For the last few weeks I have been watching economists and mathematicians on YouTube, reading Acemoglu papers I half-understood, and bookmarking IMF and Goldman Sachs PDFs I told myself I'd get back to. The question that pulled me in was simple, and it's the one everyone is asking right now: will AI destroy jobs, or will it make us richer than we've ever been?
Most answers I found were either screaming doom ("everyone is going to be unemployed by 2030") or screaming utopia ("we'll cure cancer and work 4 hours a week"). Both felt lazy. Neither side was doing math. So I did something I had been thinking about for a while — I asked an AI to write me a long, mathematically grounded research paper on the macroeconomics of AI itself. No vibes. No takes. Just task-based labor models, Hulten's theorem, CES production functions, and probability-weighted scenarios.
The result is a 62-page institutional research note. This article is my attempt to translate that note into plain English — for anyone who wants the mathematical and economic view on AI rather than the philosophical one.
Spoiler: the answer isn't doom. It also isn't utopia. The honest answer is much more interesting than either.
The one equation everything reduces to.
If you remember nothing else from this post, remember this single equation. Every serious forecast — Acemoglu's pessimistic 0.66% boost, Goldman Sachs' optimistic 7% GDP uplift, McKinsey's huge $4.4T/year estimate — boils down to the same four numbers multiplied together:
| Symbol | Parameter | Meaning | Acemoglu (conservative) | Goldman/McKinsey (aggressive) |
|---|---|---|---|---|
| τ | Task exposure | Share of economic tasks AI can touch | 0.20 | 0.50 |
| π | Profitably automatable | Of exposed tasks, how many are worth automating | 0.23 | 0.50 |
| σ | Cost savings | Average cost reduction per automated task | 0.27 | 0.30 |
| φ | Spillover multiplier | Indirect productivity gains from new tasks/products | 1.00 | 1.50 |
| ΔTFP = τ × π × σ × φ | 0.66% | 11.3% | ||
The single equation that drives every AI macro forecast in the literature. Same equation, different inputs — 17× difference in output.
That's it. That's the whole debate. Plug in conservative numbers (Acemoglu: 0.20 × 0.23 × 0.27 × 1.0) and you get a tiny 0.66% productivity boost over a whole decade. Plug in aggressive numbers (Goldman/McKinsey: 0.50 × 0.50 × 0.30 × 1.5) and you get 11.3% — almost 17 times bigger.
The forecasters who say AI will change everything and the ones who say it will barely move the needle are not disagreeing about theory. They are using the exact same equation. They are disagreeing about four numbers. That's a much smaller fight than the headlines make it sound.
The "micro–macro gap" — why studies show huge gains but GDP barely moves.
Here's the puzzle that broke my brain a little. Every controlled study of AI at work shows enormous productivity gains:
- Call centers (Brynjolfsson, Li, Raymond, QJE 2025): +14% on average. +34% for new hires.
- Coding with GitHub Copilot (Peng et al.): +55.8% faster task completion.
- Writing tasks (Noy & Zhang, Science 2023): +37% speed, +40% quality.
- Consulting (Dell'Acqua et al., HBS): +12% to +40% on tasks inside AI's "frontier."
And yet US TFP (Total Factor Productivity — the catch-all measure of how efficiently the whole economy turns inputs into outputs) shows no measurable AI signal yet. How can individual workers be 30–50% more productive while the aggregate economy looks the same?
The answer is Hulten's theorem, and it's the most important piece of economics in this whole debate. Hulten says: an economy's aggregate productivity gain equals the share-weighted sum of all the task-level gains.
In plain English: AI making one slice of the workforce 50% more productive doesn't do much if that slice is only 5% of the economy's wage bill. Imagine a kitchen where one chef gets a magic knife that chops vegetables 10× faster. The restaurant doesn't serve 10× more meals — chopping was 2 minutes out of a 45-minute cook time. Bottleneck moves elsewhere.
Roughly 15–25% of the advanced-economy wage bill is in tasks AI can meaningfully touch today. Multiply that by an average 27% cost saving and you get ~4% — and that's before you discount for adoption friction, regulation, and the fact that most enterprise AI projects fail at the integration stage. Hence Acemoglu's famously low 0.66% over a decade.
