Is AI Quietly Contracting the World Economy?

Is AI Quietly Contracting the World Economy?

A Quebec translator I read about made roughly six figures in 2023. By late 2025 her work requests had dried up. Total earnings for the year barely cleared €8,000. She is forty-four years old and watching her career end through email, while large language models charge a few cents to do what she charged a few dollars per word to do.

Multiply her by a few million.

That is the whole argument I want to walk through today. Not the productivity story. The demand-side story. The one nobody at Goldman Sachs models because their clients are long Nvidia.

The Mechanism, Stated Plainly

Here is the chain.

A small business in Berlin used to pay a translator €2,000 to localise a website. Now it pays DeepL €30 a month. The translator used to spend that €2,000 on rent, groceries, a dentist, a plumber, a meal out. Each of those people spent it again. The Keynesian multiplier is not a theory you have to believe in. It is just bookkeeping with a delay.

Strip the translator out, and the €2,000 does not vanish. It moves. It moves to a hyperscaler in Virginia, to Nvidia in Santa Clara, to TSMC in Hsinchu. It buys GPUs and electricity and circulating coolant. Some of it gets paid out as dividends to the top wealth decile, who already own everything they need.

The marginal propensity to consume of a Frankfurt translator is close to one. The marginal propensity to consume of a Microsoft retained earning is close to zero.

That is the entire thesis in one paragraph.

How Big Is The Substitution

Start with the freelance market because it is the only place we have clean data.

A study co-authored by Xinrong Zhu of Imperial College London Business School with Harvard Business School and the German Institute for Economic Research, published in Management Science, tracked 1,725,587 freelance job posts across 61 countries. Within eight months of ChatGPT’s launch, writing jobs fell 30.37 percent. Coding fell 20 percent. Graphic design fell 17 percent. The Vollna Upwork report shows writing projects down 32 percent year over year in 2025, the largest drop of any category. Eleven of twelve major work categories declined. Entry-level project share on Upwork collapsed from 15 percent to under 9 percent.

A February 2026 Ramp study called “Payrolls to Prompts” found that more than half of the businesses that spent on freelance platforms in 2022 had stopped entirely by 2025. Freelance marketplace spend as a share of total company spend fell from 0.66 percent to 0.14 percent. AI model spend went from zero to 2.85 percent of company budgets.

That is not a rotation. That is a substitution. Money that used to pay humans now pays GPUs.

The translation industry shows the same shape. CSA Research has the global language services market at $49.7 billion in 2023, down from $52 billion the year before, with per-word rates compressed 50 to 70 percent on commodity work. Machine translation post-editing pays roughly 25 percent of traditional rates for comparable time.

The BPO sector employs around 8 million people across India and the Philippines. The IMF estimates one-third of Philippine jobs are highly exposed to AI. The ILO’s 2017 ASEAN report found 89 percent of salaried workers in the Philippines BPO sector fall into the high-risk category for automation. Industry analysts now project 1 million BPO and IT outsourcing roles will be impacted by 2030, with 2 to 3 million workers seeing major disruption. TCS cut 12,000 roles in its largest reduction ever. India’s top IT firms added a net 17 employees in the first nine months of fiscal 2026. Seventeen.

Junior coders who used to be the on-ramp into the global tech middle class are getting fired by email.

The Local Economy Loses Twice

When a Manila call center agent earning $400 a month gets replaced by an AI agent, two things happen.

First, $400 a month leaves the local economy. The sari-sari store, the jeepney driver, the landlord. Gone.

Second, the dollars now flow to OpenAI, Anthropic, Microsoft Azure, AWS. Most of that revenue ends up paying for Nvidia chips made in Taiwan and electricity in Virginia. None of it lands in Quezon City.

Apply this across translators in Madrid, copywriters in Belgrade, illustrators in Buenos Aires, junior developers in Bengaluru. The capital concentration is not subtle.

The IMF says nearly 40 percent of global employment is exposed to AI, rising to 60 percent in advanced economies. Goldman Sachs said in 2023 that AI could substitute for up to a quarter of current work, exposing 300 million jobs in the US and Europe. McKinsey put the addressable annual value at $2.6 to $4.4 trillion across 63 use cases.

