AI is not fair — and that is only the beginning

If most of your answers fall on the “less” side, you are standing on the losing side of a very new inequality. The mainstream story says AI is “democratizing intelligence — anyone can use it.” Part of that is true. A bigger part is not. This essay tries to separate the two, because conflating them leaves you unable to see what is actually happening.

This essay is not about who got rich and who lost their job. It is about structure. Why the world’s top economists are deeply divided on AI and inequality. Why both sides can be right. And why I still lean toward the pessimistic camp — not for Harari’s reasons, but for a different structural reason that few people are naming.

There is a moment of self-doubt at the end. To be fair to you — and to the argument itself.

1. The argument no one is telling you about

Most readers of AI commentary are stuck between two peaks of noise: the Marc Andreessen-style hype (“AI will save the world”) and the Twitter-style doomerism (“AI will kill us all”). Both are loud, easy to share, and largely empty.

Beneath that noise, the world’s top economists are arguing seriously about a much more specific question: does AI amplify inequality, or narrow it? This argument has data, methodology, peer review. And it is going in two opposite directions.

Worth devoting a section to it — because if you do not understand how two camps look at the same data and arrive at opposite conclusions, you will not understand why this essay leans the way it leans.

The optimist camp has real data

I am not talking about Andreessen. I am talking about empirics.

Brynjolfsson, Li and Raymond (NBER 2023) tracked 5,179 customer support workers at a Fortune 500 company using an AI assistant in real work. Average productivity rose 14%. The crucial finding: new workers improved by 34%, veteran workers by close to zero. AI narrowed the gap between novices and experts; it did not widen it.

Noy and Zhang (Science 2023) ran an experiment with 453 college-educated professionals on writing tasks: time fell 40%, quality rose 18%, and the productivity gap between high- and low-skilled workers narrowed — it did not widen. Democratization again.

David Autor (MIT, 2024) in “Applying AI to Rebuild Middle-Class Jobs” argues AI may be the first technology in 40 years capable of rebuilding the middle class — by giving expertise to workers without elite credentials. The exact opposite of the “hollowing out the middle” thesis I will argue below.

Daron Acemoglu (MIT, 2024) in “The Simple Macroeconomics of AI” recalculates the hype: productivity growth from AI is roughly 0.66% over 10 years — 5–10× smaller than the popular numbers. If Acemoglu is right, my entire “AI changes everything” framing is overblown.

These are not hot takes. They are four peer-reviewed studies from leading economists, each one a piece of counter-evidence to the argument I am about to make.

The pessimist camp also has data, not just Harari

Eloundou, Manning, Mishkin and Rock (OpenAI 2023) in “GPTs are GPTs” estimate that 80% of US workers will see at least 10% of their tasks affected by large language models, 19% will see at least 50%. This is exposure, not automatic replacement — but exposure is the necessary condition for replacement.

Cazzaniga and colleagues (IMF 2024) in “Gen-AI and the Future of Work”: 40% of jobs globally, 60% in advanced economies sit in a zone exposed to significant automation or augmentation. The paradoxical point: the more developed the country, the more exposed — because knowledge work makes up a larger share of the economy.

Korinek and Stiglitz (NBER 2020) model AI-and-inequality scenarios. In most scenarios — without redistributive policy — AI raises inequality. But the authors stress: this is not destiny. There is policy room.

Why two camps look at the same data and reach opposite conclusions

The answer, in my reading, is time horizon mismatch.

The optimists measure over 6–18 months, at the level of individual jobs. Junior agents improve. Average writers write better. New developers ship more. All of this is true and measurable today.

The pessimists measure over 5–20 years, at the level of system structure. Who owns compute. Who owns the models. Who owns the data. Where productivity gains flow. How the middle of the labor market quietly disappears. Also true — but only legible after 5+ years.

Both can be right at once. AI can simultaneously narrow the short-term gap at the task level and widen the long-term gap at the system level — because the new gap is about compute and capital, not skill.

This is the argument I will develop over the next five sections. I lean pessimist, but for reasons of compute concentration rather than rapid job replacement.

2. Compute inequality: the hard evidence

For most of economic history, intellectual capability was something you could not buy directly. You could buy books, hire great teachers, attend great schools. But in the end, you still had to use your own brain.

