Investors are right to question the scale of artificial-intelligence spending, but abandoning the sector wholesale risks confusing financial scrutiny with technological failure.
The global retreat from semiconductor and artificial-intelligence stocks is less a rejection of AI than a belated demand for evidence. After years in which almost any announcement involving data centres, advanced chips or generative models was rewarded, investors are beginning to distinguish between companies that sell essential infrastructure, businesses that can convert AI into durable revenue and those merely attaching themselves to the investment theme.
That distinction is overdue. Nvidia (NVDA), Samsung Electronics (005930) and SK Hynix (000660) have been caught in a sharp technology sell-off that spread from South Korea across other Asian and US markets. South Korea’s Kospi suffered an unusually severe decline, while semiconductor shares weakened on concerns about elevated valuations, heavy borrowing for data-centre construction and faster progress by Chinese chip manufacturers. Nvidia also came under pressure as investors examined reports of its possible involvement in financing a vast Ohio data-centre project.
The instinct to reduce risk is understandable. AI infrastructure is becoming one of the largest corporate capital-allocation experiments in modern market history. Alphabet recently raised its projected 2026 capital spending range to between $190 billion and $205 billion, intensifying fears that free cash flow across the largest technology platforms could be consumed faster than new AI products generate revenue. Similar questions are hanging over Meta Platforms (META), Microsoft (MSFT) and Amazon.com (AMZN) as investors demand clearer links between infrastructure investment, customer adoption and eventual margins.
Yet the market’s current framing is too binary. AI does not need to justify every dollar of spending immediately to remain economically important. Nor does the technology’s long-term relevance guarantee that every company supplying it deserves a premium valuation. Both propositions can be true: AI may transform corporate productivity, while many AI-related investments still produce disappointing shareholder returns.
That tension has appeared in previous technology cycles. Railways changed commerce even though many railway investors lost money. The internet reshaped the global economy even though scores of listed internet companies disappeared. Mobile computing created enormous consumer value, but the gains were concentrated among a limited number of platforms, chip designers and software ecosystems. Technological adoption and investment performance are related, but they are not interchangeable.
The strongest case for maintaining selective exposure is that enterprise adoption is still in an early phase. Recent research examining S&P 500 companies found that 11% had deeply integrated AI into business processes by 2025, while another 10% were using it in producing goods or delivering services. Adoption had more than quadrupled from 2022, although deep implementation remained heavily concentrated in technology companies. The research also identified a profitability “J-curve”, suggesting that early implementation costs may precede measurable financial benefits.
That evidence argues against declaring the investment cycle finished. It also supports a more demanding valuation framework. Investors should care less about how often executives mention AI and more about whether adoption reduces labour intensity, improves pricing, raises customer retention or expands addressable markets. A company that spends billions without producing any of those outcomes should face a lower valuation, regardless of the strategic language used to defend the expenditure.
The same discipline should apply to infrastructure providers. Nvidia remains central to advanced AI computing, but market leadership cannot exempt it from capital-cycle risk. Customers are investing aggressively because they fear falling behind competitors, yet fear-driven spending can eventually create surplus capacity. More efficient models, custom chips and improved utilisation could reduce the number of premium processors required for each unit of computing output. Chinese progress in memory and chipmaking equipment could also increase competition, even if it does not immediately eliminate the technological lead held by established suppliers.
Investors should therefore avoid treating today’s market leaders as permanent monopolies. The semiconductor industry has always combined structural growth with violent cycles. Demand can remain strong while share prices decline because expectations had risen even faster. A stock does not need to become a bad company to become a poor investment at the wrong price.
The more interesting opportunity may lie beyond the obvious chipmakers. AI data centres require electricity generation, grid connections, cooling systems, networking equipment, storage and backup power. The current sell-off has already prompted some investors to re-examine companies addressing these physical constraints. Morgan Stanley has argued that power and energy-storage businesses may benefit from infrastructure bottlenecks even as broad AI-linked exchange-traded funds weaken.
That does not mean every utility, power producer or data-centre developer should be reclassified as an AI winner. Many of these businesses carry substantial debt, face regulatory uncertainty and require years of construction before producing cash. The same lesson applies across the chain: revenue visibility, balance-sheet strength and pricing power matter more than thematic association.
The Federal Reserve adds another layer of difficulty. With inflation still influenced by tariffs and energy costs, monetary policy may not provide the easy financial conditions that supported earlier technology booms. The Fed’s July monetary-policy report said inflation had moved higher after tariff increases pushed up some consumer-goods prices. Higher-for-longer borrowing costs would make distant AI profits less valuable and place greater pressure on companies funding infrastructure through debt.
The proper response is not to choose between blind enthusiasm and blanket pessimism. Investors should demand milestones. Cloud providers should show rising AI revenue relative to capital expenditure. Software companies should demonstrate that AI features support higher prices or lower servicing costs. Semiconductor suppliers should prove that demand is durable beyond a small group of hyperscale customers. Infrastructure developers should match long-lived assets with long-term contracts and sustainable financing.
The sell-off may ultimately mark the end of the easiest phase of the AI trade, when expanding valuation multiples mattered more than operating evidence. That would be healthy. Durable investment themes become stronger when weak assumptions are challenged, marginal projects lose funding and capital shifts towards companies with genuine economic advantages.
AI is now entering the part of the cycle in which execution matters more than imagination. Investors should welcome that transition. The technology does not need another round of unquestioning optimism. It needs financial discipline, and so do the portfolios built around it.
