The Economic Risks of AI: Productivity, Inequality, Power

The five companies that can afford to train a frontier model

Training GPT-4 reportedly cost OpenAI somewhere north of $100M in compute alone. Training the next generation will cost ten times that, and the list of organisations that can write the cheque is shorter than the list of countries with a space programme.

This guide is about what that fact, and four others like it, do to an economy. Not the science-fiction scenarios — the boring, measurable ones already showing up in payroll data, antitrust filings, and the training sets of the next model. The goal is to leave you able to reason about AI's economic impact with named mechanisms, not vibes.

The productivity gains are real, and smaller than the slides suggest

For the first time in the long history of "this technology will change everything," there are actual numbers. A 2023 study by Brynjolfsson, Li and Raymond at a Fortune 500 customer-support call centre found that access to a GPT-based assistant raised resolved-issues-per-hour by about 14% on average — and 35% for the least experienced agents. A separate randomised trial by GitHub on Copilot — Microsoft's AI coding assistant — clocked a 55% speed-up on a narrow task (writing an HTTP server in JavaScript). A BCG-Harvard study on consultants using GPT-4 found a 25% speed gain and a 40% quality lift on tasks inside the model's "frontier."

The honest summary is roughly 20–30% in well-bounded knowledge work — coding, customer service, drafting, summarisation — and approximately zero in tasks the model is bad at. Outside the frontier, the BCG study found consultants using GPT-4 actually performed *worse* than the control group, because they trusted confident-sounding wrong answers. The productivity gain is real. It is also conditional, uneven, and often invisible at the firm level because it is eaten by coordination overhead.

The interesting thing is *who* it helps. In every serious study, the largest gains accrue to the bottom of the skill distribution: the junior agent, the average coder, the new consultant. AI compresses the gap between novice and expert on routine cognitive work. Which sounds egalitarian — until you ask what happens to the wage structure built on that gap.

Wage compression at the bottom, escape velocity at the top

The standard story about technology and inequality, going back to Goldin and Katz, is "skill-biased technical change" — machines complement educated workers and substitute for everyone else, so the college premium grows. AI looks different. It substitutes most directly for the easier parts of educated work: writing a passable memo, generating boilerplate code, summarising a deposition. The *median* knowledge worker becomes more productive, but the price of that median work falls as the supply effectively expands.

Meanwhile, the top decile pulls away. A senior engineer who can architect a system Copilot cannot, a partner who can ask the right strategic question of GPT-4, a researcher who can verify what the model gets wrong — these people now operate with a small army of competent juniors who never sleep. Their leverage rises. The middle gets squeezed from both sides: cheaper than it was, less differentiated than the top, and competing with a tool that costs $20 a month.

This is wage compression at the bottom and escape velocity at the top, in the same labour market, at the same time. Anton Korinek at Brookings has been the clearest voice on this — his 2024 work on AI and inequality argues that without policy intervention the labour share of income, already in a 40-year decline, will fall further. The gains accrue to capital (the people who own the models and the firms that deploy them) and to the small set of workers whose judgement remains uncopiable. Everyone else competes harder for a smaller piece.

What most people miss is that this can happen even if AI creates more jobs in aggregate. Aggregate employment is not the same as the distribution of bargaining power, and bargaining power is what determines wages.

Capital concentration: five labs, one supply chain

The cost of frontier training is the most under-appreciated economic fact of the decade. GPT-4 cost an estimated $100M+. Anthropic's Claude 3 Opus and Google DeepMind's Gemini Ultra are in the same range or higher. Dario Amodei, Anthropic's CEO, has said publicly he expects the next generation to cost $1B and the one after that $10B. Even if he is off by half, the implication holds: pre-training a frontier model is a capital expenditure on the scale of a semiconductor fab or a nuclear reactor.

The list of entities that can plausibly fund this is short. OpenAI (backed by Microsoft), Anthropic (backed by Amazon and Google), Google DeepMind, Meta FAIR, xAI, and a handful of Chinese labs (Alibaba's Qwen, DeepSeek, Moonshot, Baidu's Ernie). Mistral and a few others are competitive at smaller scales. That is the entire list. Every other "AI company" — and there are tens of thousands — is building on top of one of these foundations, the way every web company builds on TCP/IP, except TCP/IP is owned by five firms.

Below that, the supply chain narrows further. NVIDIA reported roughly $30B in data-centre revenue *per quarter* in 2024 and effectively owns the GPU market for AI training. Its high-end chips depend on TSMC's most advanced nodes in Taiwan, which depend on ASML's extreme-ultraviolet lithography machines, which are made in the Netherlands and which the US has restricted from export to China. The economic geography of AI is: Dutch lithography → Taiwanese fabrication → American hyperscalers and labs → global users. There is no second source for any link in that chain.

