AI-linked stocks across Asia sold off sharply today. SoftBank, Kioxia, SK hynix, Samsung and TSMC all came under pressure as investors reacted to a growing debate around whether the industry should slow the pace of frontier AI development. Anthropic CEO Dario Amodei has called for more time to evaluate safety risks before pushing model capabilities much further, while other major AI leaders have also shown support for stronger safeguards.
The market’s first reaction is understandable: if even the AI labs themselves are saying “slow down,” does that mean the massive spending on GPUs, HBM, networking and data centers is also about to cool?
Tiger thinks the answer may be more complicated.
What may slow is the pace of frontier model training, not necessarily the overall demand for AI compute. If the next generation of models takes longer to arrive, cloud providers and enterprises may actually get more time to deploy the models that are already powerful enough for coding, search, advertising, customer service and AI agents. In that scenario, AI spending does not disappear. It simply shifts from building the next giant model toward running existing models at scale.
That is also why the recent Meta Agent story matters. The next phase of AI may depend less on whether the newest model beats the previous one by a few benchmark points, and more on whether billions of users and millions of companies start calling these models every day. If Meta pushes agents deeper into WhatsApp and Instagram, Microsoft expands AI across enterprise workflows, Google embeds Gemini further into Search and Cloud, and Amazon keeps scaling AI through AWS, the center of compute demand could gradually move from one-time training runs toward constant inference.
And inference is not “lightweight” from an infrastructure perspective. Agents still need GPUs or ASICs to execute tasks, HBM and DRAM to feed data, optical and networking equipment to connect clusters, and huge amounts of electricity to keep data centers running. So one distinction matters a lot here:
Slower frontier training does not automatically mean lower total AI compute demand.
In fact, some parts of the hardware stack could benefit from this shift. Frontier training is concentrated among a small group of labs and a narrow set of top-end accelerators. Inference is potentially much broader. Enterprises care more about cost, latency, power efficiency and cost per token, which could create more room for custom ASICs, inference-focused chips and alternative suppliers.
That said, today’s selloff also highlights a real risk: AI stocks are already priced for very strong growth. The market is not only debating safety regulation. It is also asking whether today’s valuations still make sense if model progress slows. If AI improves more gradually than expected, investors may start questioning whether current levels of capex can still generate enough return.
Tiger View
Tiger would separate the AI cycle into two stages.
The first stage was building models. The market rewarded bigger training clusters, more advanced GPUs and stronger benchmark performance.
The second stage may increasingly become using models. The market will start caring more about agents, enterprise adoption, inference volume and whether AI can generate real revenue.
So today’s correction does not automatically mean AI CapEx has peaked. The more important question is whether spending begins to shift from:
Frontier Training → Inference & Deployment
If flagship model releases slow down, but AI usage, cloud revenue and data-center utilization keep rising, then “AI slowing down” could actually mean the industry is entering a more commercial phase.
The real warning signal would be very different: cloud companies cutting capex, GPU utilization weakening, and HBM or networking orders starting to fall at the same time.
Until then, today looks more like a repricing of the assumption that AI capabilities will keep accelerating forever — not confirmation that AI demand has peaked.
Related Stocks
AI Compute: $NVIDIA(NVDA)$
Watch whether inference demand can continue to support GPU growth even if frontier training slows.
Memory: $Micron Technology(MU)$, $SK hynix(SKHY)$, Kioxia
Watch HBM, DRAM and data-center storage demand as inference workloads scale.
AI Networking / ASICs: $Broadcom(AVGO)$, $Arista Networks(ANET)$
Watch custom silicon and networking demand as AI deployment broadens.
Cloud Platforms: $Microsoft(MSFT)$, $Alphabet(GOOG)$, $Amazon.com(AMZN)$ $Meta Platforms, Inc.(META)$
Watch whether AI spending converts into cloud revenue, agent usage and monetization.
Today’s Poll
If frontier AI development really slows, where do you think capital goes next?
① Inference / AI Agents
② GPUs / HBM remain the core trade
③ ASICs / Networking benefit most
④ AI CapEx peaks and the whole chain cools
For market discussion only. This is not investment advice. Markets involve risk, and investment decisions should be made carefully.
Comments
Selling off Asian chip stocks just because AI leaders said "slow down" is a knee-jerk overreaction. Slowing down frontier training doesn't mean stopping construction. Foundation models are already big enough. The money is shifting from chasing model benchmarks to monetizing inference and agents.
High-frequency inference consumes HBM and DRAM even more persistently. The memory logic is intact. ASICs and networking will benefit as demand diversifies.
Today's broad sell-off is a textbook headline-driven panic. Don't hand over your core chips. Let Micron's Sept 30 earnings show the real demand numbers. I'm watching for dips, not joining the bears.
The biggest mistake is treating slower frontier-model development as the same thing as weaker AI demand. These are two different cycles.
If model training slows, the next growth engine could be inference: AI agents, enterprise automation, search, coding and customer service. Once millions of businesses start using AI continuously, compute demand becomes recurring rather than concentrated in giant training runs.
That shift could also benefit ASICs, networking, HBM and data-center infrastructure, especially where cost, latency and power efficiency matter.
So I wouldn’t call today’s selloff the end of the AI cycle. I’d call it a rotation from “build smarter models” to “use AI everywhere.”
The real warning sign would be falling cloud capex, weakening GPU utilization and declining HBM orders at the same time.
@Tiger_comments [贱笑]