AI ended that.

Today, an individual can buy intellectual capability the way they buy electricity. So many kilowatt-hours buy so many results. The gap between subscription tiers is not “a slightly better version” — it is productivity leverage by an order of magnitude, which I will call compute inequality.

Compute inequality is qualitatively different from every prior wave of technological inequality. For three reasons.

Reason 1: Compute is a new factor of production — and it is concentrating frighteningly

Sevilla and colleagues (2022) in “Compute Trends Across Three Eras of Machine Learning” tracked the compute used to train the largest AI models from 1950 to 2022. The finding: from 2010 onwards, training compute doubled every six months — far outpacing Moore’s Law (doubling every 18–24 months).

By 2026, a frontier model requires a training run worth $100M or more. The number of organizations on Earth capable of doing this: no more than 5–7. OpenAI, Anthropic, Google DeepMind, Meta, xAI — perhaps Mistral, perhaps DeepSeek. That is the entire list.

Compute concentration

A factor of production that has never been this concentrated in modern economic history

4–5×

Annual growth rate of training compute since 2010 — far above Moore's Law (~1.4×).

~5–7

Organizations worldwide capable of training a frontier model ($100M+ training run).

>$100M

Minimum capital for a frontier training run — up roughly 100× in five years.

For comparison: anyone can write a book. Many people can shoot a film. A handful of companies can build a car. Five organizations on Earth can train a frontier model.

This is a degree of concentration in a factor of production that has no precedent in modern economic history. Oil, electricity, the internet — all were far more distributed at the top. Yanis Varoufakis in Technofeudalism (2023) calls this structure “cloud capital” and argues we have already exited classical capitalism into a new feudalism of cloud infrastructure. I think Varoufakis is somewhat extreme, but the cloud capital metaphor for AI infrastructure is precisely right.

Reason 2: The optimists are right at the task level, wrong at the system level

Brynjolfsson shows junior agents +34% in productivity when using AI. This is a fact. But gaining 34% in productivity while standing on an AI platform someone else owns is not democratization — it is rented productivity.

A thought experiment: if the provider raises API prices 10× tomorrow, or revokes access, or changes the terms of service — does that junior agent still keep the +34%? No. The productivity gain depends on who supplies the compute.

This is the point Karl Marx might have written into Das Kapital if he had lived to 2026: workers using AI resemble 19th-century factory workers — far more productive than the artisans they replaced, but with no ownership of the means of production. History is fairly consistent: productivity gains without ownership of the means of production lead to temporary income increases, which then get eaten by competitive pressure and supplier price extraction.

Reason 3: Four tiers — not a linear scale, but a leverage curve

Compute leverage curve

Tier upgrades are not linear — leverage rises exponentially

Free-tier ChatGPT users get half-baked answers, capped at a few dozen turns a day, with no long context and no ability to handle anything sustained.

People paying $20 a month get a better model, more turns, basic web search and file handling.

People paying $200 a month — like me — get the strongest models, deep reasoning, effectively unlimited turns, web search, file creation, computer use, MCP connectors, memory, projects, past chats. Claude Code can rewrite an entire repository, ship a product in a week instead of three months. Plus an API to build custom automations, agents, and workflows that run 24/7 without sleeping.

Enterprises paying $20,000 a month and up get all of the above, plus models fine-tuned on private data, SLAs, on-premise deployment, and a solution architect team.

This is not a linear scale. It is a leverage curve. With the same hour of work, someone on a high tier produces value an order of magnitude greater than someone on the low tier — not because they are smarter, but because their tools are several rungs more powerful.

Compute inequality does not replace older inequalities. It stacks on top of them.

3. Eight axes of inequality (each with data)

When the conversation turns to “who benefits from AI,” it usually stops at “engineers who can code.” That view is too narrow. There are at least eight axes splitting now — and each has a piece of data describing how serious the split is.

Axis 1 — Technical. Not every developer wins, only developers who can build with AI. The 2024 Stack Overflow Developer Survey shows 76% of developers use AI daily, but only ~20% report “high trust” in the output. The gap between shallow and deep users is large — and it will decide who keeps a job over the next 18 months.

