Citrini’s View: Nvidia’s Moat Remains Deep; HBM4 Costs Doubling Drives Up Rubin’s Pricing, With Its High Gross Margins and Strong Pricing Power Unshaken.

Citrini analyst Jukan cited Fubon Securities' *2027 Semiconductor Outlook* report, noting that although HBM4 costs will jump sharply from HBM3e's $17–18 per GB to $31–32 per GB in 2026, pushing Nvidia's Rubin GPU price to roughly $78,000–$80,000, Nvidia will still maintain a high gross margin of 75%–80%, with its pricing power and cost pass-through capability remaining unshaken. The report also pointed out that custom ASIC-based HBM could have even higher costs, reaching $35–36 per GB, meaning HBM costs will double by 2027. Fubon Securities remains optimistic about AI market demand, arguing that despite recent concerns over AI inflation, token costs are a more critical factor driving cloud service providers' capital expenditures. At the architecture level, Nvidia's new rack is still planned to use the same number of compute chips, with scale-up within racks via cables and scale-out between racks via NPO/CPO. Additionally, Google plans to deploy 12 million to 15 million TPUs by 2028, with production capacity more than doubling from 2027 levels; Intel's EMIB production capacity is expected to rise to 24,000–25,000 units per month by the end of 2027, while TSMC is slowing SoIC expansion to prioritize scaling up CoWoS capacity. The core conclusion is that even with changes to its cost structure, Nvidia's pricing power, gross margin, and demand support in the AI computing ecosystem still form an unassailable competitive moat.