AI and the Concentration of Power
Who owns the chips, the models, the compute, and the data. The new industrial geography that will shape every technology and national-security policy decision of the next decade.
(OpenAI, Anthropic, Google DeepMind, Meta, and a handful of others)
(a two-year figure, and of the same order as two years of US public infrastructure investment)
(Nvidia; TSMC manufactures them)
A note on framing. Artificial intelligence is one of the few technical revolutions whose course is genuinely uncertain. Reasonable people who have watched the field closely for decades do not agree about how transformative the current AI systems will be over the next ten years. The page below tries to walk through the parts that are well-measured (industrial structure, capital concentration, who owns what), the parts that are genuinely open (capability over time, broader economic impact), and the parts that depend heavily on policy choices that have not yet been made.
The new industrial geography
The current generation of AI systems runs on a remarkably concentrated supply chain. The chips that train and run the models are designed primarily by Nvidia, manufactured primarily by TSMC in Taiwan, and use lithography machines made by ASML in the Netherlands. The models themselves are trained primarily by a handful of US-based companies (OpenAI, Anthropic, Google DeepMind, Meta, xAI) and a smaller number of Chinese ones (DeepSeek, Alibaba, Baidu, ByteDance, Tencent). The data centres that host the training runs are increasingly large, expensive, and energy-hungry, and they are increasingly clustered in a small number of geographic locations chosen for cheap electricity and reliable grid connections.
Each layer of this stack is a chokepoint. Each chokepoint is, in turn, a piece of strategic geography. The Taiwan chip story is covered in a separate piece on this site; here the focus is on the layers above it. The simplification that "AI is being built by tech companies" misses how much of it is now industrial - vast investments in physical infrastructure, contested by governments, with national-security implications that are reshaping export-control law, immigration policy, and energy policy on multiple continents.
The capital intensity is easy to underestimate. Training a frontier model in 2025 costs hundreds of millions of dollars. Running large numbers of users on it costs comparable amounts in operating expenses. Building the data centre to support it costs several billion dollars. Major US technology companies are now spending more on AI-related capital expenditure than the federal government spends on basic scientific research, by several multiples. The economics of a frontier AI lab look more like the economics of a semiconductor company or a refinery than like the economics of an early internet startup.
The handful of companies that matter most
Most of the world's frontier AI capability sits inside a small number of organisations. The names are unusually well-known and the structural picture is more concentrated than at any point in the modern computing industry. The numbers below are approximate and shift quarter to quarter, but the rough picture is steady.
The takeaway: AI is more concentrated than the early internet, more concentrated than the semiconductor industry, and more concentrated than any major technology since perhaps the early 20th century industrial trusts. Five labs and one chip designer hold most of the capability that matters. Whether this concentration will compound over time (the network-effects-and-capital case) or break down (the open-weights-and-efficiency case) is one of the central technological and political questions of the next decade.
What AI is doing today
Past the hype, AI is already doing real work. The work is uneven and the productivity gains are uneven, but the deployment is real and growing. A short tour of where the technology is actually being applied at scale:
Software development. AI coding assistants (GitHub Copilot, Cursor, Claude Code) are now used by a majority of professional software developers in major US technology companies. The productivity effect is real and measurable, somewhere in the range of 20-50% time saved on common tasks. Software engineering is one of the few cases where the technology is already producing direct business value at scale.
Customer service and operations. Large enterprises are deploying AI to handle initial customer inquiries, internal employee questions, document processing, and routine data work. Cost savings are real but uneven. Most projects deliver less than the headline pitch promised; some deliver substantially more.
Knowledge work assistance. Lawyers, doctors, accountants, journalists, researchers, and consultants increasingly use AI as a drafting and research tool. The pattern that has emerged is "AI as a junior associate" - useful for first drafts, summarisation, and structured tasks; still requiring expert review for anything consequential. Fully automated knowledge work remains rare and most attempts at it have stumbled.
Search and information retrieval. The Google search experience for hundreds of millions of users has been substantially reshaped by AI summaries. The economic implications for the open web are large and not fully understood; many publishers report falling traffic from search since AI summaries became the default presentation.
Healthcare. Radiology, pathology, and diagnostic support are seeing real progress, especially in rural and developing-country settings where specialist scarcity is acute. Drug discovery and protein structure prediction (DeepMind's AlphaFold work) have already changed how some research is done. Direct patient-facing AI medical care is more limited and more contested.