Axis 2 — Domain expertise. AI amplifies existing judgment. Someone with 15 years of experience using AI works many times faster — because they know which questions to ask, which results to trust, when AI is wrong. Newcomers get help with the basics, and the basics are exactly what is being replaced. Note: this seems to contradict Brynjolfsson’s finding (juniors gaining more than seniors). The two are not in conflict — they apply to two different task types. Juniors win on structured tasks (scripted customer support). Seniors win on open-ended tasks (strategy, complex judgment). AI narrows the gap on type 1 and widens it on type 2. Work is gradually shifting toward type 2.

Axis 3 — Private data ownership. Without your data, AI is mostly useless to your business. According to McKinsey 2024, 70% of GenAI’s potential value sits in use cases that require private enterprise data — not public prompting. Data-rich firms (banks, telecoms, large retailers) are pulling ahead twice as fast.

Axis 4 — English. Models are trained mostly in English. English output is markedly higher in quality — multilingual benchmarks show the strongest model losing 15–25% of performance between English and Vietnamese / Indonesian / Hindi on the same task. A 30-word English prompt typically produces better results than a 100-word Vietnamese prompt with the same intent. The EF English Proficiency Index 2024 places Vietnam in the “moderate” band — close to Mexico, below Indonesia, well below the Philippines. This is not anyone’s personal failing. It is structure.

Axis 5 — Capital. $200 a month for one tool feels normal in the US — about 0.3% of average monthly pay. For a Vietnamese person earning 15M VND a month, $200 is 32% of income. Expensive tier = fewer accessing it = higher advantage tier.

Axis 6 — Time to learn. AI changes every week. New tools ship every day. Anyone working two jobs, hustling for daily bread, with no time budget to learn new tools will fall behind — and the gap grows by simple addition.

Axis 7 — Infrastructure. Stable internet, international Visa or Mastercard, no geographic blocks. Many Vietnamese cannot buy subscriptions even with money — OpenAI does not yet support local payments, or domestic debit cards get rejected. This is a concrete infrastructure barrier, not an abstract one.

Axis 8 — Network and taste. Young digital natives, sitting inside the Twitter/X tech bubble, run a year ahead of outsiders on information. AI generates a lot — only those who can choose what is good can use it well. Taste accumulates only through time, exposure, and community — it cannot be bought, cannot be crash-learned.

This is not complaint. It is structural description. And the structure is splitting fast.

4. Three crises — not future, already happening

A century of promises and realities

Every revolution promised liberation. Every one stratified instead.

1. 1930
   Keynes

   Predicted humanity would work 15 hours a week by 2030. Productivity grew exactly as he calculated. Today’s knowledge worker still works 50+.

2. 1995
   Internet age

   Promised to democratize knowledge. In practice it created a digital divide that has never closed — those with bandwidth are a decade ahead of those without.

3. 2023
   ChatGPT moment

   Promised AI for everyone. The $20 / $200 / $20,000 split began in months. Nvidia, OpenAI, Anthropic valuations exploded.

4. 2026
   Today

   Compute inequality plus three parallel crises. Those with advantages widen the gap faster than they realize.

Compute inequality by itself is not the problem. The problem is what it leads to: three parallel crises. None is “maybe in 20 years.” All three are already underway, at different intensities, and accelerating.

The economic crisis

The mechanism is straightforward. AI raises productivity → goods get cheaper → profits accrue to the owners of the technology rather than to labor. The middle class loses income → consumer demand collapses. The paradox: more is produced, more cheaply, more abundantly — but no one has money to buy it.

This is the pattern Thomas Piketty identified in Capital in the Twenty-First Century (2014): US labor productivity rose strongly from 1979 to 2020; real average wages rose only ~17%. Most of the productivity gain flowed to capital. And that is before AI. AI will only accelerate it.

Wealth concentrating in capital

AI valuations explode while entry-level hiring shrinks

$4T+

Nvidia market cap — first company in history to cross four trillion dollars.

Mid-2025

$300B+

OpenAI valuation — roughly 10× growth in two years.

2025

−30 to −40%

Decline in junior software engineer job postings in the US.

Stanford AI Index

Big Tech has laid off more than 600,000 workers since 2023 despite record profits — because AI lets the same output ship with fewer people. Productivity gains are pooling in very few places.

The jobs crisis

It is happening now.