This is also why I now distrust any forecast that doesn't tell me which value of τ (tasks exposed) it's using. If you don't pin that number, you can produce any headline you want.
What history says: the Solow paradox is the most likely outcome.
In 1987, Robert Solow famously said: "You can see the computer age everywhere but in the productivity statistics." Factories had computers. Offices had computers. And yet US TFP growth from 1973 to 1995 was a sad 0.4% per year.
Then suddenly, from 1995 to 2004, TFP growth jumped to 1.4%. Then by 2005 it collapsed back to 0.3%. The computer revolution produced a single decade of measurable boom, and it took 20 years to show up.
Why the lag? Because new technologies need complementary investments — reorganizing factory floors, rewriting business processes, training workers, waiting for old capital to depreciate. Paul David documented this for electrification: factories had electric motors by 1900 but productivity gains didn't appear until the 1920s, when factory layouts were redesigned around unit-drive electric motors instead of the old steam-belt arrangements.
We are roughly 2–3 years into the AI equivalent. Expect a J-curve: worse before better, with the real macro productivity boom probably arriving 2032–2038. We will spend the next 5–7 years hearing both that "AI failed to deliver" and that "AI is everywhere" — both will be partially correct.
The seven scenarios — how to think about uncertainty without lying.
Anyone who claims a precise number for what AI will do to GDP by 2045 is either confused or selling something. The honest move is to spell out a small set of scenarios, assign probability weights to each, and average. That's what serious institutional research does, and it's what I asked the AI to do.
Here are the seven scenarios, with the probability weights and what each one means:
| Scenario | Prob. | GDP 2035 | Δ Unemployment | Labor Share |
|---|---|---|---|---|
| S1 — Minimal disruption | 10% | +0.6% | +0.5 pp | −1 pp |
| S2 — Moderate productivity boom | 30% ← modal | +3.5% | +1.2 pp | −3 pp |
| S3 — Severe labor shock | 15% | +8.5% | +4.0 pp | −6 pp |
| S4 — Hyper-productivity | 15% | +17% | −0.5 pp | −4 pp |
| S5 — AI oligopoly world | 15% | +10% | +2.0 pp | −9 pp |
| S6 — Deflationary collapse | 10% | +13% | +5.5 pp | −8 pp |
| S7 — Post-labor high-GDP | 5% | +45% | +12 pp | −20 pp |
| Probability-weighted mean | — | +8.4% | +2.1 pp | −5.4 pp |
The seven AI macro scenarios with probability weights. The probability-weighted GDP (+8.4%) exceeds the median (+4.2%) because the right tail (S4, S7) pulls the mean up. Mass unemployment is NOT the central case.
A few things jump out when you actually stare at this table for a minute:
- The single most likely scenario (30%) is a moderate productivity boom — basically a compressed replay of the 1995–2004 IT era. Real, but not transformative.
- The "AI takes all the jobs" scenarios (S3, S6, S7) collectively only total 30% probability. They are the tail, not the mode.
- The "AI does nothing" scenario is just 10% — almost everyone in serious economics agrees AI does something; the question is what kind of something.
- Across every non-trivial scenario, the labor share of GDP falls. That is the most robust prediction in this entire analysis.
The headline numbers.
After running 10,000 Monte Carlo draws through this scenario distribution, here is what the probability-weighted forecast actually looks like, with honest 90% confidence intervals:
| Horizon | Central GDP uplift | 90% confidence band |
|---|---|---|
| By 2030 (5 years) | +2.0% | +0.4% to +5.1% |
| By 2035 (10 years) | +4.2% | +1.1% to +11.4% |
| By 2045 (20 years) | +9.8% | +2.0% to +28.5% |
| By 2065 (40 years) | +21.5% | +3.0% to +70%+ |
Probability-weighted GDP impact relative to a no-AI baseline. Uncertainty grows roughly as the square of the time horizon.
The way to read this: in the next 5 years, AI probably adds about 2% to global GDP relative to where we'd be without it. That's roughly a year of normal growth, compressed. Not nothing, not transformative. By 2045, the central forecast is +10% — meaningful, but the band stretches from "barely noticeable" (+2%) to "civilizational" (+28%).
