Is it a failure of leadership, or a crisis of complexity? Both. Competition supplies pressure. Governance determines what the pressure is applied to

Capitalism is in crisis because competition has defocused from outcomes and refocused on accumulation. Firms still compete, intensely, at enormous expense. But competition no longer reliably converts private accumulation into broadly usable productive capability. Increasingly, the safer return lies in defending position (controlling distribution, acquiring challengers, influencing regulators and managing expectations) rather than in building capacity, the benefits of which travel through the rest of the economy.

What they compete for has shifted toward valuation, market position, individual wealth and measured performance. The decoupling of accumulation from productive value is uneven. Some sectors are well past it, some are nowhere near it, and no definitive before and after exists.

The threshold is the point at which a company can expect a higher or safer return from defending its position, duplicating a competitor, influencing a regulator, acquiring a challenger or managing a quarterly expectation than from building new productive capacity. Past that point, competition continues and intensifies. It stops reliably selecting for the thing it is defended for selecting.

Underneath the sectoral evidence sits a structural fact. A fragmented system contains thousands of firms and no durable mechanism for deciding which collective capabilities it needs. Every participant optimizes its own balance sheet, electoral prospects or institutional survival. Nothing decides what the system as a whole should be able to do in fifteen years, and nothing carries that decision across a change of government. There is a point at which that becomes a detriment and not an asset. This is not to say that individual ambition and entrepreneurship should be limited, but it does mean that incentives and direction are not being effectively applied.

Five failures, one name

Five distinct mechanisms get bundled together as one threshold. They don’t really co-occur, and treating them as a single phenomenon makes the diagnosis not very useful.

1. Duplication. Research on the Red Queen effect describes firms that must act continually because rivals are acting, with competitive responses erasing most of the relative advantage. Effort rises, position holds. Philippe Aghion, Nicholas Bloom, Richard Blundell, Rachel Griffith and Peter Howitt formalized the related finding that competition and innovation have an inverted-U relationship. Competition can be too weak to force innovation and also intense enough to discourage lagging firms while encouraging neck-and-neck firms to innovate. There’s a contradiction here because duplication also indicates that an innovation has a market beyond a single company, but there’s a balance between satisfying the market and excessive repetition.

2. Exploitation drift. James March distinguished exploration, which is experimental and uncertain, from exploitation, which refines what already works. Organizations drift toward exploitation because its returns are measurable and near. Success accelerates the drift, since existing products build internal constituencies that defend them.

3. Rent extraction. Thorstein Veblen separated industrial efficiency from business advantage and argued that restricting output or controlling a market can protect profit while inhibiting production. Mancur Olson supplied the political version: stable societies accumulate distributional coalitions that get better at protecting their share of existing wealth. George Stigler described how an industry acquires its own regulator and operates it for its own benefit.

4. Horizon compression. Stephen Terry’s work on corporate short-termism found that firms under pressure to meet earnings targets cut R&D and intangible investment, protecting the individual firm at the cost of aggregate growth. Giovanni Arrighi placed the same tendency inside a longer cycle in which mature capitalist centres shift from material expansion to financial expansion, with capital earning through ownership and leverage rather than new capacity.

5. Coordination failure. Paul Rosenstein-Rodan identified investments that fail alone and succeed together, and Kevin Murphy, Andrei Shleifer and Robert Vishny formalized the trap: no participant captures enough of the system-wide benefit to move first.

Sorting sectors by which mechanisms are active produces a more useful map than a single threshold does. Pharmaceuticals show rent extraction and exploitation drift. Legacy telecom shows rent extraction and coordination failure. AI shows heavy duplication and severe coordination failure with almost no horizon compression at the capital layer, which is why the standard short-termism account fits it badly.

The evidence supports a diagnosis and not a forecast. Research effort has risen while research productivity has fallen, with maintaining Moore’s Law now requiring more than eighteen times the researchers it took in the early 1970s. Ufuk Akcigit describes an American innovation paradox in which R&D spending expanded while growth slowed and dominant firms directed resources toward defending existing businesses. Median net investment across the OECD fell from 2.5 per cent of GDP before the financial crisis to 1.6 per cent. The OECD’s own firm-level work describes a broken diffusion machine, with leading firms advancing while technology spreads slowly to everyone else. Some of that decline reflects discoveries genuinely getting harder. The conversion machinery is weakening regardless.