The IMF (Cazzaniga 2024) estimates 40% of jobs globally, 60% in advanced economies, have at least a substantial fraction at risk of automation or augmentation. The paradoxical point: the more developed the country, the more exposed — because knowledge work makes up a larger share of the economy.

Junior developers, content writers, illustrators, translators, paralegals, tier-1 customer support — being replaced or downpriced. BPO operations in the Philippines, India, Vietnam — the knowledge-work export industries of developing economies — are the most directly threatened.

The macro pattern: hollowing out the middle. The middle of the labor market disappears. What remains is the top (strategy, relationships, complex judgment) and the bottom (manual work AI cannot yet do). The middle — historically the pillar of the middle class — is eaten from both sides at once.

David Autor (2024) believes AI may reverse this pattern — by giving expertise to workers without elite credentials, “rebuilding” the middle class. I hope he is right. But the historical track record of technological disruption does not support this scenario: after the industrial revolution, it took 50–80 years and a strong labor movement to redistribute productivity gains to workers. This time, there is no labor-movement equivalent. There is no strong politics on labor’s side.

The crisis of meaning

The most dangerous one, the least discussed, and (in my view) the most likely to materialize.

For 200 years, “work” has defined personhood. Who you are = what you do. When AI does most things better than you — including the work you love, studied for a decade, poured your passion into — the question “who am I” goes hollow.

University students ask each other “what am I studying for if AI does it better?” Artists, writers, musicians fall into depression because AI generates in seconds what they spent a lifetime learning to do. The feeling of being good at something — the feeling that gives humans a sense of worth — is being eroded at a scale unprecedented in history.

Daniel Susskind in A World Without Work (2020) distinguishes three losses from automation: income loss (solvable through universal basic income), skill loss (solvable through retraining), and meaning loss — for which there is essentially no policy solution. The third loss is the one arriving now, with no safety net for it.

Carl Frey at Oxford in The Technology Trap (2019) adds a historical warning: in the transition phase between two technological revolutions, societies always pass through severe political instability — because people who lose meaning typically seek substitute meaning in extremist ideology. Populism, ultranationalism, conspiracism — all find fertile ground in communities that feel abandoned by the new economy.

Personally, I think Harari may be too pessimistic on a few points — and he has shifted focus in his newer work Nexus (2024), no longer placing the useless class at the center as he did in Homo Deus. History has proven many times that humans always find new meaning. But the core question he raises cannot be ducked: in a world where AI does everything better, what do humans live for?

This is no longer a philosophical question. It is becoming a practical one for our children — the kids starting first grade in 2026 will graduate university into a labor market that may no longer exist in the present sense.

5. Vietnam: data and three portraits

I think the world is in the early phase of a “disordered but self-correcting” scenario, with a real risk of sliding into an “irreversible concentration of power” scenario without good policy. Vietnam is especially exposed for four structural reasons.

The scale of the problem is concrete.

Knowledge-work exports

Three economies sitting in AI's line of fire

~$200B

India's IT-BPM exports — much of it work AI can automate within 5–10 years.

~8–9%

BPO's share of Philippine GDP — one of the most BPO-dependent economies in the world.

~$10B

Vietnam's software exports — mostly outsourcing for the US and Japan.

But numbers tell only half the story. To make compute inequality concrete, look at three slices of Vietnamese life I see every week. Names changed; details composite.

Phương

32 years old

Founder · Hà Nội

Compute / month

≈ $1,000

Top of the curve

$0 $200 $1K+

Running two AI-native startups. Pays $200/month for Claude Max, $200/month for ChatGPT Pro, several hundred USD/month for APIs and specialized tools. Total personal compute budget around $1,000/month — equal to one full-time junior salary in Vietnam.

A team of 5, with output equivalent to a team of 50 three years ago. At the most recent funding round, US investors were impressed by the leverage. Phương said: “I’m not better than I was two years ago. I just have more compute.”

Phương stands at the top of the curve. Compute inequality makes Phương 10–20× more productive than the Vietnamese baseline for the same role.

Hoàng

28 years old

Outsourcing developer · Đà Nẵng

Compute / month

≈ $40

Middle of the curve

$0 $200 $1K+

Five years of working for US clients. $2,000/month — high by Đà Nẵng standards, enough to rent a decent apartment, send money home. Already paying off a car loan, planning to buy a house.