The thing nobody wants to talk about — the labor share.
Mass unemployment makes headlines. The labor share doesn't. But the labor share is where the real economic story is.
The labor share is the fraction of GDP that goes to workers as wages, as opposed to capital owners as profits, dividends, and rents. Globally it has been falling since the 1980s — by about 5–7 percentage points. Karabarbounis and Neiman showed this is mechanically driven by capital becoming cheaper relative to labor, combined with an elasticity of substitution between capital and labor of about η ≈ 1.28 — meaning capital and labor are substitutes, not complements, in production.
Now consider AI. The price of one unit of "useful cognitive work" from a frontier model has dropped roughly 10× per year for the last 18 months. There has never been a cheaper capital input in modern history. The math is brutal:
| Step | Expression | Value |
|---|---|---|
| Formula | d ln(labor share) / dt = (1 − η) · d ln(p_AI / p_L) / dt | — |
| Plug in | (1 − 1.28) × (−3% / yr) | −0.84% / yr |
| Over a decade | Compounding | ~5 pp absolute drop |
Why the labor share falls when AI capital gets cheaper.
The probability-weighted forecast: labor share drops by another 3–7 percentage points over 20 years. That sounds small. It is not. A 5 pp labor share decline translates roughly into a 7–12% reduction in payroll tax bases in advanced economies — with public debt already over 100% of GDP. That's a fiscal crisis waiting to happen, not via mass unemployment but via tax base erosion.
Meanwhile the capital share of GDP grows — and not evenly. AI-rents will concentrate in maybe 5–10 firms globally. The top 1% wealth share in the US is projected to climb from 31% today to 34–38% by 2040. Top 10% income share up another 4.2 pp. The Gilded Age comparison is not a metaphor; it's the closest historical analog we have.
Who actually loses their jobs.
The previous wave of automation (computers, 1980s–2000s) hollowed out the middle of the wage distribution — bank tellers, factory workers, secretaries. The top (lawyers, doctors, managers) and the bottom (waiters, janitors) were both safe. Economists called this job polarization.
AI breaks the polarization rule. For the first time, automation comes for non-routine cognitive work — the stuff lawyers, paralegals, junior bankers, junior coders, analysts, and copywriters do. The classic "safe" white-collar jobs are now in the substitution zone. Meanwhile, plumbers, electricians, nurses, and HVAC technicians are protected by the Moravec paradox — robotics is still 5–10 years away from doing their work at competitive prices.
The sectoral picture, in rough order of pain:
- BPO / customer support: 50–70% employment reduction probable. This is essentially a solved problem with current tech.
- Junior software engineering: Hiring already down 30–40% from 2021–2022 peaks at major US tech firms.
- Junior finance / consulting / legal: Pyramid-shaped firms (Big 4, MBB, biglaw) face structural pressure on the up-or-out model. 10–25% in junior ranks at risk.
- Content / advertising: Marginal cost of content production collapsing. Copywriters, junior designers, voice-over artists.
- Manufacturing, healthcare, education: Productivity gains, modest displacement.
- Skilled trades, in-person services, in-person healthcare: Largely protected.
The brutal news for younger people entering the workforce: the "safe path" your parents told you about — get into a top university, become a lawyer or consultant or banker, climb the pyramid — is the path most exposed to AI substitution. The historically less prestigious paths (skilled trades, in-person work) are now structurally more secure. We are not used to thinking about the economy that way.
India and the BPO economies — the most exposed national franchise.
This part hit close to home. India's IT/BPO export base is $315 billion per year. By the AI's estimate, 50–55% of that is in tasks with above-median automation probability. Cumulative revenue at risk over a decade: $30–80 billion.
We are already seeing the early signal: TCS cut headcount by 13,249 in FY24 while revenue held. Infosys headcount down 5.9%. The Nifty IT index dropped 19% in a single month early in 2026. None of these are coincidences.
The Philippines is worse off proportionally — BPO is 8% of GDP and 60–70% of it sits in highly automatable categories. Bangladesh, Costa Rica, Vietnam, Poland — every country that built its development model on labor arbitrage in cognitive routine tasks is structurally exposed.
