When The Future Becomes Collateral
What happens when a society can no longer afford the future it has promised?
When The Future Becomes Collateral
South Korea will close forty-nine schools this year. Most are in rural areas, where the decline in children has been arriving faster and with less ceremony. A school does not need to become completely empty before the arithmetic turns against it. It needs only too few pupils to justify the building, the teachers and the journey required to keep it open.
A few months later, I was reading about the Korea Composite Stock Price Index, usually shortened to KOSPI. It is South Korea’s principal share-market benchmark, roughly comparable to the S&P 500 in the United States. The index rose nearly 28 per cent during the first half of 2025, its strongest first-half performance in twenty-six years. Political stabilisation and expectations of corporate and market reform helped drive the rally, so it would be too simple to call it an AI boom. The country still had fewer classrooms worth keeping open and one of the hottest stock markets in the world.
I kept trying to treat those as separate stories. One belonged to demography and the other to finance. One concerned children who had not arrived, while the other concerned adults purchasing claims on future earnings. The separation held until I began asking what those earnings were expected to come from, and what the people buying them thought the future was now made of.
The House Moves First
Calling an investor reckless is easiest once the investment has gone wrong. The word arrives after the safer choices have disappeared from view. During the pandemic, South Korean household borrowing through banks for housing, investment and living expenses reached roughly the size of the country’s annual economic output. Housing remained the largest pressure, but younger borrowers were also using leverage to enter property and equity markets after wages and patient saving stopped looking like reliable routes into adulthood.
The geography made the arithmetic stranger. Schools were closing where children and young adults had left, while secure work and housing pressure remained concentrated around the capital. Private education consumed another part of the household budget. The OECD found that nearly four in five Korean schoolchildren used private tutoring in 2023, costing parents about 10 per cent of disposable income on average. Long and inflexible working practices remained difficult to combine with care, and women continued to absorb severe career and income losses after childbirth. The fertility rate rose from 0.72 children per woman in 2023 to 0.75 in 2024, its first increase in nine years, but deaths still exceeded births by about 120,000. A movement can be welcome without being a recovery.
The stock market begins to look different from inside that pressure. It does not need to offer a dependable route to wealth. It needs only to appear more open than the route through wages, property and slow accumulation. Samsung Electronics and SK Hynix did not explain the whole KOSPI rally, but their role in supplying high-bandwidth memory for AI systems connected Korean investors to a much larger machine being built elsewhere. A share in a memory manufacturer could be an investment and an escape plan at once. The companies were selling parts of systems expected to change the price of labour itself. The classroom and the market were not opposites. Both were responding to a future that no longer seemed likely to arrive through the usual route.
The Move Nobody Had Taught
On 10 March 2016, AlphaGo placed a black stone on the fifth line during the second game of its match against Lee Sedol. The move became known as Move 37. DeepMind later described it as a choice with roughly a one-in-ten-thousand likelihood of appearing in human play. Professional commentators initially struggled to explain it, although the stone proved decisive much later in the game.
The machine had not simply calculated a familiar strategy faster. It had found a valid move outside the habits accumulated through centuries of expert play. That did not make the machine mystical. It made the inherited map incomplete, which was awkward enough. The response in China carried a different temperature from much of the reaction in the West. Go had a deep place in Chinese culture, and AlphaGo became a public image of national technological distance. The match did not create Chinese AI policy by itself, but it supplied a rather efficient picture of falling behind.
When the State Council issued its national AI development plan in July 2017, it set staged objectives through 2030 and placed artificial intelligence inside manufacturing, medicine, cities, agriculture, defence and economic competition. Move 37 gave a visible shape to a strategic shift already gathering force. The important change was not from human intelligence to machine intelligence. It was from treating AI as a product to treating it as capacity. A product can be purchased when required. Capacity has to exist before the moment when its absence becomes decisive.
The Stone Given Away
DeepSeek-R1 arrived in January 2025 and disturbed two assumptions at once. The first was that advanced reasoning would remain confined to the largest American laboratories with the loosest access to new hardware. The second was that frontier capability would remain available mainly through expensive hosted services controlled by a handful of companies. DeepSeek released R1, R1-Zero and six distilled models, reporting results comparable with OpenAI’s o1 family on specified mathematics, coding and reasoning benchmarks.
Markets reacted before anybody had settled what the release meant. Nvidia fell just under 17 per cent in one session, losing about $593 billion in market value as investors questioned the expected scale of spending on chips, power and data centres. Some of that response was plainly theatrical. DeepSeek had not trained a frontier model on warm feelings and pocket change. It still depended on substantial hardware, previous research and a wider technical organisation. The release nevertheless showed that useful reasoning might become cheaper faster than the valuations surrounding it had assumed.
