Reality check for artificial intelligence: correction or end of cycle?

Note: For explanations of AI terminology, please see glossary at the bottom of the Viewpoint.

Putting the sell-off into perspective

Asian technology supply chains sit at the heart of the AI investment boom led by US tech giants. They have therefore witnessed exceptional earnings growth and stellar share price performances. Since the launch of ChatGPT in 2022, the MSCI Emerging Markets Information Technology Index has increased more than fivefold, and along the way, fears around an AI bubble and concerns over a lack of return on investment (ROI) repeatedly triggered corrections. So far, these pullbacks have been followed by periods of strong market recovery.

The latest sell-off seen in June 2026 followed a familiar pattern. Concerns about AI ROI, end-of-cycle fears, rising Chinese competition, and economic uncertainty surrounding renewed US-Iran tensions weighed heavily on sentiment. Elevated levels of investment leverage, given many private investors used borrowed capital, amplified the decline. In our view, however, the market reaction overstates the risks, as large-scale investment plans, moderating memory price hikes, and lower-cost Chinese AI models do not imply an oversupply environment. Instead, we view the recent correction as another opportunity to gain exposure to a compelling long-term growth theme, even if near-term weakness might persist.

The big picture: how we track the AI cycle

From a historical perspective, the AI investment cycle seems to be at a relatively early stage. Previous technology waves, such as smartphones and public cloud computing, lasted more than eight years before revenue expansion slowed down to below 20%. Less than four years have passed since ChatGPT's launch in November 2022, and we believe AI could ultimately surpass both these technology revolutions. The impact is already evident: Cloud revenue growth reaccelerated, towards 50%, while AI leaders such as OpenAI and Anthropic reportedly quadrupled their revenues in 2025. Yet enterprise adoption remains at an early stage, and agentic AI workflows are only beginning to scale, which may indicate a long growth runway.

The unprecedented surge in demand for compute capacity also led the overall semiconductor market to triple, fueling a broad ecosystem of datacenter technology providers.

A long-term perspective on the semiconductor industry highlights the recurring pattern of boom-and-bust cycles with periods of undersupply eventually followed by overcapacity, even within powerful secular growth trends. The cloud computing boom, for example, accelerated after Amazon disclosed AWS in 2015, yet still experienced significant corrections during the 2018 inventory adjustment and the slowdown related to the Covid pandemic in 2022. This underscores the importance of closely monitoring the supply-demand balance. Given US hyperscalers account for the bulk of global datacenter investments and their capital expenditure (capex) should total roughly USD 745 billion this year, we believe the key questions are whether this level of spending creates value and whether the strong outlook is supported by satisfactory return expectations.

Chart 4: AI Tracker – no bubble, but financing and datacenter pipeline to watch

Sources: Bernstein, Bloomberg, BofA, Morgan Stanley, Goldman Sachs, OpenRouter, Ramp, Silicon Data, UBS, Vontobel; data as of August 5, 2026

Infrastructure investments continue, albeit increasingly via debt and equity financing

We expect hyperscalers’ capex to grow by more than 80% this year, and recent developments may support a further upward revision of the outlook. Alphabet recently increased its capex guidance by another USD 15 billion, although this exceeds its operating cash flow for the first time in history, and even indicated "significant" further expansion next year. Amazon is also accelerating its investments, expecting “compelling returns” after a three-year payback period and a "striking" demand expected for 2028.

The reason is simple: Compute capacity appears to remain well short of rapidly expanding demand. If demand continues to exceed supply, higher investments could remain both economically attractive and strategically important to preserve competitiveness. While increasing reliance on debt financing requires close monitoring, we view it as a yellow rather than a red flag. In fact, we consider both debt and equity issuance supportive of further growth in the near term, and we expect hyperscalers to increase debt as long as their investment-grade credit ratings, usually a sign of high creditworthiness, remain intact. The initial public offerings (IPO) of Anthropic and OpenAI provide additional funding for capacity expansion. Another yellow flag is the rise of so-called circular financing, where for instance, hyperscalers take stakes in major customers. While this increases leverage both up- and downwards the ecosystem, it also supports investment growth in the near term.