The cruel irony: the "development ladder" that India and China climbed (manufacturing → BPO → high-value services) is being kicked out from under countries currently on its lower rungs. Sub-Saharan Africa may find its development pathway foreclosed before it could even start the climb.
India's escape route, if it exists, is to pivot from using human labor to deliver AI-substitutable services toward building and exporting AI products. We have the English-speaking ML talent. We do not yet have the capital, compute, or industrial policy. Without those, India's net GDP impact over a decade ranges from −1% to +1%. With aggressive policy: +3% to +6%. That's the size of the policy bet.
The bottlenecks nobody factors into headlines.
Every Twitter take about AI economic impact assumes deployment is free and instant. It isn't. The hard physical constraints are real and binding for at least the next decade.
Electricity. Global data center electricity demand: ~415 TWh in 2024, projected to hit 945 TWh by 2030 (IEA). US data centers alone will use more electricity than aluminum, steel, cement, and chemicals production combined by 2030. Building new transmission takes 5–12 years. New nuclear takes 10–15+. The US grid has minimal slack — the typical interconnection queue is 4–7 years. AI deployment will be electricity-rationed through 2030. This is not a forecast; it is already happening.
Advanced packaging and HBM memory. TSMC's CoWoS packaging is sold out through 2026. NVIDIA pre-booked 800,000–850,000 wafers for 2026 alone. China is structurally capped at 200,000–300,000 frontier-equivalent chips per year (vs. millions for the US) for at least 3–5 years due to export controls.
Organizational adoption. McKinsey data: 60–70% of enterprise AI projects fail to deliver expected returns. Not because the models are bad, but because rewriting business processes is hard, slow, political, and ego-bruising. This is the modern Solow paradox in real time.
Trust and verification. Frontier models still hallucinate 1–5% on factual questions. In medicine, law, finance, infrastructure, the cost of verifying AI output partially offsets the productivity gain. Many "AI productivity gains" measured in narrow studies don't survive the move to high-stakes deployment.
Inflation, interest rates, and the central bank's headache.
Here's something most people miss: AI is disinflationary in the things AI can do (knowledge services, software, content, basic analysis) but inflationary in things it can't (housing, in-person care, electricity, copper, GPUs, skilled construction labor).
Net first-order CPI impact through 2030, by my best estimate: −0.3 to −0.8 percentage points per year in advanced economies. This is the largest structural disinflation since the China shock of 1995–2007. Central banks are going to struggle to hit their 2% inflation targets without running policy easier than would otherwise be appropriate.
The flip side: real interest rates rise about 50–100 basis points because higher productivity growth and the AI capex boom both push up the natural rate of interest. So we get a strange combination — lower inflation, higher real rates, elevated asset prices (especially in AI-platform equity), and a tax base that's shrinking faster than the corporate tax base is growing.
What would change the forecast.
One thing I respect about how this paper was framed: it commits to explicit reversal triggers. These are the data points that, if they happen, would force a serious rethink.
- US TFP growth sustained above 2% per year for 3+ consecutive years → upgrade S4 and S7 (the hyper-productivity scenarios).
- Hyperscaler capex drops more than 30% year-over-year for 2 quarters → upgrade S1, downgrade S4/S7. The capex bubble would be deflating.
- Global labor share recovers by more than 2 pp over 3 years → downgrade the AI-oligopoly and deflationary-collapse scenarios.
- China achieves indigenous 3nm AI chip production at >1M units/year → completely rewrite the geopolitical section.
- A general-purpose robot deploys at < $50K/year all-in cost → S7 (post-labor) becomes much more likely.
None of these have happened yet. Several could happen within 2–3 years. That's why I plan to revisit this post sometime in 2027 and see how much of it I want to retract.
The strongest counterargument — and why I take it seriously.
The single best argument against the central forecast above is the Solow paradox combined with organizational rigidity. The case in plain English:
We have been here before. The 1980s and early 1990s were full of breathless predictions about computers transforming the economy. The transformation happened, eventually — but it was smaller than predicted, slower than predicted, and ended sooner than predicted. From 1973 to 1995, despite massive IT investment, US TFP grew at 0.4% per year. The "computer revolution" produced exactly one decade of measurable boom (1995–2004), then dissipated even as Moore's Law continued.