Open weights do not mean the entire process has been given away. The released parameters allow others to run, inspect, modify and fine-tune the trained model, but they do not necessarily include the original training corpus, every filtering decision or the complete infrastructure that produced it. The Open Source AI Definition makes that distinction explicit. Electricity and memory have also declined to join the open-source movement.
What changes is the location of control. A closed service keeps the model, pricing and permitted behaviour with the provider. An open-weight model can sit inside another organisation’s infrastructure, beside its own information and subject to its own engineering decisions. This looks like a poor strategy only when the model layer must capture most of the value. China’s 2017 plan was explicit about using AI to upgrade wider industry, including manufacturing, vehicles, robotics and physical infrastructure. Giving away a stone can be sensible when it changes the shape of the board. Cheaper models may reduce margins on scarce intelligence while increasing demand for hardware, energy, applications, factory systems and technical standards. The release is not necessarily the gift. The position created around it may be the thing being purchased.
The Worker Who Needs No Room
South Korea already has the world’s highest manufacturing robot density. The International Federation of Robotics recorded 1,012 industrial robots for every 10,000 manufacturing workers, more than twice the density found in several other major industrial economies. Its electronics and automotive sectors made automation sensible before demography made it urgent.
A robot does not replace every worker one for one. It may extend the people who remain, take over a narrow task or allow a factory to preserve output with fewer hands. It can weld, lift, inspect and repeat without requiring a flat near the factory, a nursery place or several years of private tutoring before becoming productive. This is excellent news for the robot. The person remains more difficult. A worker needs wages that can support a household, time outside work and some reason to believe that the next decade will be less precarious than the current one. Automation may help an employer preserve production while doing little to repair those conditions.
That produces a loop which is rational at every individual step. Difficult work and housing conditions weaken family formation. Fewer future workers strengthen the case for automation. Automation protects output and raises the value of firms supplying it. The market then offers investors exposure to the machinery expected to compensate for the missing labour. The proposed cure benefits from the persistence of the condition. South Korea is making related wagers at two levels. Nationally, it is betting that machines can preserve production as the population ages. Privately, its markets offer claims on companies expected to supply those machines. What appears from a distance as an industrial strategy begins to resemble insurance against a future the household economy has made harder to produce.
The Nursery Trades Higher
The American AI industry is financing a larger version of that insurance policy. Alphabet, Amazon, Meta and Microsoft were expected to spend about $320 billion in capital expenditure during 2025, roughly thirteen times their combined spending a decade earlier. Companies once valued for turning software into revenue with little additional physical cost have become rather enthusiastic builders of power-hungry industrial estates.
The investment may prove justified. AI does not need to become conscious or invent a warp drive for the economics to work. It needs to reduce enough labour, accelerate enough processes and create enough new demand to cover the machinery beneath it. A real technology can still be financed badly. Railways changed countries while ruining railway investors, and the internet outlived companies whose valuations assumed every visitor would become revenue. DeepSeek makes the distinction harder to avoid because intelligence may become cheaper while the infrastructure required to produce and distribute it continues to grow.
That would not end the AI industry. It would move the value. Model providers might capture less, while chipmakers, energy suppliers, data-centre operators, application companies and organisations holding proprietary information capture more. South Korea makes that movement unusually visible because American data-centre spending passes through demand for Korean memory and into expectations surrounding Korean industry. The productivity belongs to the future. The factories, data centres, savings and debt committed to it belong to the present. The future has become collateral before it has become income.
Cronus Builds a Factory
Production and continuity are easy to confuse because both can be expressed as growth. A robot can preserve the number of vehicles leaving a factory. A language model can reduce the time required to examine a document or answer a customer. Neither decides who owns the productivity, who receives the saved time or whether the people living beside the system can afford to form another household. The aggregate number may improve while the domestic calculation remains unchanged.
The ancient story of Cronus captures the precise shape of that failure. In Hesiod’s Theogony, Cronus was warned that he was fated to be overthrown by his own child. Governed by the terror of a rising generation that would inherit his power, he chose to devour his offspring as they were born. The act was a desperate attempt to freeze time at its peak-to preserve his own authority by destroying the very possibility of succession.
The modern system acts on that same fear, though it is quieter and, as a result, easier to place inside a policy document. It, too, consumes the future to protect the present. It devours affordable homes, stable work, secure time and confidence in adulthood. It then notices that too few children are arriving and commissions machines to replace them. The machine did not create the condition. It receives the work left behind by it.
Move 37 gave a public image to a strategic shift already gathering force in China. DeepSeek has shown that the model layer may become difficult to keep scarce. South Korea is already highly automated while its rural schools close and its households search for another route into security. The United States is building the physical infrastructure for intelligence on the assumption that future productivity will justify the present cost. It probably will arrive in some form. The harder question is what that productivity will preserve once it does. The factory can continue after the classroom empties. Its lights may even be brighter.
Comments