What it has not yet done. Despite many predictions, AI has not produced large measurable productivity gains at the macroeconomic level. US productivity growth has accelerated modestly since 2022 but the connection to AI is unclear and contested. The deeper transformation - the one that would justify the trillion-dollar capital investment - has not yet shown up in the aggregate statistics. Whether it will, and on what timescale, is one of the largest open questions in economics today.
What it has done to work so far
The aggregate labour-market picture is calmer than almost any public discussion of it. Studies using administrative payroll records and large surveys have found little evidence of economy-wide job loss or wage decline through the period of fastest adoption. Unemployment in the countries with the heaviest deployment has stayed low. If the question is whether AI has caused mass displacement, the answer from the data so far is no.
The distributional picture is where something is visibly happening, and it is concentrated in one place. Erik Brynjolfsson's group at Stanford has tracked it using payroll records from the largest US payroll processor, covering millions of workers. Employment among 22-to-25-year-olds in the most AI-exposed occupations is down by around 13% relative to workers in less exposed roles since generative AI came into wide use, and closer to 16% once firm-level shocks are controlled for. Employment for experienced workers in the same occupations held steady. The mechanism matters as much as the number: the decline comes from a collapse in hiring rather than from layoffs, and it is concentrated in roles where AI automates the task rather than assisting with it.
The mechanism proposed for this fits the pattern better than a general-displacement story does. Current systems are good at codifiable tasks that do not require accumulated judgment - retrieval, summarising, drafting, formatting, scheduling, first-pass research - and those tasks are disproportionately what junior staff are hired to do. For experienced workers the same capability is complementary: it removes the parts of the job that were never the valuable part. The result is not fewer jobs in aggregate but a reshaped career ladder, with the bottom rungs thinner than they were.
That has a consequence worth stating even though nobody has measured it yet. Experienced practitioners are made, not recruited, and the making has historically happened by doing junior work badly and then better. A profession that automates its apprenticeships is borrowing capability from its own future, and the bill arrives a decade later when the senior cohort retires without a replacement pipeline behind it. Several professions have started to say this out loud - law, accountancy, software - without yet changing how they train.
Two honest caveats. The series is short, and it overlaps with a post-pandemic normalisation in hiring and a period of high interest rates, both of which independently reduce entry-level recruitment; disentangling them is not yet possible with confidence. And "exposure" in these studies measures how much of an occupation's task content current models can perform, which is a reasonable proxy and not the same thing as an employer deciding to substitute. The direction of the finding is consistent across several independent datasets. The magnitude is not settled.
Three regulatory philosophies, and no common rulebook
Governments have stopped debating whether to regulate AI and started producing rules. What they have not produced is anything resembling a shared approach, and the three largest jurisdictions have chosen structurally different answers. This matters for the industrial concentration described above, because rules determine which companies can operate where and at what compliance cost - which is usually a barrier that favours the largest incumbents.
The European Union legislated first and most comprehensively. The AI Act, adopted in 2024, sorts systems into risk tiers and applies obligations before deployment rather than after harm. It is arriving in phases. Bans on the prohibited practices and the AI-literacy duties took effect in February 2025, obligations on general-purpose models in August 2025, and most transparency requirements in August 2026. The heaviest obligations are those on high-risk uses - biometrics, critical infrastructure, education, employment, migration and border control - and their timing has moved: an amendment agreed politically in 2026 deferred them, with the published schedule still catching up to the change. That the deadlines slipped at all is itself informative about how hard the regime is proving to implement, and anyone relying on a specific date should check the current text rather than a secondary summary, this one included.
The United States has no federal AI statute and shows no sign of passing one. What exists instead is a patchwork: well over a hundred state laws enacted, agency guidance, and voluntary commitments. Colorado passed the first comprehensive state framework, requiring deployers of high-risk systems to take reasonable care against algorithmic discrimination. California's frontier-model law requires developers of the largest models to publish risk frameworks and report safety incidents, with penalties up to a million dollars per violation. For a company operating nationally this produces the compliance burden of a federal regime without the predictability of one, which is the outcome that both industry and its critics say they did not want.
China regulates by application rather than by statute, through dozens of targeted rules issued mostly by the Cyberspace Administration. Recommendation algorithms, synthetic media, generative services, AI agents and AI companions each have their own instrument, with content-control obligations running through all of them. The approach is faster to update than a comprehensive law and produces rules that are specific enough to enforce, at the cost of coherence and with purposes that include political control as well as safety. It is worth noting that this makes China both heavily regulated and regulated for different ends than the other two.
