolios specifically, sequence market entry and capex by country-and-segment AI readiness — not sector-wide — since specialist-heavy tertiary care and MedTech manufacturing will diverge from front-line delivery and EMDE diagnostic chains on very different timelines.
• For India specifically, weight the country’s real capital-account controls and informal-labor scale into any adoption-timing model — don’t take the paper’s small-open-economy, formal-labor-market assumptions at face value for an economy this large and this heterogeneous.

11. What the Authors Admit They Left Out

Credit where it’s due — the paper spends a whole section on its own limits. It doesn’t capture the rents that flow to countries that produce AI (chips, IP, platforms), which likely go to the same economies already positioned to adopt early — meaning real-world divergence could end up worse than modeled here, not better. It treats capital as one homogeneous stock, when in reality a lot of existing infrastructure isn’t easily repurposed for AI. It assumes capital moves perfectly freely across borders, a much safer assumption for small economies than for giants like China or Japan. And it leaves out trade, exchange rates, and sector-by-sector differences entirely — channels that could push divergence either wider or narrower depending on how they play out.

The authors’ own framing is worth taking seriously: read this as an illustration of a mechanism and its direction, not a forecast of any one country’s path.

Selected References

1. Benzell, S. G., Kotlikoff, L. J., LaGarda, G., & Ye, V. Y. (2021). Simulating Endogenous Global Automation. NBER Working Paper №29220.
2. Cazzaniga, M., Jaumotte, F., Li, L., Melina, G., Panton, A. J., Pizzinelli, C., Rockall, E. J., & Tavares, M. M. (2024). Gen-AI: Artificial Intelligence and the Future of Work. IMF Staff Discussion Note SDN/2024/001.
3. Acemoglu, D., & Restrepo, P. (2019). Automation and New Tasks: How Technology Displaces and Reinstates Labor. Journal of Economic Perspectives, 33(2).
4. Moll, B., Rachel, L., & Restrepo, P. (2022). Uneven Growth: Automation’s Impact on Income and Wealth Inequality. Econometrica, 90(6).
5. Autor, D. H., Levy, F., & Murnane, R. J. (2003). The Skill Content of Recent Technological Change. The Quarterly Journal of Economics, 118(4).
6. Jaumotte, F., et al. (2026). Bridging Skill Gaps for the Future: New Jobs Creation in the AI Age. IMF Staff Discussion Note SDN/2026/001.
7. Che, N., Diez, F. J., Oeking, A., & Xin, W. (2025). Dissecting Medium-Term Growth Prospects for Asia. IMF.
8. Budina, M. N., et al. (2023). Structural Reforms to Accelerate Growth, Ease Policy Trade-Offs, and Support the Green Transition in EMDEs. IMF.
9. Che, N., Xin, W., & Yoshida, T. (2026). AI and Economic Divergence in Asia. IMF Working Paper WP/26/166.

The Great AI Fork: What the IMF’s New Model Says About Asia’s Next Two Decades was originally published in Bootcamp on Medium, where people are continuing the conversation by highlighting and responding to this story.