Companies keep getting AI cheaper. Does that mean the AI cycle is over? The go… · @carrioresearch · TwKit

@carrioresearch

Companies keep getting AI cheaper. Does that mean the AI cycle is over? The good side 1. The savings are turning into more usage. In this survey, some teams at the fintech company halved their AI spend while using more tokens. Falling unit prices with rising volume is good news for whoever sells the volume, and memory is first in line: more inference means more KV cache, which means more memory. 2. Self-hosting open models still means buying GPUs and memory. The pharma group estimates self-hosting runs about 90% cheaper than calling an API. That moves compute demand from the cloud into the company's own data center. Total demand doesn't shrink. 3. Supply-side numbers are still accelerating. On Oct. 15 TSMC guided Q4 revenue to US$52.3–54.3bn, up 14% sequentially at the midpoint, and raised its 2026 growth outlook to about 44%. 4. Earnings estimates for optics and packaging/test are still moving up. What to watch 1. The software and model-API layer is under real pressure. 2. If unit prices fall faster than usage grows, where does the money come from? The return on hyperscaler capex ultimately depends on model-lab and cloud revenue. There's no first-hand data yet to settle it either way, and it's the biggest unknown. 3. Some of the capex growth is just higher prices. In April Microsoft said about $25bn of its roughly $190bn 2026 budget comes from higher component prices, and Meta tied its July raise to component pricing too. The dollars are growing faster than the hardware they buy. 4. TSMC softened its wording a notch. Its three-year capex language went from "even more significantly higher" to "significantly higher," and its caution on consumer demand now runs "going into 2027" because of rising component prices. Looking ahead The cycle is shifting gears, from building compute for training to scaling inference while cutting cost. How I rank who benefits: • Volume layer: memory, optical interconnect, power. First. • Compute chips: Nvidia. Second. • API and software layer: last. Thanks @FundaAI for the survey! $MU $SNDK $NVDA $TSM $MSFT $META $LITE $COHR $AAOI $TER $AMKR $BE

曝光 236 · 评论 1 · 点赞 2 · 书签 1