The unravelling of the AI trade
The AI trade’s renewed selloff last week is playing out as we have been fearing. While sell-side consensus appeared to remain fixated on quarterly earnings beats, it quickly became clear to investors that earnings results and near-term forecasts were increasingly looked at more as transient, as likely financing bottlenecks, which we have been emphasising as yet another key problem to sustained growth, came back to haunt the market after a brief initial excitement surrounding the Anthropic IPO filing.
Within our own market in Japan, with earnings across most semiconductor-related names, semiconductor production equipment (SPE) makers, and broader AI infrastructure rising sharply, the underlying mechanics of this correction are coming into plain view. Even though several large players have disappointed more bullish forecasts on the back of supply constraints, memory pricing issues, and weaker data centre-related compute deliveries, what we are witnessing is a severe price-earnings decompression. For now, this dynamic is poised to create classic value traps. As sentiment shifts decisively toward a more cautious narrative, share prices either drift sideways or fall outright even as reported earnings continue to move higher.
Memory stocks, and specifically Kioxia (285A) which has been another of our big contrarian short picks since June, which trades at under 3x forward earnings and saw its cautious post-market forward guidance disappoint just as we had anticipated, represent the most acute manifestation of this multiple contraction. While single-digit forward P/E multiples are hardly unusual for the memory space during mature phases of the cycle, trading below 3x is extraordinary by historical standards. Crucially, companies with meaningful High Bandwidth Memory (HBM) exposure rather than pure-play NAND portfolios, such as SK Hynix, are languishing around similarly depressed valuations, underlining that massive exposure to AI-driven DRAM demand has simply failed to insulate HBM names from a similar multiple derating.
In our experience these types of structural deratings almost never reverse on technical positioning alone. They typically require one of two catalysts, whether visibility improves substantially around credit conditions and overall debt absorption capacity, or earnings finally roll over so that compressed multiples zoom back up and signal that we have reached the genuine reset juncture. Until that clearing event occurs, optically low P/E ratios remain a big value trap rather than an entry signal.
This brings us to the core vulnerability surrounding financing and credit market absorption. Credit fund managers note that while yields offered by hyperscalers and related infrastructure issuers look attractive relative to US Treasuries on an isolated basis, with global AI-related debt issuance expanding rapidly across bond markets and private credit facilities, there are real limits to how much concentrated AI exposure credit portfolios can physically absorb. In our view, the most probable outcome is that debt issuance will have to retrench next year and with it, capital outlays. This appears inevitable. Only when spending realigns with underlying free cash flows can we reassess the sector and look to reload for the medium term.
Compounding these financing bottlenecks are significant physical and commercial headwinds which we have previously underlined as key problems facing the sector. Chief among them is the intensifying public and municipal opposition to data centre developments, a pushback that has expanded well beyond the United States into global jurisdictions. As regional power grids and local communities push back, the structural cost of compute will likely remain high, further elevated by the secular inflation of plus 20% within the AI supply chain which, as we have previously highlighted is eating into these huge budgets, adding even more to overall costs.
As we have also underlined for months, as cost of compute is rising, price of compute remains under pressure. With enterprise buyers facing steep token pricing from leading frontier labs, prompting an accelerating shift toward lower-cost open-weights models and competitive Chinese architectures, the commoditisation pressure at the application layer poses a direct threat to the premium economics required to justify the hyperscalers’ buildout.
Layered on top are escalating concerns surrounding model misalignment, frontier safety liabilities, and enterprise product risks, all of which threaten to drag on the cadence of new flagship model rollouts and dampen the velocity of infrastructure deployment across the board. These structural friction points will not disappear on the back of an engineered IPO valuation or another round of solid quarterly beats from the supply chain.
Underscoring the sheer scale of this mismatch, what has stood out for us is Gartner’s latest warning that over $3.5 trillion in cumulative AI infrastructure investments could be exposed to impairment and asset write-downs by 2030 should actual enterprise monetisation and end-market revenues fail to justify the capital outlay. As liquidity tightens and balance sheet discipline reasserts itself, the market is beginning to wake up to that reality.
