What TradingKey 美股財經 said in this video:
- AVGOneutrallong-termBroadcom's AI semiconductor revenue is growing rapidly, but mainly captures newly added demand from the overall expansion of the AI computing market rather than taking existing share from NVIDIA. Broadcom and NVIDIA revenue cannot be directly compared because Broadcom sells chips while NVIDIA sells whole racks.Source
- Otherneutrallong-termOpenAI's first custom chip Jalapeno is inference-only and does not do training; its announcement also states that training and inference will continue to heavily use NVIDIA and other accelerators, confirming NVIDIA's training moat has not yet been touched by custom chips.Source
- AI Capexneutrallong-termThe incremental return on investment for cloud vendors' AI capital expenditure is declining. NVIDIA's capital wall of guarantees and investments to support customers forms a new moat but moves industry cyclical risk onto its own balance sheet. Ultimately, one must watch whether more external customers pay with their own cash for AI services; if total AI investment declines, even a strong moat will not prevent valuation contraction.Source
- NVDAneutrallong-termNVIDIA's competitive advantage definitely exists; its long-standing software ecosystem and acquired network interconnect system integration form a moat. As long as AI computing demand keeps expanding, revenue will likely keep growing, but what is truly being eroded is pricing power with large customers and profit quality, and the stock price may already reflect optimistic expectations.Source
- AI Capexneutrallong-termCustom-developed ASIC chips are accelerating encroachment into standardized, highly repetitive inference tasks, but frontier training and complex system integration remain hard to shake in the short term. There is a migration gap between inference and training, so custom chips mainly capture new market growth rather than replacing NVIDIA's core training position.Source