Read those numbers as someone other than a sell-side analyst. They are mostly a transfer estimate, not a wealth creation estimate.

What The Optimists Say

I have to do this part honestly because the optimist case is not stupid.

Goldman Sachs models a 7 percent boost to global GDP over a decade and a 1.5 percentage point lift to US labor productivity. The Briggs and Kodnani paper from 2023 is still the consensus bull take. McKinsey is in roughly the same neighborhood. The IMF in January 2026 raised its global growth forecast for the year to 3.3 percent, citing AI investment as a key driver. Marcello Estevão at the IMF wrote in March 2026 that AI-related investment now accounts for a large share of US GDP growth.

This part is empirically true. Apollo Global Management calculates hyperscaler capex around $646 billion in 2026, roughly 2 percent of US GDP. Bank of America puts AI capex from the top five hyperscalers at $399 billion in 2025, rising past $600 billion. White House AI Czar David Sacks, citing a May 2026 Morgan Stanley report, said AI was 75 percent of US Q1 2026 GDP growth. Pantheon Macroeconomics found that without AI capex, US private fixed investment in 2025 would have been negative.

So yes. The buildout is currently keeping the American economy out of recession.

The Jevons paradox argument is also real. Cheaper code means more code gets written. Cheaper translation means more content gets localised. AI-specialised freelancers on Upwork command 25 to 60 percent higher rates than general practitioners. Per Upwork’s Q4 2024 earnings release, freelancers working on AI-related projects earned 44 percent more per hour than those on non-AI work. CEO Hayden Brown calls these the “AI generalists” and says they are the most sought-after profile in the ecosystem.

The historical analogues are well-rehearsed. ATMs were supposed to kill bank tellers. Teller employment grew for two decades after ATMs because cheaper branches meant more branches. The internet was supposed to kill travel agents. It did kill them, but it created software developers, SEO specialists, growth marketers. Disruption with reinstatement.

What Is Different This Time

Three things, and I am going to be blunt.

First, Acemoglu. The Nobel laureate ran the same task-based math and got a 0.66 percent total factor productivity bump over ten years, and a GDP gain between 1.1 and 1.6 percent. His point is not that AI does nothing. His point is that only about 5 percent of tasks can be profitably automated in the near term, and the macroeconomic effect of automating them is small, while the displacement effect on the workers doing those tasks is concentrated and severe. He explicitly notes that “productivity gains from AI are unlikely to lead to sizable wage rises.” Other recent work pushes harder. Falk and Tsoukalas’s “AI Layoff Trap” model concludes that “at the limit, firms automate their way to boundless productivity and zero demand.” Korinek and Stiglitz model a temporary labor demand collapse. Roubini, in a Bloomberg interview this February, invoked Marx directly: workers losing wage share lose the purchasing power to buy what the economy produces.

Second, the AI companies are losing money. OpenAI posted a $13.5 billion net loss in the first half of 2025 alone, against $4.3 billion in revenue. Internal projections show 2026 losses of $14 billion against $13 billion of sales, with operating losses of $74 billion in 2028. HSBC analysts estimate OpenAI faces a $207 billion funding shortfall through 2030. Every dollar a translator loses is not currently a profitable dollar in OpenAI’s pocket. It is a subsidised dollar, paid for by SoftBank, Microsoft, Nvidia, sovereign wealth funds, and the bond market.

The circular financing is not subtle. Nvidia announces a $100 billion investment in OpenAI, OpenAI buys Nvidia chips. Oracle commits $300 billion in cloud to OpenAI, then plans roughly $40 billion in Nvidia GB200 chips to deliver it. CoreWeave gets equity from OpenAI, rents Nvidia GPUs back. Michael Burry is shorting it. Sam Altman has called it a bubble. Ray Dalio compares it to dot-com.

If the bubble pops, the displaced translator does not get her job back. The substitution is sticky. The capital is not.

