The Price of Abundance

Part one of a diptych on how China and America finance capacity nobody is paid to count, and who ends up holding it.

ChangXin Memory Technologies lost money for nine years. In July it raised 57.9 billion yuan on Shanghai's STAR Market, the largest offering in the board's history. At 8.66 yuan a share the company was worth close to 580 billion yuan before a single lot had traded. CXMT closed its first day at 49 yuan on a freely tradable float of 6.73 per cent, implying a total market capitalisation of 3.28 trillion yuan. The company itself received 57.9 billion yuan, much of it intended for new production capacity in an industry whose present profits depend on scarcity.

Every institution involved can defend its part in that conversion according to the measure assigned to it. Taken together, those measures allow each participant to succeed while leaving the market they are enlarging outside anyone's account.

The public institutions that carried CXMT's losses measured their return in capability. The national guidance fund's own literature describes it as an early bearer of risks no private institution will touch, operating on a whole-portfolio logic in which one company's failure need not invalidate the programme. Nine years of losses, including an inventory writedown of 11.5 billion yuan in 2023 alone, enter that account as the cost of acquiring capability. Hefei measures the same investment by industrial presence. Its vehicles hold at least a third of CXMT and control the main fabricating subsidiary through concerted-action agreements that turn 30.68 per cent of the economics into 73.01 per cent of the votes. A working fab is a municipal return, whatever memory earns once the shortage ends.

The institutions bringing that capability to market face much narrower tests. The underwriter's horizon is written into the offering terms. CICC, the authorised joint lead underwriter, took a fee and an over-allotment option worth 15 per cent of the issue, giving it thirty days to buy CXMT stock back at no more than 8.66 yuan — a safety net with a calendar on it. Exercise the option, or let the month run out, and that authority ends. The exchange and the regulator had to decide whether CXMT qualified for listing. Utilisation had climbed from 87 per cent to roughly 96, while CXMT forecast first-half revenue of 110 billion to 120 billion yuan and profit attributable to shareholders of 50 billion to 57 billion yuan, against a 2.3 billion yuan loss in the same period a year earlier. Approval was consistent with the figures before them.

What none of these institutions has been asked to decide is whether the capacity financed by the flotation will remain profitable once it reaches the market. That question belongs to the industry as a whole. An individual project can be defensible and a listing can meet every regulatory test while dozens of similarly justified investments create a market in which nobody earns an adequate return. Responsibility for that aggregate risk has no obvious institutional home.

China's industrial bureaucracy knew the shape of this problem long before CXMT came to market. In October 2013 the State Council issued Document 41, naming five industries in severe excess, from steel and cement to shipbuilding. Government departments were barred from approving new capacity projects in those sectors, while banks were instructed to withhold credit from projects lacking the necessary permissions. Construction that had not begun was forbidden to start. Work already under way could proceed only after central review.

The restrictions extended beyond administrative approval. Companies behind unauthorised projects could neither issue bonds nor list on a stock exchange. New capacity had to be matched by the retirement of old plant, with tighter requirements across the Beijing-Tianjin-Hebei, Yangtze River Delta and Pearl River Delta regions. Document 41 also instructed the relevant ministries to build a dynamic monitoring system and an early-warning mechanism for industrial excess within five years. Beijing had named the danger and assembled a substantial body of preventive powers before the latest wave of overinvestment began.

Document 51 followed in 2015 and attempted to reinforce those powers through the financing system. It raised the minimum capital investors had to put into a project, capping how much could be borrowed. Steel and electrolytic aluminium were set at 40 per cent, with cement at 35. Coal and several energy-intensive chemical industries received a 30 per cent floor. Polysilicon was included among them, and financial institutions were told to enforce the earlier restrictions wherever severe excess had already been identified.

Polysilicon consequently faced a specific financing barrier from 2015. Chinese manufacturers nevertheless went on to build somewhere between three-quarters and nineteen-twentieths of world capacity, roughly twice global demand. Module prices fell by about half in 2023 and continued downward. The price of polysilicon dropped from more than 300,000 yuan a tonne to around 30,000, while thirty-one listed solar companies lost a combined 12.6 billion yuan in a single quarter of 2025.

The capital-ratio requirement remained in force throughout. It governed the financing of identifiable projects presented to banks, while local governments could support an industry through their own investment vehicles and provincial credit. Restricting the leverage available to one proposed plant did nothing to remove the political reward attached to making a city a solar manufacturing hub.

By 2025 Beijing was intervening again through an anti-involution campaign covering ten sectors. The target was competition that kept expanding output long after the profit in it had disappeared. Two-year plans sought to hold output growth below its 2024 level, while a draft national energy-consumption standard set a compliance deadline for inefficient polysilicon producers, with closure the penalty for missing it. The planning commission and industry ministry also endorsed a proposed 50 billion yuan fund to buy obsolete capacity and remove it from production. Public money was being considered to retire plant that public policy had helped to finance.