Potential path to higher profitability becomes clearer

The durability of the AI cycle ultimately depends on profitability and returns on invested capital (ROIC). Hyperscaler cloud businesses on aggregate have accelerated to almost 50% growth with operating margins of 35-40%. Newer chip generations, in-house silicon such as Google's TPU, and selling access to in-house models running on that capacity could support margins as the AI mix grows. Hyperscalers earn strong returns on new capacity, but not yet on the whole capital base. A well-utilized, current-generation datacenter could earn around 20-30% ROIC when rented to customers, but much of the total capital is still under construction, running older chips, or absorbed by internal training. Therefore, we think returns are initially just around the cost of capital, or slightly below, and we expect they can rise over time, as capital utilization increases. Rental prices remain supportive.

Frontier AI labs also exceeded expectations in terms of both user and revenue growth. OpenAI just surpassed one billion monthly users after less than four years – Meta took nine years to achieve this – while Anthropic's last reported annualized revenue run rate of USD 47 billion surpassed software companies like SAP or Salesforce, which took decades to achieve a similar scale. Model makers are generally still loss-making, even if some estimate that Anthropic is already turning towards positive operating profit thanks to its enterprise focus and a less aggressive capacity expansion compared to OpenAI. In any case, we see a possible path to higher profitability over the longer term. Expensive training and free users initially weigh heavily on profitability. But gross margins on inference – which is the phase when a trained model applies its knowledge and generates revenue – are reportedly around 70% in the enterprise business. So, as inference grows relative to training, margins and returns on investment could improve. And although OpenAI currently seems to face more challenges with a customer base skewed to consumers, we see that potential improvement could come from higher enterprise adoption, advertising opportunities, and deeper vertical integration.

Competition from much cheaper Chinese AI models remains a key concern, but DeepSeek demonstrated that lower token prices can boost token volume consumption. For frontier models, usage growth outpaced price declines by far, which explains the exponential growth in revenues of frontier labs. Moreover, while the recent launch of Chinese model Kimi 3 from Moonshot AI surprised the market with its strong performance, it also comes with higher pricing, comparable to that of US frontier models, as it requires high compute capacity. This underscores that, contrary to common perception, leading Chinese open-weight models are not for free and not that cheap after all. It also demonstrates the so-called scaling laws that model performance predictably improves with more compute, training data, and model size.

Supply shortage may take longer to resolve

Technology cycles typically turn when supply catches up with demand and concerns about oversupply emerge. Alongside the massive spending plans of hyperscalers, investments in the multi-billions recently announced by chip makers stoke such fears. TSMC raised its capital expenditure outlook and announced an additional USD 100 billion investment in advanced chip manufacturing facilities in the US. Meanwhile, Korean memory producers Samsung Electronics and SK Hynix unveiled long-term expansion plans worth over USD 1.7 trillion and USD 750 billion, respectively.

However, much of these investments will be made beyond 2030. With semiconductor projects taking years to complete and ramp, supply may remain tight in the meantime. As a result, memory prices rose roughly tenfold since early 2025, and customers are increasingly willing to commit to long-term contracts at elevated prices to lock in future supply. Supply constraints appear to remain widespread across the AI supply chain, extending beyond semiconductors to critical components and input materials. We therefore closely monitor pricing and capacity. Further downstream, datacenter deployment continues to face bottlenecks related to power supply, building permits, and capacity delays, slowing timely deployment. Hence, this is another key area for us to track progress.

Current valuations attractive and far from bubble territory

Taken together, we believe that the AI cycle should still have further room to run. While valuations were elevated before the recent price correction in technology stocks, they remained well below the levels seen in the dot-com era from 1995 to 2001, when the rapid rise of the Internet caused a massive stock market boom driven by internet-based startup companies, which ended in a dramatic crash. Unlike some previous speculative episodes, today's technology rally has been supported by strong earnings growth and sustained infrastructure demand. After the sell-off, we believe that current valuations look attractive and that the long-term drivers of the AI cycle remain broadly intact.