Internationally there is less than the volume of summitry suggests. The Council of Europe's framework convention, in force since November 2025, is the first legally binding international treaty on AI, though its obligations are general and its membership does not include the countries where most frontier development happens. A United Nations dialogue on AI governance held its first session in Geneva in July 2026. Neither constrains what the major labs do. The practical situation is two regulatory blocs with incompatible assumptions, a large third group choosing between them, and no forum with authority over the handful of companies described at the top of this page.
The paths from here
AI is unusual among the topics on this site because the range of plausible outcomes over the next decade is genuinely wide. Each path below is a coherent reading held by some subset of careful researchers and analysts; they are not mutually exclusive in the short run.
Capability scales steadily; deployment slowly transforms knowledge work
The current generation of models keeps improving year over year, with the rate of improvement gradually slowing as the easiest gains are exhausted. Knowledge work is reshaped over five to ten years rather than overnight. Productivity growth picks up modestly, mostly in software, finance, professional services, and customer-facing roles. The structure of the economy in 2035 looks recognisably similar to today, with substantial efficiency improvements in many sectors.
Will it happen? This is the base case among most economists who have studied the technology carefully. It assumes the current scaling laws keep producing useful gains but with diminishing returns; the deployment friction is real but workable; and the societal adjustment happens at a pace history has seen before with previous technology revolutions.
A capability jump produces substantially more capable systems
Sometime in the next five to ten years, AI systems become substantially more capable than today - able to do most knowledge work to an expert standard, capable of multi-step planning that current systems struggle with, and able to do real scientific research. The economic and political consequences are large and arrive faster than most current institutions can adapt.
Will it happen? Genuinely uncertain. The major lab leaders mostly believe something like this is coming, with timelines ranging from "within five years" to "within fifteen years." Many serious AI researchers think the current scaling approach has fundamental limits and that this scenario requires architectural changes that have not yet been figured out. The probability is high enough to take seriously and low enough that betting everything on it would be foolish.
The closed frontier versus the open-weights challenge
The cost of training frontier models keeps rising, which could leave only a handful of US and Chinese companies able to keep up - an oligopoly structurally similar to cloud computing, with national-security separation between Western and Chinese ecosystems. The countervailing force is the open-weights ecosystem (Meta's Llama, DeepSeek, Mistral), where algorithmic efficiency has been improving roughly twice as fast as raw compute scaling. If that holds, capable open models become a commodity - as Linux did for operating systems - and the closed labs keep some advantages but lose pricing power.
Will it happen? Both forces are already visible. The capital costs of the closed frontier are rising sharply; the DeepSeek release of late 2024 was a significant data point on the open side. The race between closed-frontier compute and open-weights efficiency is the single most contested question in the field, and which side wins shapes who ends up controlling the technology.
Power concentration triggers serious regulation
The combination of corporate concentration, national-security implications, labour-market disruption, and visible misuse triggers a serious regulatory response in the US and Europe. Compute thresholds, training-data audits, model-evaluation requirements, and deployment licences become standard. The pace of capability deployment slows; safety and accountability practices improve.
Will it happen? Already starting in Europe (the EU AI Act) and beginning in the US (executive orders, state laws, FTC enforcement). The pace and depth of regulation in the US specifically is a live political question. Most serious analysts think some version of this is likely; the specifics will shape outcomes for the next decade.
Geopolitical decoupling separates Western and Chinese AI
Export controls, security reviews, capital restrictions, and divergent regulatory environments produce two largely separate AI ecosystems. Western and Chinese labs increasingly cannot use each other's chips, training data, or talent. Each ecosystem has its own foundation models, deployment platforms, and corporate champions. By 2030 they look like two largely separate technology stacks.
Will it happen? Already underway. US chip export controls, Chinese restrictions on Western AI services, divergent regulatory frameworks, and visa restrictions on AI researchers are all moving in this direction. Full decoupling is not yet here but the trajectory is clear. The cost of this for both sides will be larger than is currently appreciated.
Disappointment cycle hits hard
AI capability progress slows or plateaus. The economic returns on the trillion-dollar investments fail to materialise. Major investors take losses. The hype cycle deflates and the field consolidates. Useful applications continue but the transformative promises fade for at least several years until the next wave of progress.