Companies select, not calculate

The usual defence of markets assumes firms are rational actors reading price signals efficiently. Herbert Simon’s work on bounded rationality established otherwise. Organizations decide with incomplete information and limited cognitive capacity, they satisfice, and they optimize whatever measurement determines compensation, status and survival. Growth adds layers, filters information and makes bad news expensive to transmit.

Capitalism’s actual strength is decentralized selection. Many imperfect organizations try many things, and competition is supposed to reward the ones that find better methods. Markets select for whatever conduct produces the greatest return under the prevailing rules. When invention pays best, competition produces invention. When distribution control, acquisition, lobbying or expectation management pays better and pays sooner, firms adapt to that instead, and they do so without anyone deciding to stop innovating.

Concentration of personal control compounds the problem. Lucian Bebchuk and Kobi Kastiel argue that the benefits of founder control decline over time while its entrenchment costs rise, and that controllers retain the incentive to preserve the arrangement after it stops being efficient. Such firms can still innovate at extraordinary scale. A system that depends on unusually insightful individuals lacks a correction mechanism for the point at which those individuals become mistaken, distracted or politically committed.

What the AI numbers show

The 2026 Stanford AI Index reports United States private AI investment of $285.9 billion in 2025 against China’s $12.4 billion, a ratio of roughly 23 to 1. Stanford cautions directly that private investment understates China, whose government guidance funds deployed an estimated $184 billion into AI firms between 2000 and 2023 outside that tally. Read the ratio as a private-capital gap.

By March 2026, the measured gap between the leading American and Chinese models had narrowed to 2.7 per cent on the benchmark set Stanford tracks, down from a spread of 17.5 to 31.6 points in May 2023. American companies produced 59 notable models in 2025 against China’s 35, with China’s count doubling from 15 the year before. The United States funded 1,953 new AI companies, more than ten times the next country, and hosts 5,427 data centres. China produced 23.2 per cent of global AI publications, received 20.6 per cent of citations against 12.6 per cent for the United States, and filed the majority of AI patents worldwide. This year China will produce 100,000 humanoid robots while the US will produce between 3000 and 6000.

The structure inside those numbers matters more than the totals. Competition is fragmented at the model and application layers, where dozens of firms train overlapping systems, bid against each other for the same small pool of researchers and reproduce each other’s infrastructure. Concentration is extreme underneath, where nearly every leading AI chip still comes out of a single Taiwanese foundry. Twenty-three times the private capital bought a 2.7 per cent lead on the benchmarks and a dependency neither system has resolved.

Multi-decade assets on a quarterly technology

AI capex is the opposite of short-termism in the sense economists use the term. Hyperscalers are absorbing serious near-term earnings damage for bets that pay out over a decade, which is precisely what firms under earnings pressure stop doing. The mechanism operating here is different, and it splits by layer.

The model layer moves in weeks. Nobody is investing toward a known endpoint. Releases are reactive, roadmaps get rewritten on a competitor’s launch, and talent bidding responds to whichever architecture looked best last quarter. That layer is pure duplication under uncertainty.

The compute layer is contested. Whether accelerators are long-lived assets or a rolling expense is a live argument, conducted in public over depreciation schedules of five to six years, and the answer determines whether a large share of current capital expenditure is investment or consumption.

The physical layer is much longer. Power purchase agreements, substations, transformers, land, water and fabrication capacity commit twenty to forty years. Grid interconnection queues alone run longer than most model generations.

What the sector displays is a horizon mismatch. Assets with multi-decade lives are being committed on the basis of a technology whose requirements change quarterly. Firms are not being too short-sighted. They are making very long commitments against a target that will not hold still, and the resulting misallocation is a coordination problem rather than a patience problem.

Where innovation is flourishing

Innovation has obviously not stopped. More than 90 per cent of notable AI models in 2025 came from industry. Private laboratories continue to produce capability at a pace no public agency matches.

