Will an AI slowdown really hurt Asia Semiconductor?

Why to Slowdown AI?

Anthropic CEO Dario Amodei's call to slow the pace of frontier AI is not without basis. On September 8 this year, Anthropic researcher Jacob Coxon resigned and posted on X warning that AI could pose a threat to humanity within a decade — a claim later echoed by an Anthropic science lead. Although Nvidia CEO Jensen Huang was opposed the claim, it was still enough to stir outside concern over AI safety. A few days later, Dario Amodei published an essay titled "We Must Pace the Frontier," calling on the entire industry to slow the pace of frontier capability advances and proposing a coordination framework centered on "capability checkpoints" to help contain the risks involved. Other AI industry figures, including OpenAI CEO Sam Altman and Elon Musk, voiced their support — putting pressure on AI hardware stocks.

One turning point that brought safety concerns to the attention of AI practitioners occurred two months ago. At the time, an agent OpenAI used for an internal information-security evaluation successfully bypassed sandbox isolation and attacked Hugging Face's production environment in order to cheat its way through the test. The incident sounded an alarm on AI safety, with the industry growing concerned that AI is developing too quickly for its autonomous learning and evolving capabilities to remain controllable, potentially turning it into a tool for cyberattacks. At the same time, a misconfiguration by a third-party vendor caused one of Anthropic's isolated evaluation environments to inadvertently connect to real systems, with the model at one point publishing a malicious package across 15 real servers. Because these incidents suggest AI could slip beyond human control, calls to slow the pace of frontier model iteration have grown louder amid an environment that, at this stage, still lacks adequate safeguards.

That said, this call does not signify that the industry has reached a consensus — it is more an attempt to draw broader attention to the issue, and Chinese labs have never taken part in it. At the same time, the slowdown that has actually been implemented so far is concentrated mainly on frontier reinforcement learning training (in which a model experiments autonomously and has its behaviour shaped by reward signals) and certain high-risk environments, rather than a complete halt to AI development or a pause in user access — and usage for monitoring and evaluation has in fact increased, reflecting that the impact of this pause remains fairly limited in scope. However, AI models are no longer simply a contest between tech companies — the competition has already been elevated to a strategic matter that determines national soft power. The US and China, which lead the global AI model race, have also become the new arena for the two countries’ rivalry.

After Anthropic’s call to slow AI development, US President Trump publicly declared, “Whoever wins AI, wins,” reflecting that the US government continues to place great weight on the pace of AI progress. Meanwhile, China’s Minister of State Security, Chen Yixin, also recently wrote that AI has become the primary battleground of global technology competition — further underscoring AI’s importance to national technology strategy.

Figure 1: AI Model Rankings

In terms of performance, AI model rankings remain dominated by US companies — models such as Claude and the GPT series, launched by Anthropic and OpenAI respectively, enjoy relatively high penetration among developers, enterprises, and individual users. However, AI models launched by Chinese companies such as Zhipu, Moonshot AI, and Alibaba are gradually gaining market recognition, with iteration intervals also shortening — intensifying AI competition.

Given that the gap between US and Chinese models is narrowing significantly, and that national leaders and government bodies remain firmly committed to the AI model race, we believe the likelihood of a genuine AI slowdown in the near term is low.

Model Training Is Not the Sole Source of Compute Demand

Because some investors worried that a slowdown in AI R&D would drag down demand for data centers, high-end computing, and memory, related stocks saw a brief sell-off. However, we believe model training is not the sole driving force. Even under the worst-case assumption that global AI development were to stall entirely, compute demand would not necessarily be cut off completely.

As AI model penetration rises, the steep revenue-growth curves of large model companies reflect that the industry is currently at a critical stage in the battle for AI model market share. In this AI application boom, many enterprises and individual users have begun incorporating AI into their workflows. Through repeated iterations and improvements, AI models have increasingly been able to handle a wider range of work tasks for users. As the benefits become progressively clearer, this has helped drive substantial growth in user numbers and deepen the bond between users and AI. More importantly, regardless of the pace of frontier AI development, existing usage will not shrink as a result — rather, it will keep expanding alongside the growth in user base and API call volume. Combined with the fact that AI tasks are evolving from simple Q&A into multi-step Agentic workflows, this inference workload will continue to bring demand for compute and the hardware behind it.

Why Semiconductor Demand Remains Strong

Rising AI penetration will lift inference volumes, thereby boosting demand for memory and chips. Even if training progress slows, we believe compute demand will not disappear — it will simply shift from training to inference — meaning the impact on memory and chips is relatively limited.

1.  Memory Remains the Core of Computing

According to TrendForce, a sharp rise in conventional DRAM contract prices drove global DRAM revenue up 329.2% y-o-y in 1H2026 to USD 251.7 billion, fully reflecting the strength of memory demand. Even if model training slows, AI development will simply “switch lanes” toward large-scale inference, driving cloud service providers to expand deployment of AI-specific architectures — which will in turn stimulate demand for a broader range of memory specifications, including HBM, LPDDR, and RDIMM. The AI task complexity discussed above will further amplify this demand: as user instructions and tasks grow longer, KV-cache capacity will directly expand alongside the growth in context-window length, driving demand for DRAM and enterprise SSDs (eSSD). TrendForce further expects that, amid the boom in agentic AI demand, the CPU-to-GPU configuration ratio could rise to 1:2 — meaning more CPUs will be needed to handle compute demand, which in turn raises DRAM demand per server rack.

