Why Malinvestment, Not Chip Oversupply, Could End the AI Trade

The artificial intelligence boom is driving the most dramatic capital expenditure (capex) cycle in history, but a significant warning has emerged regarding its sustainability. Chris Wood, Jefferies’ Global Head of Equity Strategy, suggests that the eventual end of the AI trade will not be triggered by a shortage of chips, but by a crisis of confidence in investment returns.

The Looming Threat of Malinvestment

Unlike traditional semiconductor cycles that end due to sudden supply gluts or inventory buildups, Wood argues that the AI era faces a unique structural risk: malinvestment. The primary danger lies in the possibility that hyperscalers and leading AI labs will be unable to generate adequate returns on the massive capital they are deploying.

Wood highlights a concerning "circular funding" pattern within the ecosystem. For instance, Nvidia has been involved in financing entities like OpenAI, which in turn use that capital to purchase more Nvidia chips. While this creates a powerful momentum in the short term, it creates a feedback loop that could unwind sharply if investors begin to doubt the long-term monetization and earnings visibility of the AI stack.

Massive Capex and the Concentration of Risk

The scale of investment currently being witnessed is unprecedented. TSMC, the world’s leading foundry, has raised its 2026 capex guidance to approximately $56 billion, up from $41 billion last year. Projections from Fubon Research suggest this could soar to between $65 billion and $70 billion by 2027.

This surge is already transforming regional economies. In Taiwan, the impact of AI-related demand is evident, with real GDP growth hitting 14.55% year-on-year in Q1 2026. Furthermore, AI-related demand is expected to account for roughly 31% of TSMC’s total revenue in 2026, underscoring how heavily the global economy is becoming concentrated in AI infrastructure.

The Commoditisation of AI Models

A secondary pressure point is the rapid commoditisation of Large Language Models (LLMs). As efficiency improves and costs drop, the "premium" edge of Western AI providers is being challenged. Wood points to the rise of Chinese models, such as Z.ai’s GLM-5.2, which reportedly offers performance comparable to top-tier Western models at just one-quarter of the cost.

Data supports this shift; on the OpenRouter platform, top Chinese AI models processed 21.37 trillion tokens in late June, a massive leap from 4.37 trillion in April. This volume is significantly higher than the 5.76 trillion tokens processed by top US models, signaling a crowded and price-sensitive landscape.

Shifting Focus to "Picks and Shovels"

Despite these warnings, Wood does not predict an immediate collapse. Instead, he suggests a strategic pivot toward the "picks and shovels" of the industry—specifically DRAM and memory suppliers. Due to Jevons Paradox, as compute becomes more efficient and cheaper, total consumption actually rises, benefiting hardware providers.

Major players like Micron are already securing their position through structural changes, such as signing strategic five-year agreements that cover significant portions of their DRAM and NAND volumes. This provides memory makers with greater pricing power and stability, even if the broader AI software layer struggles to prove its profitability.

Key Takeaways

  • The Real Risk: The AI trade is more likely to end due to "malinvestment"—the failure of hyperscalers to earn sufficient returns on massive capex—rather than a traditional chip oversupply.
  • Model Commoditisation: The rapid rise of low-cost, high-performance Chinese AI models is putting immense pressure on the economics of premium Western AI providers.
  • Hardware Resilience: Memory and DRAM manufacturers (like SK Hynix and Samsung) remain the most resilient beneficiaries due to their ability to lock in long-term sales agreements and command pricing power.