The current contract has 12 months left. Last month, the client wrote: “From the next contract, we expect AI-augmented productivity. We’re consolidating from 8 devs to 4.” Hoàng was kept — but the salary stays the same while expectations double.

Hoàng now uses Cursor and Claude Code, learning day and night. English is decent but not native — has to draft a prompt 2–3 times to get the result Phương gets in one. Has an international Visa, has stable internet — but at $2,000/month, paying $200/month for Claude Max is a wrenching decision.

Hoàng sits squarely in the middle. Compute inequality may keep Hoàng employed — or unemploy Hoàng within 18 months. The curve is splitting fast under his feet.

Linh

20 years old

Second-year university student · Cần Thơ

Compute / month

$0

Bottom of the curve

$0 $200 $1K+

Foreign-language major. English is moderate — enough to read coursework, not enough for natural conversation. Old Android phone. Slow Wi-Fi at the rented room. Parents are farmers.

Uses free ChatGPT to translate texts. Doesn’t know what Claude is — a roommate recommended Gemini, said “free, fewer caps than ChatGPT.” No international card, doesn’t think about paying $20/month — that is three days of meals.

When I asked Linh what she thought about AI, she said: “I’ve heard AI replaces people. I don’t really understand. I just worry whether I’ll have a job after graduation.”

Linh is not on the curve — she is standing at the bottom, unaware the curve exists.

---

Phương and Linh live in the same country, the same year. The Labor Code applies to both. The same education system trained them. Same election, same taxes, same policies.

But compute inequality makes the gap between them larger than the gap between Phương’s parents and Linh’s parents — and it is widening every month.

6. The real question — and a moment of self-doubt

While writing this essay I felt a strong temptation: end with five things “you should do.” Promise that if we all try, it will be fine.

I will not. For two reasons.

One, most “five things you should do” are reflexive personalizations of a structural problem. Learn AI every day, use the right tools — all good for you. But if you are one of the 8% who already have the conditions to do those things, the advice is redundant; if you are not in the 8%, the advice is meaningless. Structural inequality is not cured by individual advice.

Two, the real question is not “is AI fair” — clearly no. The real question, for those of us on the favored side of the wave, is:

Where I might be wrong

Before closing the essay, a moment of self-doubt. To be fair to readers who read this far — and to the argument itself.

If Brynjolfsson is right and GenAI continues to narrow the gap between novices and experts in the long run — not just in customer service but across the entire economy — this essay overstates pessimism. If Acemoglu is right and AI’s productivity boost is only ~0.66% per decade rather than transformative — my compute concentration argument is much weaker. If Autor is right and AI genuinely rebuilds the middle class — the entire “hollowing out the middle” section points the wrong way.

I do not claim 100% certainty. I bet on the pessimist scenario for three reasons:

1. Compute concentration is structural, not transitional. Five organizations controlling all frontier compute = a structural root, hard to reverse without strong anti-trust action or a technical breakthrough that drops compute cost 100×.
2. History shows productivity gains rarely distribute evenly without strong policy or organized counterweights (unions, regulation, anti-trust). All three are at their weakest in 50 years.
3. Vietnam lacks the political-economic strength to force redistribution the way Germany or the Nordics can. This argument applies especially to Vietnam — and to most developing countries.

If any one of these is wrong, the argument weakens significantly. I admit that — and still lean my way.

Four principles I am using on myself

Not advice for you. Just principles I am testing — and will report back on in 12 months.

One, pay compute forward to those who cannot yet pay. When I build products with 10× AI leverage, I commit at least 10% of the output to those who cannot afford their own $200/month — open source, free tier, or subsidized. Rented productivity carries a moral obligation.

Two, teach taste, not tools. Tools change every month. Taste does not. Vietnam lacks taste more than it lacks tools — and taste is transmitted through mentorship, not YouTube tutorials.

Three, invest in the next generation’s English. Axis 4 (English) is the axis Vietnam can close fastest with the right investment — and it carries the largest amplifier coefficient. One dollar invested in a child’s English today = one hundred dollars of advantage in 15 years.

