In his latest newsletter Greed & Fear, Wood describes the ongoing AI build‑out as “the most dramatic capex cycle” he has ever seen, driven by hyperscalers and foundries racing to ramp data‑centre and compute capacity. TSMC, for instance, has lifted its capex guidance for 2026 to about US$56 billion from US$41 billion last year, with Jefferies’ Taiwan partner Fubon Research now projecting US$65–70 billion of capex in 2027.
This surge in investment is already translating into boom‑like macro conditions in Taiwan, with real GDP growth hitting 14.55% year‑on‑year in 1Q26 and export orders up 53.4% year‑on‑year in the three months to May. Wood notes that AI‑related demand now accounts for an estimated 31% of TSMC’s revenues in 2026, underscoring just how concentrated the cycle has become in AI infrastructure.
Also Read | 3 AI stocks outweigh all of India: Why this concentration is sounding EM alarm bells
Jevons Paradox and “Picks and Shovels” Winners
Wood frames the AI demand story through the lens of Jevons Paradox: as token costs fall and efficiency improves, total compute consumption rises instead of falling. “The increased demand triggered by cheaper prices should be good for the picks and shovels plays,” he writes, arguing that DRAM and memory suppliers are the primary equity beneficiaries three and a half years into the AI capex arms race.
He cites Micron CEO Sanjay Mehrotra’s comment that “memory has evolved from a peripheral component into the core engine driving productivity in the AI era,” adding that the big three DRAM makers now have sufficient leverage to lock in long‑term sales agreements. Micron has already signed 16 strategic customer agreements covering roughly 20% of its DRAM volume and a third of its NAND volume, typically with five‑year tenors — evidence, in Wood’s view, of structural change in the industry.
The Commoditisation of AI Models
A key contextual risk is the rapid commoditisation of large language models, particularly in the consumer and, increasingly, corporate markets. Wood highlights the launch of GLM‑5.2 by Hong Kong‑listed Z.ai, formerly Zhipu AI, noting that informed sources describe the new model as “almost equal to Anthropic” for corporate use at just one quarter of the cost per token.This comes against a backdrop of a backlash against “tokenmaxxing” and explosive growth in cheaper Chinese models on platforms such as OpenRouter. In the week ended 21 June, top Chinese AI models processed 21.37 trillion tokens on OpenRouter, up from 4.37 trillion in late April, versus 5.76 trillion tokens for the top US models. That shift in volume, he argues, is already signalling a commoditised landscape and mounting pressure on the economics of premium Western AI providers.
Malinvestment in AI
Wood is explicit that the key vulnerability in the AI trade is not a classic semiconductor oversupply shock but the eventual recognition that the hyperscalers and leading AI labs will fail to earn a satisfactory return on their investment. “GREED & fear is personally convinced that concerns about malinvestment will be the most likely trigger for an end to the AI trade, or at least for a protracted pause to refresh,” he writes.The danger, in his view, lies in circular funding arrangements and aggressive capacity expansion built on optimistic monetisation assumptions. He points to structures such as Nvidia financing OpenAI so that OpenAI can in turn buy more Nvidia chips — a feedback loop that works as long as investors are willing to bankroll the ecosystem but could unwind sharply once doubts over long‑term returns take hold.
Why Traditional Supply‑Side Risks Are Secondary
Historically, semiconductor cycles have tended to end with abrupt increases in supply and inventory gluts. Wood believes the current cycle is different. “The key point to note for now is that this is the way the cycle is most likely to end rather than because of a sudden increase in supply, as has traditionally been the case in semiconductors,” he argues, emphasising that DRAM makers now command far greater pricing power.
The dominant three DRAM producers have been able to negotiate multi‑year strategic customer agreements, and Wood relays market chatter that Korean memory makers are already regretting locking in long‑term terms because they expect chip prices to rise further. In such a structurally tight industry, the more plausible end‑of‑cycle trigger is not oversupply but investor capitulation over capital discipline and earnings visibility in the AI stack above memory.
Despite these structural concerns, Wood emphasises that there is “zero sign of AI capex slowing” yet. He links ongoing spending to US banking deregulation under the Trump administration, citing Alvarez & Marsal’s estimate that recent regulatory changes will unlock US$2.5 trillion in additional lending capacity across the US banking system, including US$1.1 trillion unlocked in the last two quarters.
Portfolio Implications For Wood
Importantly, Wood is not calling for an immediate collapse in AI‑linked equities; instead, he is re‑positioning towards hardware and memory names that he believes will remain long‑term beneficiaries even if the AI trade endgame is defined by capital‑return disappointments higher up the stack.
He is lifting exposure to tech hardware across GREED & fear’s model portfolios, adding SK Hynix and Kioxia with initial 4% weightings in the global long‑only book and increasing the allocation to Samsung Electronics. Alphabet and Alibaba are being removed from the global portfolio, reflecting a deliberate tilt away from big‑cap platform plays towards “picks and shovels” beneficiaries of the AI capex cycle.
(Disclaimer: Recommendations, suggestions, views and opinions given by the experts are their own. These do not represent the views of The Economic Times)