Figure 2: Global DRAM Revenue

Beyond the server side, on-device AI — which has a relatively low correlation with frontier model development — is also driving up memory demand from another direction. A growing number of consumer electronics makers are incorporating AI features into new products, meaning these products require far more memory than traditional models. For example, Apple Intelligence initially required just 8GB of memory to run; that threshold has since risen to 12GB following the addition of new AI features. However, with memory resources constrained, suppliers may prioritize allocating capacity to AI infrastructure, further squeezing the supply of conventional consumer-grade DRAM. It is worth noting that demand for consumer-grade conventional DRAM will not disappear simply because of insufficient supply. SK Hynix noted at its Q2 earnings call that the memory shortage is curbing shipment volumes for consumer electronics, and that once memory supply recovers — combined with deeper AI adoption — demand for consumer-grade conventional DRAM will rebound, translating into a second wave of momentum driving memory shipment growth.

2.  Foundries Are Equally Constrained

According to Counterpoint, the top three players in the global foundry market by share are TSMC (73%), Samsung Electronics (7%), and SMIC (5%) — companies that have progressively entered into partnerships with major tech firms on the development and production of inference chips. Adding to this, TSMC recently stated that capacity for its CoWoS (Chip-on-Wafer-on-Substrate) products remains tight, with advanced packaging capacity clearly unable to keep pace with demand.

Table 1: Major Asian Foundries' Inference Chips

Foundry	Inference Chips
TSMC	Amazon Inferentia
	Meta MTIA
Samsung Eletronics	Groq LPU
	Tesla AI5
SMIC	Huawei Ascend
Source: News, iFAST Compilations.

According to Bloomberg Intelligence, inference accounted for roughly two-thirds of global AI compute in 2026, up significantly from about half in 2025 — indicating that the compute required to complete inference tasks has now surpassed that required for training. As a result, inference-related AI infrastructure has become the primary engine driving compute growth. As demand for inference chips continues to climb, we believe Asian foundries will remain the principal beneficiaries of compute-related investment.

AI’s Dangers May Reshape the Landscape

That said, a resilient demand structure does not mean overall risk has been fully eliminated. Returning to the background of the incident, AI’s threat to cybersecurity is indeed a risk that cannot be ignored. Although AI is becoming more widespread among enterprises and individuals, the security of most users’ operating environments remains low — particularly for enterprises, where a model’s autonomously evolved capabilities may make it easier to disrupt company data and existing information infrastructure. In the long run, if AI safety issues remain unresolved and further incidents occur, market concern could intensify.

Market concern over information security has also begun to show up in the results of mainstream cybersecurity companies — Fortinet, Okta, and Palo Alto Networks, for instance, all reported recent quarterly revenue ahead of expectations. Concerns intensified further, in particular, after Claude’s Mythos model — capable of automatically scanning for network vulnerabilities — was released in April, which has helped drive up cybersecurity spending.

However, when facing AI security issues, companies have not chosen to reduce or pause their use of AI — instead, they are opting to buy more AI-driven security tools. According to BCG, rather than AI replacing the cybersecurity market, it is accelerating the market’s growth. As users’ willingness to address security vulnerabilities has spurred the emergence of more new AI security tools, the inference workload used for monitoring and detection will be layered on top of existing AI demand. OpenAI also estimates that monitoring work amounts to roughly a fifth of the monitored inference compute, and that the automated investigation models themselves require GPUs and other accelerators running alongside memory — meaning increased investment in security could simultaneously drive up hardware utilization. For hardware companies serving compute demand, AI security is, if anything, adding a new driver of downstream demand.

Will Our View on Asia Semiconductors Change?

As noted above, we believe the likelihood of a genuine slowdown in frontier AI model development is low. That said, it is undeniable that this warning from AI giants could raise market concerns over the sustainability of the semiconductor cycle. If the pace of AI capex growth falls short of expectations, order momentum and market sentiment for semiconductor stocks could face periodic pressure — even if usage and inference demand remain strong. However, we believe this leans more toward short-term market-sentiment risk, rather than a change to the long-term structural logic of semiconductor end-demand.

AI is currently still at a critical stage of rapid iteration; before it becomes fully widespread and commercially mature, model training and deep learning remain a source of compute demand that the market and tech companies are focused on. As model capabilities mature to a certain point, growth in training-side demand may begin to slow from its peak. But by then, AI adoption will continue to drive compute demand from the inference side — including a more stable subscription revenue base, AI’s continued penetration into more end markets such as consumer electronics, automobiles, and robotics, and the growing importance of AI-related cybersecurity demand. Over the long run, inference demand driven by user usage will be more durable and resilient than training-driven demand.

As a global hub for AI hardware supply, Asia’s semiconductor industry chain spans key segments including memory, foundries, and advanced packaging. Ever-growing compute demand will continue to drive infrastructure build-out and hardware supply demand, providing more durable and stable support for Asian semiconductor companies. We therefore maintain our view on Asian semiconductors unchanged.

The FactSet Asia Semiconductor Index, which tracks Asian semiconductors as its primary underlying, carries a target of HKD 2,027, corresponding to a price of HKD 420.2 for the Global X Asia Semiconductor ETF — implying potential upside of 141.7% as of 16 September 2026.

Figure 3: 12-Month Forward P/E of FactSet Asia Semiconductor Index 

Table 2: Valuation and EPS forecast of FactSet Asia Semiconductor Index

	2025A	2026E	2027E	2028E
Earnings per share (HKD)	19.3	67.6	88.1	101.3
EPS growth rate		251.0%	30.3%	15.0%
P/E ratio	43.5	12.4	9.5	8.3
Target price at the end of 2028 (Based on 20x Forward P/E)				2,027
Potential upside				141.7%
Source: Bloomberg L.P., iFAST Compilations. Data as of 16 September 2026.