Four, do not flinch from political questions. AI inequality is ultimately a political problem, not a technical one. Vietnamese founders who talk compute inequality but avoid policy = complicit with the status quo. I am still learning how to speak about public policy without sounding sanctimonious — hard, but necessary.

This is the most important conversation of our generation. Do not let it happen only in conference rooms in Silicon Valley. Do not let it happen only among the people already on the winning side.

---

Appendix: Primary references

Empirical research on AI and productivity

• Brynjolfsson, Erik; Li, Danielle; Raymond, Lindsey R. (2023). “Generative AI at Work”. NBER Working Paper 31161. — Field study of 5,179 customer support workers; juniors +34%, seniors near zero.
• Noy, Shakked; Zhang, Whitney (2023). “Experimental evidence on the productivity effects of generative artificial intelligence”. Science Vol 381, Issue 6654. — N=453 experiment; productivity gap narrowed.
• Eloundou, Tyna; Manning, Sam; Mishkin, Pamela; Rock, Daniel (2023). “GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models”. arXiv:2303.10130. — Estimated exposure for 80% of US workers.

Macroeconomic analysis

• Acemoglu, Daron (2024). “The Simple Macroeconomics of AI”. NBER Working Paper 32487. — Pushback against AI productivity hype; estimates ~0.66% TFP boost per decade.
• Autor, David (2024). “Applying AI to Rebuild Middle-Class Jobs”. NBER Working Paper 32140. — Argues AI may rebuild the middle class.
• Korinek, Anton; Stiglitz, Joseph E. (2020). “Artificial Intelligence, Globalization, and Strategies for Economic Development”. NBER Working Paper 28453. — AI-and-inequality scenarios.
• Cazzaniga, Mauro et al. (2024). “Gen-AI: Artificial Intelligence and the Future of Work”. IMF Staff Discussion Note 2024/001. — 40% of jobs globally, 60% in advanced economies, exposed.
• Piketty, Thomas (2014). Capital in the Twenty-First Century. Harvard University Press. — Long-run analysis of productivity distribution and capital vs. labor returns.

Compute concentration

• Sevilla, Jaime; Heim, Lennart; Ho, Anson; Besiroglu, Tamay; Hobbhahn, Marius; Villalobos, Pablo (2022). “Compute Trends Across Three Eras of Machine Learning”. arXiv:2202.05924. — Training compute growing 4–5× per year since 2010.
• Vipra, Jai; Korinek, Anton (2023). “Market Concentration Implications of Foundation Models”. Brookings. — Concentration in the AI value chain.

Books and essays on meaning and society

• Harari, Yuval Noah (2016). Homo Deus: A Brief History of Tomorrow. Harvill Secker. — The useless class concept.
• Harari, Yuval Noah (2024). Nexus: A Brief History of Information Networks. Random House. — Updated framework eight years on.
• Susskind, Daniel (2020). A World Without Work. Metropolitan Books. — Three losses: income, skill, meaning.
• Frey, Carl Benedikt (2019). The Technology Trap: Capital, Labor, and Power in the Age of Automation. Princeton University Press. — Historical warning about transitional instability.
• Varoufakis, Yanis (2023). Technofeudalism: What Killed Capitalism. The Bodley Head. — The “cloud capital” argument.
• Hao, Karen (2025). Empire of AI. Penguin. — Reporting on OpenAI and global compute concentration.

Industry reports

• Stanford HAI (2024–2025). Artificial Intelligence Index Report. — Data on AI investment, hiring, labor impact.
• McKinsey (2024). The Economic Potential of Generative AI. — 70% of GenAI value sits in use cases requiring private enterprise data.
• EF Education First (2024). EF English Proficiency Index. — Vietnam in the “moderate” band, near Mexico, below Indonesia/Philippines.
• Stack Overflow (2024). Developer Survey. — 76% of developers use AI daily; ~20% report high trust in output.
• NASSCOM (2024). Indian IT-BPM Industry Strategic Review.
• IBPAP (2024). Philippine IT-BPM Industry Roadmap.
• VINASA (2024). Vietnam Software Industry Report.

Valuations and market caps

• Nvidia, OpenAI, and Anthropic figures drawn from public filings and announced funding rounds reported in Bloomberg, Financial Times, and The Information, 2023–2025. Numbers are approximate as of writing (April 2026) and may have shifted.