The early-warning mechanism commissioned in 2013 had been due in 2018. The forceful measures arrived after the investment had been made and prices had collapsed. The state could discipline a sector once falling prices and losses made its excess impossible to ignore. Preventing local institutions from producing that excess while industrial expansion still brought political rewards proved harder.

Naming a sector did not change the incentive beneath it. Michael Pettis makes the same point structurally: suppress involution in one named sector and the pressure relocates, potentially to batteries, ships, steel or petrochemicals. Local governments kept looking for industries to grow, and capital moved into the one that had not yet been classified as excessive. The utilisation rates that caused five industries to be named in 2013 ranged from 71.9 to 75 per cent. Chinese manufacturing as a whole was operating at about 74 per cent in 2025. Whether that figure described strategic expansion or dangerous excess depended increasingly on which industry produced it.

Memory is the clearest example. Beijing regards domestic semiconductor capacity as insufficient for national purposes, whatever the eventual commercial balance between supply and demand. Investment in memory continues to count as a contribution to technological self-reliance while comparable investment elsewhere is condemned as involution.

In September 2023 Xi Jinping told cadres to develop 新质生产力, new quality productive forces, prompting local governments to compete for the industries associated with them. Two years later he criticised excessive investment in some of those same fields. The fourth plenum communiqué of October 2025 continued to demand faster progress towards manufacturing and technological self-reliance while the 反内卷 campaign was under way. Cadres could be criticised for the volume of their investment without being released from the strategic direction that encouraged it.

Semiconductors were absent from both the 2013 restrictions and the 2015 capital schedule. The Semiconductor Industry Association told the US Trade Representative last year that the anti-involution campaign was not directed at China's chip industry. Domestic-content requirements across several levels of government create demand for Chinese chips while the state restrains supply in industries already judged excessive. Document 41 gives Beijing the power to prevent a company in a designated sector from listing on a stock exchange. Memory has not been designated, and CXMT's flotation proceeds inside that exemption. The same paperwork that would have grounded a shipbuilder waves a chipmaker through.

The exemption secures access to capital without securing the return. A household purchasing CXMT relies on Beijing continuing to treat memory as a strategic deficiency and on global producers maintaining the restraint that supports DRAM prices.

The scale of what the exemption permits can be measured. CXMT ran about 100,000 DRAM wafer starts a month at the beginning of 2024 and around 290,000 by the end of 2025. Analysts modelling its trajectory expect roughly 350,000 by the end of 2026, within touching distance of Micron's estimated capacity and enough to make China the world's second-largest DRAM production base. Global industry capacity is on course to rise by something near a fifth by 2027 against commodity demand growth in the low twenties, a margin thin enough that the destination depends almost entirely on how much of the new Chinese output arrives on schedule.

More telling is what Beijing is doing with the technology it has bought. CXMT is reported to be under official pressure to share its DRAM process with JHICC, Swaysure and YMTC's subsidiary XMC, each of which has built domestic memory capacity or intends to. The instrument that would restrain aggregate supply is being used to multiply the number of producers. Self-reliance is measured in the count of capable firms; the market is measured in bits. Nothing in the apparatus converts one number into the other.

Against that stands the oligopoly the new capacity is entering. Three companies have controlled around 90 per cent of the market for a decade and, since 2016, have generally avoided the ruinous price competition that characterised earlier cycles. High-bandwidth memory has made that discipline unusually profitable. HBM absorbs close to a quarter of the leading manufacturers' wafer input while producing only about 9 per cent of their total DRAM bits, since stacking and yield loss make each bit consume roughly three times the wafer area required for DDR5. As capacity moves towards HBM, less remains available for ordinary server and consumer memory. The shortage travels down through the product range.

CXMT's first-day valuation assumes that squeeze will last long enough for the company to expand without destroying the prices that justify its expansion. In earlier cycles the shortage ended when one producer decided that gaining market share was worth accepting lower prices, forcing the others to follow. Samsung's chairman is reportedly arguing for a more aggressive approach now. AI demand may prolong the present cycle, but some of the scarcity supporting that valuation remains at the discretion of its competitors.

The composition of CXMT's earnings shows how much of it does. Analysts decomposing the first quarter of 2026 put the company's bit shipments up around 11 per cent while average selling prices rose by roughly 57 per cent, following quarterly price increases of a similar order through the second half of 2025. The move from a 2.3 billion yuan loss to a forecast 50 billion yuan of half-year profit is very largely a price event. Investors buying at 49 yuan are capitalising a number produced by the shortage rather than by any decisive change in the company's position within it. The prospectus does not conceal this. Average selling prices fluctuated by 55 per cent in 2024 and 34 per cent in 2025, and CXMT warns that a reversal in the balance between supply and demand could make its first-half growth unsustainable. The reversal is described in the document selling the shares.

CXMT accounted for 7.67 per cent of global DRAM bits by its own account and ranked fourth; independent estimates put it between eight and nine per cent. It supplies Xiaomi, OPPO, vivo and Transsion, while PC manufacturers have reportedly booked its capacity into 2027. The case offered to investors is that a company this size cannot yet move the global price. Memory prices, however, are set at the margin. A producer need not dominate the industry to push them down, especially in the lower-value products where previous collapses began and where CXMT is strongest.

