It is not often that markets punish companies for making too much money. But that’s been the fate of memory chip manufacturers, whose record earnings triggered a massive sell-off. This isn’t a case of market narratives “getting ahead” of fundamentals; Samsung’s Q2-2026 operating profit, to take the most prominent example, rose 19x(!) to $58 billion on an unprecedented 52% operating margin. The problem is the fundamentals themselves, which complicate customer economics to an extent that increases the probability of a bust ahead.
Like all other industries, the economics of AI need to be evaluated in terms of optimization, accounting identities, and budget constraints. When profits at memory chip manufacturers look this juicy, it naturally raises questions about the profitability of other links in the value chain. Can AI infrastructure providers (hyperscalers and neo-clouds) earn adequate returns on the capital they’re committing to serve frontier model providers (Anthropic and OpenAI)? Can those AI labs, in turn, convert API usage and enterprise adoption into revenues sufficient to justify their commitments? If not, “excess” memory profits constitute a tax on the downstream value chain, potentially slowing AI adoption and creating incentives for capacity additions and architectural workarounds.
The capacity trap
At the end of Q2-2026, the combined market capitalization of Samsung and SK Hynix reached 135% of South Korea’s GDP (Figure 1), up 5x from year-ago levels. The stock market begs these manufacturers to add capacity. And the companies (in conjunction with the Korean government) seem ready to oblige, intending to build four new fabrication plants at a combined cost of more than $500 billion. SK Hynix expects memory shortages to persist until those fabs come online.
Figure 1: Combined value of two memory chip manufacturers equal to ~ 135% of South Korea’s GDP
But this is an industry historically characterized by violent cyclicality. The Japanese corporate sector first consolidated and then exited memory chip manufacturing entirely because capacity additions tended to arrive at precisely the moment demand turned down. While there are reasons to suspect this Time is DifferentTM – prior memory booms were driven by price-sensitive, substitutable end markets (PCs, phones, conventional servers), while this one (high-bandwidth memory (HBM) for AI accelerators) exhibits no such elasticity – any market veteran could also provide, in excruciating detail, reasons to have believed the last time was also different.
The risks here are twofold. First is the non-trivial possibility that the application and enterprise layers fail to generate enough downstream value, potentially because customers won’t absorb the token prices necessary to recover higher training and inference costs. In this scenario, hyperscalers rationalize their capex budgets faster than capacity additions can be unwound. Second, there’s the chance that transformer architectures get reengineered in ways that reduce memory bandwidth requirements. Here, the dollar value of AI capex remains on its current trajectory, but HBM per unit of AI output falls sharply.
Public filings suggest that AI compute backlogs now exceed $2.2 trillion (Figure 2). For AI infrastructure providers, these contracts represent both an asset (the discounted present value of compute rental payments from AI labs) and a liability (the discounted present value of the hardware purchases, data center rent, and utility outlays necessary to deliver the promised compute).
Figure 2: $2.2 trillion revenue backlog at major AI infrastructure firms also creates capex obligations
The implied "equity" in the infrastructure layer looks thin: under reasonable assumptions, the net present value of the rental payments, after the capex required to deliver them, equals only about 15% of the backlog.1 A deceleration in AI revenue, a derating of the labs, or a loss of capital-markets access could raise the risk premium applied to these longer-dated, unsecured cash flows enough to make these “remaining performance obligations” more trouble than they're worth, imperiling demand for memory chips.2
More importantly, these backlogs are denominated in dollars of compute rather than units of memory. An AI lab’s multi-billion-dollar commitment to Microsoft, for instance, makes no reference to how much HBM stands behind it. The same demand inelasticity that generates those 50% margins also motivates hyperscalers to devote their financial resources, scale, and talent to engineering new, less memory-intensive accelerator designs. Unlike the prior memory demand shifts that played out over years, this one could prove larger and more abrupt given market concentration and adoption speeds.
Commodity market lesson
The stock market’s “memory trade” faded because this level of profitability compromises the economics on which future memory orders depend. It’s a situation comparable to those commodity producers often face, where prices and margins can rise too high for their own good. Added capacity could be even more important for memory makers’ customers because of the confidence it signals in the sustainability of the broader AI capex boom. But don’t be surprised if these costly capacity additions turn into “white elephants” when the industry’s next cyclical downturn inevitably arrives.
1 Source: Carlyle analysis. Each provider's backlog is decomposed by contract type and valued as gross billings × gross margin × a collection-and-survival factor × a duration discount factor × (1 − liquidity haircut). Central-case assumptions: blended gross margin of ~30–40% for AI-infrastructure contracts (higher for diversified SaaS/enterprise cloud, lower for thin-margin neo-cloud capacity); a weighted-average ~3.5-year conversion horizon; a 14% discount rate reflecting the long-dated, unsecured, and counterparty-concentrated nature of the cash flows.
2 Memory manufacturers' long-term "take-or-pay" agreements cover only a minority of volume. Micron says its Strategic Customer Agreements span roughly 20% of DRAM, leaving the uncontracted majority fully exposed. And these contracts run only a few years before repricing or rolling off, delaying rather than preventing the crash from any eventual supply glut.
Sign up to receive The Carlyle Compass in your inbox.
Economic and market views and forecasts reflect our judgment as of the date of this presentation and are subject to change without notice. In particular, forecasts are estimated, based on assumptions, and may change materially as economic and market conditions change. Carlyle has no obligation to provide updates or changes to these forecasts. Certain information contained herein has been obtained from sources prepared by other parties, which in certain cases have not been updated through the date hereof. While such information is believed to be reliable for the purpose used herein, Carlyle and its affiliates assume no responsibility for the accuracy, completeness or fairness of such information. This material should not be construed as an offer to sell or the solicitation of an offer to buy any security in any jurisdiction where such an offer or solicitation would be illegal. We are not soliciting any action based on this material. It is for the general information of investors of Carlyle. It does not constitute a personal recommendation or take into account the particular investment objectives, financial situations, or needs of individual investors.
Humanize & Differentiate
Insights & Trends>Macroeconomic Insights