Will it happen? Possible, especially if scaling-law gains slow faster than expected and major productivity benefits fail to materialise. AI has been through several disappointment cycles before; the pattern is not unfamiliar. Most current researchers think this is less likely than continued steady progress, but there is room for serious people to disagree.
Steady capability scaling is the base case. Some version of regulatory tightening is likely. Geopolitical decoupling is well underway. Whether a major capability jump arrives, and when, is genuinely uncertain. The single most important variable for individual readers is probably whether AI lives up to the productivity promises in their specific industry over the next five years - which depends as much on specific deployment work as on raw model capability.
Where serious analysts disagree
AI is one of the topics where the disagreements among serious researchers are wider than the public conversation suggests. Each reading below is held by named figures whose work is worth engaging directly.
Transformative AI is close and the world is unprepared
Continued scaling is likely to produce systems that can do most knowledge work to expert level within five to ten years. The economic, political, and security consequences will be larger than any technology shift since the industrial revolution. Current institutions are nowhere near ready, and the time to prepare is now.
Held by: the leadership of most major AI labs (Sam Altman at OpenAI, Dario Amodei at Anthropic, Demis Hassabis at Google DeepMind), and a substantial fraction of AI researchers within those labs. Their position is partly substantive and partly self-interested - their companies' valuations depend on it - but the underlying technical case has serious merit and is taken seriously by careful outsiders.
Current systems hit a fundamental wall
The transformer architecture and the current scaling approach have intrinsic limits that are not solved by adding more compute. The systems are good at pattern-matching from training data and worse at the structured reasoning, planning, and learning from limited examples that genuine intelligence requires. Without architectural breakthroughs, the rate of capability gain slows substantially over the next few years.
Held by: Yann LeCun (Meta's chief AI scientist until the end of 2025, now running his own research company), Gary Marcus (NYU), Subbarao Kambhampati (Arizona State), and a number of senior AI researchers with long track records. Their critique of current systems is technical and specific; whether they are right depends on questions that are actively being tested and may take a few more years to settle.
The economic productivity case is overstated
Despite massive investment and visible deployment, AI is not yet showing up in the economic productivity statistics in the way that the headline projections suggest it should. Productivity gains in narrow domains (coding, customer service) are real but the diffusion to broader economic impact is slower than the capital expenditures imply. The gap between promised returns and actual returns may close, but it may also widen.
Held by: Daron Acemoglu (MIT), Robert Gordon (Northwestern), and a strand of economic-historical thinking that has been sceptical of new-technology productivity claims for decades. Their data has been right more often than wrong about productivity timing in past technology cycles.
Open-source AI will keep the field competitive
The combination of open-weight models from Meta, the algorithmic efficiency gains demonstrated by DeepSeek and others, and the long tail of academic and corporate research will keep frontier AI capabilities accessible to a broader community than the major closed labs. The "AI oligopoly" framing assumes capability is bottlenecked by capital; the actual picture has more competition than the corporate-concentration narrative suggests.
Held by: Yann LeCun, parts of the Meta AI organisation, and the broader open-source AI community. The DeepSeek release of late 2024 was a significant data point in their favour. The case is partly substantive and partly normative - they want this future, which makes it worth checking against actual evidence rather than just shared aspiration.
The catastrophic-risk case deserves to be taken seriously
The strong version of the safety argument is that sufficiently capable AI systems could pose risks at the level of major civilizational disruption: large-scale misuse for cyber and biological attack, persuasion at unprecedented scale, autonomous systems pursuing goals that drift from human intent, or in the most concerning scenarios, systems that humans can no longer reliably oversee. The labs' published safety work is real but is widely viewed as inadequate to the worst-case scenarios it is meant to address, and the commercial pressure to deploy faster than safety would suggest is structural.
Held by: Geoffrey Hinton (formerly Google, who left specifically to speak about these risks), Yoshua Bengio (Mila), Stuart Russell (UC Berkeley), parts of the Anthropic safety team's published work, and a community of AI safety researchers. Hinton and Bengio have publicly estimated non-trivial probability of catastrophic outcomes within decades. Their case is partly technical and partly precautionary; reasonable people disagree about how seriously to weight low-probability but high-consequence scenarios, but dismissing the position as alarmism misreads the technical argument and the credentials of the people making it.
