AI Investment Sustainability Put to the Test…For Chip Stocks, Focus Shifts from Earnings to Shareholder Returns

Noh Geun-chang, CEO of Semicon Research Lab, viewed positively the improvement in revenue and profitability evident in Anthropic’s public disclosures. Revenue in the first half rose 14-fold year on year. The company remained loss-making under generally accepted accounting principles, but its adjusted operating margin was put at about 10%. The gross margin was about 85%, and Noh said profitability was improving as computing costs also declined.

Noh highlighted Anthropic’s position at the point where end consumers and corporate customers use frontier AI models. Semiconductor companies are further removed from end users, so investors may question the durability of their results even when earnings are strong. But he said evidence of revenue and profits at frontier-model companies helps validate AI demand. He added that the public filing made it possible to examine a debate that had until then relied on rumors and narratives against actual data.

Noh said growth continued after revenue slowed in August, supported by OpenAI’s Astra and other developments, and that Anthropic’s revenue was rising again in October. He also said combined monthly revenue at OpenAI and Anthropic continued to grow steadily. However, he said further examples were needed of companies that had adopted AI actually outpacing competitors and generating profits before investors’ doubts about monetization would be dispelled.

He assessed that business-to-business (B2B) demand was more likely than end consumers paying AI subscription fees to drive demand initially. Companies are still at the stage of experimenting with AI as a matter of course, and need time to change their workflows and organizations. Noh said that as more companies such as Palantir demonstrate actual gains in industry-specific productivity and competitiveness through AI, resistance to the costs of frontier models could ease and valuations across related companies, including semiconductor firms, could recover together.

Noh Geun-chang said supply constraints remain, making it premature to conclude that demand for AI semiconductors has weakened. Earnings announcements from Lam Research and ASML also indicated that equipment supply could be less smooth than expected, and bottlenecks in equipment are making it difficult to rapidly increase memory production capacity. Noh estimated that net annual additions of about 300,000 DRAM wafer starts would be needed to meet demand, and said this could exceed the combined expansion plans worldwide.

Noh said Samsung Electronics and SK hynix could differentiate themselves in premium products such as HBM and high-speed DRAM. Rising memory prices can increase revenue simply by raising the value of output, but expanding supply could lead to lower margins, making pricing and product mix important. He said HBM and high-speed DRAM command higher premiums than standard products and could lift overall profits if demand grows.

He said that because the market questions the durability of semiconductor companies’ earnings, companies need to explain their growth paths, supply constraints and responses to a slowdown in price increases in concrete terms during investor relations. They need to explain how they will maintain profitability if production does not rise as much as expected and what product strategies they will pursue even if memory prices fall over the longer term. Simply waiting for earnings to be reported before assessing the results and investing is not enough; he emphasized that IR should involve communicating forecasts and scenarios in advance.

Noh said shareholder returns can signal a company’s confidence and the durability of its profits. He explained that presenting specific return plans based on free cash flow and combining them with measures such as share cancellations could affect return on equity and a re-rating of the P/B ratio. He said Samsung Electronics and SK hynix were in the rare position of recording both low P/E and low P/B ratios, and that they should candidly explain to investors both the basis for their growth and unfavorable scenarios.

Noh Geun-chang said financing difficulties at individual neocloud companies such as CoreWeave do not automatically mean that the AI industry as a whole is in crisis. CoreWeave’s shares of revenue and U.S. data-center capacity were each put at around 3%. Hyperscalers are larger, and even if an operator runs into difficulty, another party could emerge to acquire its data-center and GPU assets.

Neocloud companies use borrowed money to buy GPUs and lease data-center computing capacity, leaving tangible assets in place even when financial pressures mount. Noh said CoreWeave’s and Nebius’s financing pressure could become an issue around 2027 and 2030, respectively, but cautioned against interpreting financial difficulties at a few companies as a collapse in overall demand. He added that the industry’s financing structures are not uniform, taking into account overseas operators with government shareholdings and other operators such as Naver.

Noh also suggested that the semiconductor cycle could prove more resilient than the market expects. Concerns about memory prices and supply volumes remain, but demand for HBM and high-speed products, as well as bottlenecks in equipment supply, could allow Korean companies to outperform. He urged shareholders of Samsung Electronics and SK hynix to be patient, saying their low valuations are difficult to match in past examples.