Semiconductor shares have retreated after what SemiAnalysis described as the industry’s strongest first half on record, exposing how quickly a rally tied to AI infrastructure can reverse when leverage, pricing expectations and capital-spending assumptions begin to shift. In South Korea, the KOSPI had fallen about 40% from its peak, according to the semiconductor research firm, while leveraged retail positions added forced selling pressure during the decline.
The selloff was the focus of a July 29 SemiAnalysis podcast featuring Doug O’Loughlin, an analyst at the firm, and Dylan Patel, its founder. Their discussion placed the downturn in a familiar chip-industry pattern: demand accelerates, supply appears scarce, valuations stretch, and even a modest change in the expected pace of pricing or capacity growth can trigger a sharp reassessment.
O’Loughlin compared the market’s earlier behaviour with Taiwan’s late-1980s equity bubble, when some bank shares traded at roughly 500 times earnings. His point was less about a direct comparison between banks and chipmakers than the speed of the preceding move. Rapid gains can leave little room for disappointment when earnings, memory prices or AI demand do not continue rising at the pace traders had assumed.
Memory pricing expectations have shifted
Memory chips have been one of the clearest beneficiaries of the AI buildout, particularly high-bandwidth memory used alongside advanced AI processors. O’Loughlin said memory prices rose by roughly three times last year, but SemiAnalysis now expects increases of about 30% to 50% next year.
That would remain a substantial rise in most industries. In a market priced for continued extreme scarcity, though, a move from tripling prices to gains measured in tens of percentage points can change how traders value memory producers and their suppliers.
O’Loughlin linked part of that shift to SK Hynix moving a larger share of production into long-term agreements, or LTAs. Such contracts can give manufacturers more predictable revenue and provide customers with better access to constrained components. They can also reduce the visibility of spot-market price changes that participants have used as a signal for the strength of the memory cycle.
The memory market has repeatedly moved between shortage and oversupply. When availability tightens, customers may order more chips than they immediately need to protect their supply chains. Producers respond with new capacity, but fabs take years to construct and equip. If demand later normalizes while additional output arrives, factory utilization can drop sharply. O’Loughlin described a possible fall from full utilization to 50%, a level that can put heavy pressure on pricing as manufacturers seek cash flow from expensive facilities.
AI demand remains harder to model than chip supply
The debate is complicated by AI usage, which Patel argued is growing fast enough to make conventional demand forecasts unreliable. He cited internal usage data following the rollout of a coding agent, saying the number of active users rose from nine to 90 while token consumption per user increased roughly tenfold over three to four months. That combination produced a 100-fold increase in AI spending in his example.
Tokens are units of text or data processed by an AI model. Rising token use generally means more computing demand, though the cost of each token can decline as models and hardware become more efficient.
O’Loughlin said the range of possible demand outcomes remains unusually wide. The central question, in his view, is whether AI usage settles at a level 10 times higher than current levels or reaches 100 times higher. Supply is easier to map through announced fabs, data centres and equipment orders, but translating user adoption into a durable revenue stream is far less straightforward.
That uncertainty has become central to the valuation of companies building AI infrastructure. Semiconductor makers can point to orders, backlog and capacity constraints, while cloud operators must show that their data-centre spending will eventually produce revenue beyond experimental deployments and subsidised services.
Capital needs are meeting physical limits
SemiAnalysis estimated that the AI ecosystem has accumulated about $1 trillion in capital expenditure against approximately $150 billion in annual revenue. O’Loughlin used a hypothetical 50% profit margin to calculate an implied return of around 7.5%, arguing that annual revenue would need to reach roughly $500 billion to better support that spending base.
The calculation is an illustrative framework rather than a forecast, but it captures the pressure facing the sector. High spending can be justified when revenue scales rapidly and infrastructure remains productive for years. The economics become less attractive if hardware is replaced quickly, utilisation remains low, or competition drives down the price of AI services.
O’Loughlin also described a much larger later-stage scenario in which $5 trillion in investment generates $50 billion in revenue, producing an exceptionally long payback period. The contrast underlines why markets are increasingly scrutinising whether AI demand translates into profitable workloads rather than simply rising chip shipments.
Funding is one constraint. O’Loughlin said hyperscale cloud companies had issued about $450 billion in debt during the year, a sum he described as surpassed only by borrowing associated with the United States and China. He added that pensions and annuity providers are among the major buyers of such debt, even as retirement savings increasingly move toward 401(k)-style structures rather than traditional pension pools.
Labor and grid construction present another barrier. O’Loughlin estimated that the United States faces a shortage of 100,000 electricians and said training can take about 18 months. He cited mid-level electrician compensation of around $250,000 a year, with pay reaching $400,000 to $500,000 for workers taking extended hours. His figures reflect the intense competition for skilled labour around data centres, power equipment and industrial construction.
Taiwan’s exposure keeps the stakes high
Taiwan remains at the centre of the AI supply chain. O’Loughlin said Taiwan Semiconductor Manufacturing Co., directly and indirectly, accounts for about 20% of Taiwan’s gross domestic product. He also said Taiwan’s GDP rose 25% during the year, attributing much of that growth to chip output.
That concentration gives the country an outsized role in the global race to expand AI computing capacity, while also amplifying the economic impact of swings in semiconductor demand. It places memory pricing, advanced packaging, power availability and capital spending into the same market narrative rather than treating them as separate issues.
The episode also touched on U.S. restrictions involving China’s access to advanced GPUs. O’Loughlin said the ROSA Act passed the House of Representatives by a 300–20 vote before stalling in the Senate, and argued that corporate lobbying had opposed measures that would restrict remote access to AI computing resources.
For cryptocurrency markets, the semiconductor pullback offers a reminder that AI-linked narratives can travel faster than the infrastructure needed to support them. Tokens associated with AI projects may respond to excitement around data centres and model launches, but the underlying hardware cycle is shaped by memory contracts, electricity connections, construction workers, debt markets and the eventual revenue earned from computing capacity.
Worried about leveraged volatility across semis and crypto? Learn risk-aware strategies in our guide to crypto crash navigation.
Disclaimer: The content on this page is provided for general informational purposes only and does not represent the views or financial advice of Toobit. We make no guarantees regarding the accuracy or completeness of this information and shall not be held liable for any errors, omissions, or outcomes resulting from its use. Investing in digital assets involves risk; users should independently evaluate their financial situation and the risks involved. For further details, please consult our Terms of Service and Risk Disclosure.