Third, the energy bill is hitting voters. The IEA expects global data center electricity consumption to roughly double to 945 terawatt-hours by 2030. US data centers alone accounted for around half of all US electricity demand growth in 2025. PJM capacity prices went from $28.92 per megawatt-day in 2024/25 to $329.17 in 2026/27 to $333.44 in 2027/28, a roughly tenfold jump, with PJM’s own market monitor Monitoring Analytics attributing $9.3 billion of the 2025/26 increase, 63 percent of it, to data centers. NRDC estimates $100 to $163 billion in cumulative PJM ratepayer costs through 2033 absent regulatory action. Pepco residential customers in DC saw bills rise around $21 a month from June 2025. AEP Ohio’s standard offer jumped 36 percent. BGE customers in Maryland saw $32 added in September 2025. Maine residential rates are up 22.6 percent year over year, though the Maine PUC chair attributes most of that to natural gas and LNG exports rather than data centers directly.

This is not just an externality. It is a transfer. Households are subsidising the buildout that is displacing their own labour income.

Read that twice.

The Geographic Map

Globally, I see roughly this shape.

The United States captures most of the upside. Hyperscaler capex is keeping GDP positive. Nvidia, Microsoft, Google, Meta, Amazon, Oracle: a handful of companies, roughly 40 percent of S&P 500 market cap, are the recipients of the wealth transfer. The labor market downside is concentrated in entry-level white-collar work. Goldman Sachs’ Briggs and Dong concluded in March 2025 that “aggregate labor market impacts are still negligible,” which is bull-case framing for “we cannot see the productivity yet either.”

The European Union takes more of the demand-side hit and less of the capex upside. Stronger labour protections slow the displacement. The ECB has flagged that productivity gains from AI are concentrated in US firms while EU consumers and SMEs absorb the cheaper substitutes. Translators, lectors, content marketers, junior architects, and contract lawyers are all under direct pressure. The EU has neither the data centers nor the model providers in serious scale.

Emerging markets get hit hardest on the labour side and gain almost nothing on the capex side. India and the Philippines lose BPO and IT services revenue. Latin American designers and developers, who served the US market through Upwork at lower rates, lose those clients to ChatGPT and Cursor. Africa never had a knowledge economy at scale to lose, but it also will not get one now.

Net global picture for 2026: the IMF’s 3.3 percent number is plausible only because AI capex is keeping the headline up. Strip out the buildout and the underlying economy is materially weaker. My own back-of-envelope: AI substitution is probably reducing global household income from professional services by $200 to $400 billion in 2026. With a multiplier of around 1.5 in advanced economies, that is $300 to $600 billion of demand drag. AI capex of around $700 billion masks it in the GDP figures. Net visible effect: small positive. Net structural effect: a wealth transfer from millions of professionals to about ten companies.

That is not contraction in the headline number. It is contraction in the median household.

What Could Change My Mind

If the productivity gains show up clearly in measured output for non-AI sectors over the next two years, Acemoglu is wrong and Briggs is right. So far, three years in, they have not.

If AI inference costs collapse far enough that the model providers turn cash-flow positive without further capital infusions, the bubble fear goes away. Anthropic looks plausible on this front. OpenAI does not.

If displaced workers reskill into AI-augmented roles at scale and recover wages, the demand drag dissipates. The Upwork data suggests this is happening at the top decile of freelancers. It is not happening at the median.

Bottom Line

AI is not contracting headline GDP in 2026. It is contracting the part of GDP that flows through millions of household budgets, and replacing it with a part that flows through ten balance sheets and a power grid.

The capex boom hides this. The bubble math hides this. The 7 percent Goldman number hides this. The Acemoglu number does not. The PJM auction does not. The TCS layoffs do not. The Quebec translator does not.

If the buildout pays off, we get a productivity boom and a redistribution problem. If it does not, we get the redistribution problem without the productivity boom.

Either way the translator is not getting her clients back.

The local economy was always the answer. We just stopped noticing.

Final Words

I would love to hear where you think I’m wrong on this. The optimists have history on their side: every wave of automation eventually created more jobs than it destroyed. I am not sure this one will, and I am not sure the timing works even if it does.

You can find me on LinkedIn at linkedin.com/in/ivanturkovic, on X at x.com/ithora, and on Threads at threads.com/@ithora. If you want to send a longer reply or push back in detail, the contact form at ivanturkovic.com/contact is the best place. If you are building something at the intersection of AI and fintech and want to talk strategy or architecture, you can book a call at cal.eu/ivan-turkovic/30min.

So tell me: in your local economy, who has lost income to AI in the last twelve months, and where did that money actually go?