GPUs should remain critical for training and rapidly evolving workloads, where flexibility and software ecosystems matter. But inference is different: once workloads become predictable and massive, every watt and every dollar per token matters.
That’s why custom silicon is becoming strategically important. OpenAI’s Jalapeño, developed with Broadcom, is a good example of this shift toward workload-specific optimization.
What interests me most is the “picks-and-shovels” layer. Broadcom isn’t simply competing with NVIDIA; it can benefit when hyperscalers build their own accelerators because those chips still need advanced connectivity, networking and silicon expertise. Broadcom’s Q3 FY2026 AI semiconductor revenue reached $16.7B, up 221% YoY.
My takeaway: the future may not be GPU or ASIC. It may be GPU + ASIC + networking — with each optimized for a different part of the AI workload.
@Capital_Insights [贱笑]
Disclaimer: Investing carries risk. This is not financial advice. The above content should not be regarded as an offer, recommendation, or solicitation on acquiring or disposing of any financial products, any associated discussions, comments, or posts by author or other users should not be considered as such either. It is solely for general information purpose only, which does not consider your own investment objectives, financial situations or needs. TTM assumes no responsibility or warranty for the accuracy and completeness of the information, investors should do their own research and may seek professional advice before investing.