CXMT was built from the remains of a company killed by exactly such a cycle. Qimonda, the memory arm spun out of Infineon, entered insolvency in January 2009 after an earlier DRAM price war. Its collapse ended Germany's last serious position in the industry. CXMT founder Zhu Yiming later acquired roughly 2.8 terabytes of Qimonda's technical files and spent something close to $2.5 billion turning them into a Chinese manufacturing platform. Engineering that had lost its commercial home in Germany acquired another life in Hefei.

Qimonda's history shows what a memory downturn can destroy. It does not make DRAM another version of solar, and the differences deserve their weight. Once a solar-production process has been mastered, additional capacity quickly produces modules that buyers struggle to distinguish. Memory fabrication requires continuing improvements in yield and qualification as each generation arrives, so a new fab adds physical capacity before it necessarily adds a producer capable of competing with the technological leaders. The Chinese ramp also faces a constraint solar never had. Advanced immersion lithography remains an import, domestic scanners are not expected in volume production before late 2026 at the earliest, and proposed American legislation would tighten access further. The most plausible case against a glut is that the equipment will not arrive fast enough to produce one.

The Korean industry supplies the successful precedent. Samsung entered memory as a state-supported latecomer and survived losses widely expected to finish it before becoming dominant. Creditor banks rescued Hynix after 2001, and its subsequent success turned that intervention into a profitable act of patience. These cases give substance to Hefei's treatment of CXMT's early losses as the purchase price of industrial capability.

The conditions facing CXMT are less forgiving. Samsung advanced while the Japanese incumbency was fragmenting, and Hynix was rescued as the memory industry consolidated in its favour. CXMT confronts the winners of those episodes, technically ahead and made stronger by a decade of oligopoly. Hefei is trying to repeat the Korean wager against companies that are themselves the product of it.

Zhongji Innolight went to Hong Kong three days after CXMT, seeking as much as $7 billion, and shows what the same boom looks like where the aggregate question does not arise. Zhongji was already the heaviest weighting in the CSI 300 and the world's largest optical-interconnect supplier by revenue. Its first-quarter net profit exceeded its earnings for the whole of 2024. It is raising money to serve demand it already holds, at a technological lead it has already demonstrated, in a market where its expansion displaces nobody's assumptions but its own. No public institution carried it to that position and none is measuring its return in capability. Whatever Zhongji builds, the risk of building too much sits with the company that built it.

CXMT is the opposite arrangement. Its expansion is underwritten by institutions whose returns are already booked, into a market whose balance is being set by three firms it does not control and by a domestic technology-diffusion programme it does not direct. The distinction is not between a strong company and a weak one. It is between a firm that owns the consequences of its own capacity decision and a firm whose capacity decision has been distributed across a dozen accounts, none of which records the total.

That distribution is testable. The lock-up schedule will eventually show whether CXMT's state shareholders behave as financial investors: substantial selling by Hefei's vehicles or the guidance fund into a strong market would mean they, too, had been managing the cycle rather than acquiring capability. An output ceiling or binding energy limit imposed on memory while the shortage continued would be stronger evidence that Beijing had learned to restrain aggregate capacity before collapsing prices forced it to act. A retreat from the technology-transfer programme would be stronger still. None of the three has happened.

A favourable outcome remains possible. Accelerator demand may consume bits quickly enough for CXMT's new capacity to arrive without producing a glut, allowing the fourth firm to improve its technology while continuing to earn an adequate return. Hefei's industrial objective and the shareholder's financial interest would then converge. No previous computing cycle has sustained a comparable compound increase in bit consumption for much longer than three years, and the first-day valuation leaves little room for the present one to weaken.

The timing of the flotation admits a simpler explanation than any deliberate transfer of risk. CXMT sought a listing when its operating figures were strongest, and regulators approved an issuer that met their requirements. The sequence required no unusual ingenuity in Anhui.

The uncomfortable part is how little incompetence it required. The fund and the municipality could record success long before the underwriter completed the offering. Because each institution kept its own account properly, the eventual excess cannot be traced to a defective transaction or a negligent official whose removal would repair the system.

Hefei will retain engineers trained during the nine loss-making years and a fab that is now up and running. Those are real industrial gains, and a fall in CXMT's share price would not erase them. If new capacity eventually destroys the margins that financed it, the financial cost will remain. The experience of solar suggests that part of the loss can stay with local investment vehicles and the guidance fund while the state banking system carries the rest.

Public institutions can postpone acknowledging that cost while the benefits acquire concrete form in clean rooms and trained hands. Whether the capability justifies the money consumed can remain unsettled for years. Households buying today, mostly at prices well above 8.66 yuan, have no comparable interval, and once scarcity passes the industrial gain and the financial loss will meet in the price blinking on a screen.