AI\u2019s trade is shifting from scarcity towards economics as model and inference costs fall, broadening adoption through open-source systems, cloud platforms and APIs while weakening pricing power for some developers and hardware suppliers. Chinese competition is pushing price and efficiency, with Alibaba launching its largest model and DeepSeek offering ultra-low-cost pricing, encouraging use cases where \u201cgood enough\u201d beats frontier performance. Hyperscalers are still building capacity, yet procurement is becoming more selective as buyers focus on cost per task; consensus capex for Microsoft, Alphabet, Amazon, Meta and Oracle rose from about $485bn at the start of 2026 to roughly $730bn by July, while UBS sees growth of 76% in 2026, easing to 25% in 2027 and 6% in 2028. Demand may pivot towards customised accelerators, inference optimisation, networking, power and cooling, rather than maximum GPU intensity.<\/p>\n\n
For Korea, cleaner positioning may support a tactical rebound, but it does not guarantee a fresh earnings cycle. J.P. Morgan estimates the leveraged ETF unwind is complete and hedge funds are around 90% through deleveraging, yet the risk is that memory expectations peak before demand does. TrendForce still expects DRAM undersupply, with demand growth potentially outpacing supply into 2027, though HBM and advanced server DRAM may stay tight while commodity DRAM and NAND remain more cyclical. The broader message is that returns will accrue to distribution, proprietary data and platforms alongside efficient infrastructure, not merely to the bottlenecks of the first phase.<\/p>\n\n
We are observing a critical transition in the options market as the AI trade moves from raw infrastructure scarcity to cost efficiency. Derivative traders should prepare for a decline in the implied volatility of major hardware suppliers as the desperate scramble for chips cools down. We suggest shifting focus toward options strategies that exploit this transition, particularly as hyperscaler capital expenditure growth begins to decelerate.<\/p>\n\n
Recent industry data shows that while combined capital spending from tech giants like Microsoft, Alphabet, and Meta is projected to reach nearly $730 billion this year, the rate of growth is expected to slow to 25% next year. This deceleration means the high premium on hardware call options is becoming much harder to justify. We recommend targeting bull call spreads on mega-cap software and digital platforms that are successfully monetizing cheaper AI models.<\/p>\n\n
Looking at South Korea, we see signs that peak memory expectations are already priced into the KOSPI, even though high-bandwidth memory remains tight. Historical cycles show that memory equity valuations frequently peak several quarters before actual corporate revenues do. To manage this risk, we favor using put options or bear put spreads on major memory manufacturers to hedge against a valuation contraction.<\/p>\n\n
As the industry shifts focus to cheaper inference over massive training models, custom silicon and specialized infrastructure are gaining traction. This shift is reflected in the rising options volume for mid-cap networking and power management firms. We can capture this trend by selling puts on select infrastructure providers that benefit from this efficiency drive, allowing us to collect premium on stable cash flows.<\/p>\n\n
Implied volatility skew for major semiconductor ETFs has recently started favoring puts over calls, indicating that institutional players are actively hedging their hardware exposure. This suggests that the easy money in pure hardware momentum has run its course for now. We believe trading short-term calendar spreads on these hardware names will allow us to capitalize on high premiums without committing to directional bets.<\/p>\n\n
The aggressive pricing of open-source models by Chinese tech firms is accelerating global cost deflation and squeezing margins for model developers. This price war makes long-term call options on pure-play AI developers highly risky and speculative. Instead, we prefer buying long-dated call options on established enterprise software platforms that can easily integrate these cheaper models to boost their own profit margins.<\/p>\n\n\n\n