Many AI applications, per Acemoglu's "so-so automation" hypothesis, will automate tasks at slightly lower quality and slightly lower cost, displacing labor without commensurately raising productivity. That produces inequality without growth — the worst of both worlds.
Add Baumol's cost disease (sectors AI can't touch — care, housing, education, in-person services — become a larger share of the economy, dragging average productivity growth down) and the energy/compute bottlenecks above, and you get a picture where AI is real, important, but ultimately a moderate boost — closer to S2 than to S4 or S7.
The paper gives this counterargument roughly 30% aggregate weight in the probability distribution. That feels right to me. Anyone weighting it at 0% (the doomers) or at 90% (the AI skeptics) hasn't actually engaged with the other side.
What I'd do if I were running a country, a company, or a career.
The institutional posture the paper recommends is simple but useful, and I think it generalizes well: plan for the central case, hedge for the right tail, build resilience against the left tail.
If you're a government: Modernize your tax base now — shift from labor income taxes toward capital and consumption. Invest in reskilling. Build safety nets that can absorb 2–4 percentage points of transitional unemployment. If you're an AI-laggard country, don't try to build a frontier model. Import the capability, focus your capex on energy, broadband, education, and the infrastructure that lets adoption actually happen.
If you're a company: Deploy AI aggressively, but invest in the complementary stuff — process redesign, data quality, training. The companies that win the next decade aren't the ones with the best models, they're the ones whose business processes are already AI-shaped before the rest of the industry figures it out.
If you're a worker: Three pieces of practical advice from the math. One — get good at using AI to extend your own productivity, because the wage premium for AI-fluent workers in any field is already showing up in the data. Two — be very careful before entering a high-substitution-risk career path (BPO, routine cognitive, junior pyramid-firm roles). Three — the trades, in-person care, and roles requiring tacit human judgment are structurally protected for longer than the prestige hierarchy would suggest.
If you're me (a DevOps + ML apprentice): Build the skills that orchestrate AI rather than compete with it. Learn the infrastructure layer. Learn how models are deployed, evaluated, monitored, and integrated with the rest of the business. That layer is growing, not shrinking, and the people who can bridge ML and ops are some of the most valuable workers in the economy right now. This is also partly why this blog exists.
Closing — the honest answer.
So: will AI destroy jobs or create prosperity?
The honest answer, after going through 62 pages of math: both, partially, and unevenly.
AI will not produce mass unemployment in the literal sense. The probability of sustained 15%+ unemployment in major economies over the next 20 years is around 15%. The labor market reallocates, painfully, but it reallocates. That's the historical record and it is the most robust prediction in the analysis.
AI will produce prosperity — a probability-weighted ~+10% GDP uplift over 20 years. But the gains will accrue disproportionately to capital owners, AI-complementary high-skill workers, and a small number of countries (US dominant; China, UK, Israel, Singapore, Gulf states secondary). The labor share keeps falling. The top decile captures most of the upside.
So the headline question — destroy jobs or create prosperity — is the wrong question. The real question is: can our institutions distribute the gains widely enough, fast enough, that the transition doesn't produce a populist backlash that breaks the system before the prosperity arrives?
That is a political question, not a mathematical one. The math just tells you the size of the bet.
The defensible institutional posture is: plan for the central case, hedge for the right tail, build resilience against the left tail. The macroeconomy is about to undergo its most important structural transformation since the Industrial Revolution. The mathematics of growth, distribution, and capital all change. Institutions, policies, and investments designed for the pre-AI economy will systematically underperform.
Adaptation is not optional.
A note on method: this post is a translation of a 62-page institutional research note I had an AI write for me on the macroeconomics of AI. I read the full paper, fact-checked the numbers I could fact-check, and rewrote everything in my own words and structure. What you read here is a synthesis — the AI did the heavy lifting on the math and the literature review; I did the curation, the editing, and the opinions. If you want the full 62-page version with all the equations, footnotes, and references, ping me and I'll send it. It's worth a read if you're into the topic.