The geography has shifted. China entered the top ten of WIPO’s 2025 Global Innovation Index for the first time and hosts 24 of the world’s top 100 innovation clusters against 22 for the United States, with Shenzhen, Hong Kong and Guangzhou ranked first globally. South Korea placed fourth and Singapore fifth. China accounted for a little over half of global industrial robot installations in 2024, with the International Federation of Robotics putting the share near 54 per cent and Stanford reporting 51.1 per cent on its own accounting.

Both states hold enormous coercive authority. What differs is where it gets spent. Beijing’s goes into productive capacity. Washington’s goes into protecting incumbent positions. What the successful cases share is institutional: durable objectives, sustained investment in technical capacity, working links between government and industry, performance discipline attached to support, and coordination across research, finance, production, infrastructure and demand.

The same pattern appears inside Western capitalism wherever missions exist. Defence, space, advanced medicine, semiconductors and parts of energy all depend on public bodies that finance early research, specify difficult objectives, absorb initial risk and buy the first generation of output. Mariana Mazzucato describes this as market shaping rather than market repair. Alexander Gerschenkron showed late industrializers substituting coordinated institutions for gradual market evolution, Chalmers Johnson documented MITI’s coordination of Japanese industrial transformation, and Alice Amsden emphasized the reciprocity in South Korea’s version, where firms receiving support were held to production and export targets.

What coordination gets wrong

Planning fails in ways that are as well documented as market failure, and China supplies a clear recent catalogue.

Wuhan Hongxin absorbed billions in local government backing and collapsed in 2020 without producing a chip. Tsinghua Unigroup, once the vehicle for national semiconductor ambition, defaulted on its bonds and went through restructuring. Corruption investigations in 2022 reached the senior management of the national semiconductor fund itself. The model boom produced more than two hundred registered large language models, most of which no longer exist. Provincial data centre construction ran ahead of demand, leaving compute idle in places selected for political rather than operational reasons. More than a hundred electric vehicle brands entered a price war severe enough that regulators began pressing publicly against it.

Duplication, capture and defensive competition are available to coordinated systems too. The difference between Wuhan Hongxin and MITI is not the presence of a plan. It is Amsden’s reciprocity: conditional support, measured performance and the willingness to terminate failures. Coordination without termination can produces similar inefficiency to fragmentation.

Friedrich Hayek’s objection survives all of this. Knowledge is distributed, prices carry local information no central authority possesses, and planning is weakest where objectives are uncertain and progress depends on open-ended experimentation. AI at the model layer is Hayek’s territory. Nobody knows which architecture wins, which is exactly why racing firms are the right mechanism there and central selection would be a mistake.

Planning at the wrong altitude

Regulation is the obvious remedy and the least available one. Competition enforcement, limits on bottleneck control, restrictions on acquisitions intended to neutralize challengers, interoperability requirements and governance reform all require legislation that the interests it would enforce are effective at preventing.

Procurement looks like the alternative, since public purchasing already represents more than 14 per cent of GDP across the European Union and comparable shares elsewhere, spent under existing authority by officials who set the terms. The limitation is what sits above it. Every procurement has a plan, if not a vision. That plan is scoped to the transaction, sized to the replacement cycle, run by people whose mandate is to buy a thing rather than build an industry, and finished when the contract closes. Conditions attached to uncoordinated purchasing produce compliance paperwork. The machinery exists, is fully staffed, and is occupied with something small.

Capitalist economies do plan proactively, and the trigger is threat perception. DARPA, Apollo, the interstate system, Operation Warp Speed and the recent semiconductor programs all obtained what ordinary policy cannot: multi-decade horizons, tolerance for expensive failure, continuity across changes of government, and the authority to coordinate finance, research, production and demand at once. Security framing buys those conditions. Nothing else reliably does.

The pattern of what gets built follows the pattern of what gets classified as a threat. Semiconductors were reclassified and received a strategy. Generation capacity, transmission, water and technical workforce were not, and did not, though AI at scale depends on all four. Deciding national capability through threat perception works, has worked repeatedly, and selects for what frightens people rather than for what a society needs.

Political parties compete the way firms compete

Party-based democracy runs the same selection mechanism as the corporate sector, and it has drifted the same way. Competition is over position rather than delivery. New governments obtain credit by announcing programs, not by competently continuing their predecessors’ programs, so missions get reversed, renamed, split across departments or starved between budgets. Opposition is rewarded for preventing outcomes rather than producing them. The duplication, the defensive manoeuvring and the effort spent on relative advantage rather than absolute capability all reappear, with the same waste.

