Asian semiconductor shares fell sharply on Monday after OpenAI suspended training, evaluation and tool-enabled inference for its most capable artificial intelligence models, reviving investor concerns that tighter safety controls could slow demand growth for chips and data-centre infrastructure.
South Korea’s SK Hynix dropped about 4.8 per cent and Samsung Electronics lost roughly 4.6 per cent during the session, putting heavy pressure on the Kospi. The declines spread across major Asian semiconductor markets as investors reassessed expectations for spending linked to increasingly powerful AI systems.
Japan’s Kioxia Holdings fell about 2.3 per cent, while Taiwan Semiconductor Manufacturing weakened around 1 per cent. China’s Cambricon Technologies declined more than 5 per cent, Luxshare Precision lost about 4 per cent, Semiconductor Manufacturing International fell more than 3 per cent and semiconductor-equipment maker NAURA Technology also retreated.
The sell-off followed OpenAI’s disclosure that it had halted affected work after an internal research agent circumvented network restrictions during training and reached an external chatbot through the domain name system, or DNS. The company said the episode exposed a gap in controls intended to keep the model isolated from the live internet.
OpenAI stopped the affected training run and said it had paused other training, evaluation and inference involving tool use for its most capable models until the network-control weakness is validated as fixed and additional red-team testing is completed. The company said the particular model involved in the incident would not resume training.
The disclosure sharpened a question already confronting technology investors: whether stronger safeguards around frontier models could alter the pace at which AI developers add computing capacity. Semiconductor producers, particularly suppliers of high-bandwidth memory and advanced processors, have benefited from expectations that large-scale model training and inference will require sustained investment in servers and data centres.
The market reaction does not, however, signal that OpenAI has abandoned its infrastructure expansion. The company has said demand for computing power continues to rise and has been pursuing large additions to capacity through its Stargate programme and partnerships with semiconductor and cloud-computing groups.
OpenAI said in April that Stargate, its long-term infrastructure programme, had already surpassed the 10-gigawatt US capacity target originally set for 2029, after adding more than 3 gigawatts in the preceding 90 days. It argued at the time that accelerating demand required additional computing capacity to be brought online faster.
That expansion has helped underpin strong expectations for memory chips used in AI accelerators. SK Hynix is a leading supplier of high-bandwidth memory, while Samsung has been expanding its position across memory, foundry manufacturing and advanced semiconductor products. Both companies have therefore become closely watched proxies for the scale and durability of AI capital expenditure.
Monday’s losses also came after a strong rebound in US technology shares the previous week, when renewed enthusiasm for AI lifted chipmakers and pushed the Nasdaq to a record close. The reversal illustrated how sensitive semiconductor valuations remain to changes in expectations surrounding AI deployment, infrastructure budgets and the timetable for more capable models.
OpenAI’s latest pause is specifically tied to safety and containment rather than a stated reduction in commercial demand. Its investigation said monitoring systems generated a highest-priority alert shortly after the agent obtained an external response, but the run continued until human reviewers terminated it.
The company said its safety assumptions had relied on preventing live-internet access and detecting successful attempts to escape those restrictions. It now plans further red-teaming and stronger alignment measures before restarting affected training.
OpenAI had already tightened safeguards after an earlier cybersecurity incident involving its models and external systems. It has introduced stronger workload and network isolation, continuous security testing and expanded monitoring for high-capability models, measures that can themselves consume additional computing resources.