The risk is not that this stack fails. It is that whoever controls a link in it controls the price of intelligence. Who builds frontier AI is no longer a research question; it is an antitrust and industrial-policy question, and the regulators are several years behind.

The US-China duopoly, and everyone else as customers

Geography compounds capital concentration. The US has the chips, the capital markets, and four of the top five frontier labs. China has the state subsidies, the talent pool, the second-largest set of frontier-class models, and — for now — restricted access to the best NVIDIA hardware, which is forcing genuine architectural innovation (DeepSeek-V3 trained on roughly $6M of compute is the proof of concept that constraints breed efficiency).

The European Union has produced exactly one frontier-adjacent lab, Mistral, and a regulatory framework, the AI Act. The framework arrived first. The UK has DeepMind, but DeepMind is owned by Google. India, Japan, South Korea, the UAE, Israel — all serious technology economies — have released national models that are competitive at the open-source frontier but not at the closed one. None of them have a domestic substitute for GPT-4-class capability without licensing it from a US or Chinese provider.

This is the first time in the modern industrial era that a general-purpose technology has a duopolistic geographic origin. Electricity, the internal combustion engine, the transistor, the internet — each diffused widely from invention to ubiquity. AI may diffuse in *use* but the production of frontier models is consolidating, not diffusing. The strategic position of every country that is not the US or China is roughly the position of every country that is not Saudi Arabia or Russia in oil: you can be a sophisticated consumer, but you import the upstream.

The interesting question is whether open-weights models — Llama 3, DeepSeek, Qwen, Mistral — change this. They lower the floor: a competent national model is now a matter of fine-tuning, not pre-training. They do not raise the ceiling: the frontier still belongs to the labs that can spend $1B on a training run.

Model collapse: the risk almost no one is pricing in

Here is the economic risk the slide decks ignore. Frontier models are trained on the public internet — roughly the entire crawlable web, plus books, plus code. As of 2024, by some estimates, more than half of new web content is partly or wholly AI-generated. By 2027 it will likely be the majority. The training set for the next generation of models will be, in significant part, the output of the current generation.

A 2023 paper by Shumailov et al., "The Curse of Recursion," demonstrated experimentally what happens when you train a model on its predecessor's output and repeat: the tails of the distribution collapse. Rare events, unusual phrasings, minority dialects, niche knowledge — all the texture that makes a language model useful — get smoothed away. The model becomes more confident and less correct. They called it *model collapse*.

The economic consequence is that the open public commons — the human-written internet — becomes a finite, depleting resource, like a fishery. Each generation of AI extracts value from it and partially poisons it. The labs know this. It is why they are paying Reddit, the Associated Press, Reuters, Stack Overflow, News Corp and Le Monde for licensed human-generated training data. Data is the new oil only if you grant the analogy: extractive, geographically concentrated, and prone to depletion.

What most people miss is that model collapse is not a future risk. The watermarking and provenance infrastructure that would let you reliably distinguish human from AI text at scale does not exist. The contamination is already in the training pipelines of the next generation. The labs with proprietary access to fresh human data — Reddit's deal with Google, Bloomberg's terminal corpus, X's posts, every newspaper that has signed — have a structural moat that did not exist three years ago. Compute used to be the bottleneck. Increasingly, it is uncontaminated text.

What this means for you

If you employ knowledge workers, the productivity gain is real but conditional. Deploy AI on bounded, verifiable tasks where a confident wrong answer is cheap to catch. Avoid it on the frontier of judgement, where it will sound right and waste a quarter. Measure outcomes, not adoption. The firms that win this decade will not be the ones that bought the most seats; they will be the ones that redesigned a workflow around the tool, the way the firms that won the 1990s were not the ones that bought the most PCs but the ones that rebuilt around email and ERP.

If you sell your labour, ask honestly which 30% of your work a $20-a-month tool now does as well as you. That share will be priced down. The remaining 70% — taste, judgement, accountability, the parts where being wrong has consequences — is where your wage will live. Specialise upward. The middle of every knowledge profession is the dangerous place to stand.

If you are a citizen, the policy questions to watch are antitrust on the cloud-and-model layer (Microsoft–OpenAI, Amazon–Anthropic, Google–Anthropic), export controls on the chip layer, and copyright on the training-data layer. The first two will determine market structure for a generation. The third will determine whether the public commons survives the model that ate it. None of these are being seriously addressed at the speed the technology is moving.

Ask Gradient

For deeper, citation-backed answers on the questions this guide raised, try our Ask Gradient RAG — it answers from a curated corpus of AI research, lab publications, and economic analysis. A good question to start with: *"What does the latest empirical evidence say about AI's effect on entry-level wages in software engineering since 2023?"* Pair it with our companion guide, What is an LLM, if the underlying technology is still fuzzy.