Why it’s too late to jump on the chip bandwagon

Taiwan Semiconductor Manufacturing Co

21/05/2026 12:04

The author is a Reuters Breakingviews columnist. The opinions expressed are his own

By Edward Chancellor

LONDON, May 22 (Reuters Breakingviews) - As investment in artificial intelligence surges, demand for semiconductors has gone through the roof. Prices for logic and memory chips have climbed faster and higher than ever before. The share prices of chip providers like Micron Technology MU.O and Samsung Electronics 005930.KS have risen in tandem. Supply is tight and analysts are excitedly hailing a new semiconductor supercycle. Investors should heed the advice of the late financier James Goldsmith: “By the time you see a bandwagon, it’s too late.”

The semiconductor industry historically suffered from a turbulent capital cycle. Building semiconductor plants, known as fabs, is an expensive business. As prices picked up companies increased capacity and new entrants entered the industry. When demand eased, semiconductor makers kept churning out chips, forcing down prices and leading to losses across the industry. Investors suffered. In the two decades between 1994 and 2014 the Philadelphia SE Semiconductor Index underperformed the Nasdaq Composite Index while experiencing greater volatility.

Around this time, however, the industry got its act together. A report from Bernstein Research, published in May 2013, observed that the memory chip makers were “in the midst of unprecedented structural changes; consolidation and increasing barriers to entry, combined with escalating technological uncertainty and reduced demand elasticity, pave the way for a new memory paradigm that no longer rewards aggressive investment.” The number of memory companies reduced from 12 players in the early 1990s to just three companies. Micron, Samsung and SK Hynix 000660.KS have more than 95% of the market.

The suppliers of semiconductor equipment have also consolidated, according to the research firm Pelham Smithers Associates. The Dutch firm ASML ASML.AS has a near monopoly on the extreme ultraviolet lithography machines used to manufacture the most advanced chips. Competition has declined in other parts of the supply chain, too. As a result, there are multiple bottlenecks to overcome when building a new fab. This benign shift in the capital cycle has boosted the profits of both equipment suppliers and chipmakers. Micron, says Smithers, has only lost money once in the past ten years, whereas over the previous decade it generated losses 40% of the time.

The AI revolution has exposed these supply constraints. Large language models require vast numbers of logic chips, known as central processing units, to work alongside the graphics chips made by Nvidia NVDA.O, another near-monopolist. Memory chips are required to train the frontier models and store the data generated by AI. Smithers claims that the amount of data that has been generated and stored globally is up more than twenty-fold over the past decade. Thanks to new AI-generated content it is expected to grow by around 30% in the current year.

Given the inelasticity of new supply, it’s not surprising that the price of Dynamic Random-Access Memory (DRAM) chips, which store temporary data, has risen six times over the past year. Micron’s share price is up by even more. Samsung and SK Hynix are also enjoying windfall gains. Investors have taken notice. Semiconductor, technology hardware and equipment makers now account for nearly 40% of the total capitalisation of the MSCI Emerging Markets Index, twice as much as three years ago. Retail investors are jumping on board: the Roundhill Memory exchange-traded fund has attracted billions of dollars of assets and climbed fourfold since its launch in early April.

Nvidia is locking up scarce chip supply with nearly $100 billion of purchase commitments. Dylan Patel, founder of the independent research firm SemiAnalysis, believes DRAM prices could double or triple from their current levels. Pelham Smithers describes a semiconductor supercycle that could last for years. What could possibly go wrong?

Well, for a start the semiconductor capital cycle is shifting. SK Hynix, Samsung and Micron are all building new fabs, although these facilities are not expected to start operations for another 18 months or so. Taiwanese chip giant TSMC 2330.TW will invest twice as much in U.S. dollar terms next year as in 2024. Micron’s ratio of capital expenditure to depreciation – a key capital cycle measure – will rise to 3 times by 2027, more than twice its 10-year average. Furthermore, two Chinese chipmakers, Yangtze Memory Technologies (YMTC) and ChangXin Memory Techologies (CXMT) are preparing to raise capital from their forthcoming initial public offerings to expand capacity.

The biggest risk, however, may be on the demand side. The fate of semiconductor companies depends on U.S. tech giants continuing their investment splurge on new data centres. Morgan Stanley predicts that Amazon.com AMZN.O, Alphabet GOOGL.O, Meta Platforms META.O, Microsoft MSFT.O and Oracle ORCL.N will collectively invest more than $1 trillion in 2027. This capex frenzy is consuming nearly all their free cash flow. Furthermore, the implied internal rate of return on these investments is negative for all the firms except for Amazon, according to a recent report by Panmure Liberum. The UK investment bank argues that these so-called hyperscalers will need to boost their revenue by between $2-5 trillion to justify their current spending plans. That’s a tall order.

The economics of AI are dubious, to put it mildly. Large language models are extremely expensive to build and costly to run. So far, AI companies have heavily subsidised the price of tokens – the basic unit of compute for AI models. OpenAI’s cost of inference – the process whereby the models apply their pre-training to new data – vastly exceeds its revenue, says Panmure. Losses cannot spiral upwards indefinitely. OpenAI recently closed its video creation model, Sora, because it was costing more than $5 billion a year, according to Julien Garran of The MacroStrategy Partnership. Microsoft’s GitHub, a platform of software developers, announced last month it was moving from a fixed monthly charge to usage-based billing.

Since OpenAI launched its ChatGPT chatbot in late 2022, companies have been desperate to prove their AI chops, regardless of cost. But the models, beset by hallucinations, have proved unreliable and productivity gains to date have proved elusive. As token prices rise to cover costs, demand is likely to contract. The hyperscalers will then be forced to backtrack on their grandiose investment plans. At that point, the semiconductor bandwagon will come to a screeching halt.

Follow @Breakingviews on X.

Semiconductors initially lagged tech stocks, but have soared recently https://www.reuters.com/graphics/BRV-BRV/gkvlkyzlapb/chart.png

Chipmakers dominate the MSCI Emerging Markets Index https://www.reuters.com/graphics/BRV-BRV/gdvzalemgpw/chart.png

(Editing by Peter Thal Larsen; Production by Shrabani Chakraborty)

((For previous columns by the author, Reuters customers can click on CHANCELLO/edward.chancellor.bv@gmail.com))