The observable output is rhetoric. Where competition selects for position, language becomes the product, and the volume and temperature of political speech rise while the delivered capability does not. Escalating rhetoric is what a system sounds like when both halves of it have stopped competing over outcomes at the same time.

Technical capability is now moving faster than the institutions deciding about it can absorb. The problem is not public comprehension. It is that the bodies holding the authority have limited technical capacity, and their decision cycles are shorter than the time required to understand what they are deciding about. Ministers rotate on eighteen-month cycles through portfolios covering technologies that take years to learn.

What that situation requires is direction capable of outlasting the person holding it. A system dependent on exceptional individuals has no correction mechanism for the moment those individuals are wrong, which is the same objection that applies to founder-controlled firms. Vision matters. Vision embedded in institutions with multiyear funding, stable mandates, technical staff and published performance measures is what converts it into capacity, and it is the part that market democracies have spent thirty years dismantling.

The Canada–United States trade war has become an almost literal demonstration of this selection failure. Two economies that spent decades integrating production are now dismantling parts of that integration in order to prove that neither can be compelled by the other. After negotiations collapsed over last-minute American demands that Canada says would restrict its independent trade policy and compromise matters of sovereignty, language and culture, the United States imposed 50 per cent tariffs on roughly $20 billion of Canadian goods. Canada responded with dollar-for-dollar retaliation, despite acknowledging that doing so will raise costs and reduce choice for Canadians. The asymmetry matters: the United States converted economic integration into coercive leverage, and Canada did not freely choose a pointless contest. But once partnership became leverage, declining to compete was no longer a viable option. Concession would demonstrate that further pressure works; retaliation imposes economic damage but preserves bargaining credibility and sovereign agency. Political and corporate capacity is consequently diverted into tariffs, exemptions, compliance, rerouting, inventory and retaliation rather than new production. The American system rewards the projection of dominance; the Canadian system rewards visible resistance to it. Both incentives are politically intelligible, and together they degrade a productive relationship from which both countries benefited. Competition is selecting for coercion and resolve rather than economic outcomes. The destruction is real, but so is the compulsion.

The foundation is legible

Coordination requires an objective clear enough to measure, which appears to disqualify a technology that reorganizes itself every quarter. The disqualification holds at the model layer and dissolves underneath it.

Generation and grid capacity, transmission, fabrication, domestic compute, technical workforce and the standards that let systems interoperate are all required regardless of which architecture wins. Their requirements are stable on exactly the timescale where private coordination fails and public coordination works. Canada is deciding right now how much of that layer to build and how much to buy, with the same decision facing every mid-sized economy that will not host a leading laboratory. The decision is being made through procurement, connection agreements and land approvals rather than through anything announced as industrial strategy.

So are we stuck?

The remedy that addresses the cause requires multiple interventions, including regulation that will likely not gain sufficient sustained support to pass. Procurement conditionality operates without legislation and only matters downstream of a decision no single person or body is positioned to make. Coordinated systems waste less in some sectors and more in others, and their advantage depends on a willingness to terminate failures that no system sustains for long.

What remains is threat perception, which does produce proactive planning in capitalist economies, reliably and at scale, and which arrives only after something has enough impact to become concerning enough to reclassify. When threat perception is the most reliable trigger for coordinated investment, a society will build a strange set of things: capabilities shaped disproportionately by what it fears rather than by the full range of what it needs.


Is this a failure of leadership? Partly, but really it’s a sign of increasing societal complexity, and the fact that we have evolved to the point where traditional mechanisms for managing competition, managing innovation, dealing with regulatory needs around industry and innovation have been reduced just as the number of forces requiring this kind of intervention have grown. We may need to become more deliberate in what we incentivize innovation and entrepreneurship around.

We may need to plan better where we need society to go and the goals that we are trying to meet. In many cases we are failing at meeting the most basic human needs, and very little innovation is devoted to trying to address that. The inequity in capitalist society is visible everywhere and this should be the first problem we try to tackle.