Most of this is hype on top of capable but limited tools
The dismissive reading argues that current AI systems are sophisticated pattern matchers being marketed with language that overstates what they actually do. The hundreds of billions in capital expenditure, on this account, will produce a useful set of tools but nothing remotely matching the sales pitch. The "agentic AI," "AGI within five years," and "transformative impact" framings are partly genuine technical optimism and partly the exit strategy for an investment bubble that needs continued belief in transformative outcomes to justify current valuations. Past technology cycles produced both real productivity gains and significant capital destruction; this one will likely follow the same pattern, with the painful corrections still ahead.
Held by: Emily Bender (University of Washington), parts of the linguistics and AI-ethics community who have argued the systems are "stochastic parrots," Gary Marcus when he focuses on the bubble dynamics, and a growing number of financial analysts pointing to revenue versus capital-expenditure mismatches at major AI companies. The case is partly substantive and partly skeptical of the financial story; the empirical question of how much economic value the current systems actually generate at scale is being settled in real time, and the early returns are mixed.
The spread between these readings is unusually wide for a topic that gets so much coverage. Reasonable people who have followed the field closely sit in very different places on the range from "potentially civilizationally catastrophic" to "largely a financial bubble," and the range between those positions is not a sign that some participants are uninformed. It is a sign that the underlying questions - what current systems can actually do, how fast they will improve, what economic and security effects they will produce, and how seriously to weight low-probability worst-case scenarios - are genuinely difficult and turn on technical judgments that even experts disagree about. What can be said from the available evidence: AI is a real technology revolution producing real but uneven economic effects so far; the capability trajectory over the next ten years is genuinely uncertain; the corporate concentration is sharp but is being challenged from the open-weights side; the geopolitical implications are large and arriving faster than democratic institutions can adapt; and researchers on the safety side of the field mostly report that safety work is not keeping pace with capability work, which is a claim about their own field that the capability-focused labs dispute.
What this means for you
AI is one of the topics where the personal practical implications are unusually concrete and time-sensitive. Some observations:
If your work involves writing, coding, or analysis
Get fluent with the major AI tools. The productivity gain available is real and currently substantial - 20% to 50% on many specific tasks - and the gap between people who use AI tools well and people who do not is widening. The skill is not "knowing AI." It is knowing which tasks AI is good at, which it is bad at, and how to verify its output. The half-life of specific tools is short; the underlying skill of working well with AI assistants is durable.
If you work in a job AI might replace
The honest assessment is that some routine knowledge-work tasks will be increasingly automated over the next five to ten years. The pace and breadth depend on choices that have not yet been made by the labs and by employers. The most resilient roles combine human judgment, relationship work, physical presence, or specific accountability that AI cannot bear. If you are early or mid-career, weighting your skill development toward those qualities (rather than toward the tasks AI does best) is the most useful response. If your role is heavily exposed, taking action now while alternatives are still many is much better than waiting until they are few.
If you invest
The AI investment landscape is unusually concentrated. The major beneficiaries (Nvidia, the cloud providers, the foundational labs and their parents) have already had enormous price moves. Whether the returns from here justify the current valuations depends on the productivity case actually delivering at scale, which is exactly the contested question the previous section discussed. Diversified exposure through broad technology funds is reasonable. Concentrated single-name bets in the space are riskier than they look. The historical pattern of major technology revolutions is that the first wave of investors loses money and the second wave makes it; whether AI follows that pattern is unclear.
If you worry about AI safety
The safety conversation is real but is also frequently distorted by the loudest voices. Worth engaging with the careful version: AI labs themselves publish detailed work on safety and alignment, several universities have serious research programs on it, and a small number of policy organisations work on practical questions of evaluation, deployment, and accountability. Lay-level worry tends to oscillate between dismissing the issue entirely and assuming science-fiction scenarios are imminent; the careful middle is the most useful place to read from.
If you are thinking about your children's education
The skills AI is currently bad at are exactly the ones humans need to thrive in a world where AI is good at the others: judgment under uncertainty, learning from limited data, physical and social skill, ethical reasoning, original creative synthesis, and the ability to navigate ambiguity. Education focused on these durable skills - rather than on the kinds of structured knowledge work AI is increasingly competent at - is a reasonable response to the technology's actual trajectory. Schools, like other institutions, are adapting slowly; supplementing what schools provide is increasingly important rather than optional.
The mechanics behind this
The AI story sits on top of three deeper mechanisms covered elsewhere on this site. If the analysis above depends on ideas you want to understand first, these fundamentals make the conversation more legible:
