英伟达在 CES 2026 提前披露下一代 Rubin：已进入「全面量产」，预计于 2026 年下半年大量出货并供货；首批云端客户包含 Microsoft Azure 与 CoreWeave。公司也说明其通常在 GTC 发表晶片更新，而今年 GTC 订于 3 月 16–19 日举行。

Rubin（搭配 Vera CPU）相较 Blackwell 的量化提升包括：推理算力约 5 倍、训练算力约 3.5 倍；并可把推理 token 成本最多降到 Blackwell 的 1/10（降低至多 10 倍），以压低训练与推理总成本。Rubin 也将首度整合 HBM4，资料传输速率最高可达 22 TB/s。

市场反应偏冷：英伟达股价周一盘后下跌 0.13%，此前收在 188.12 美元。黄仁勋在约 3,000 名听众前强调「实体 AI／机器人」的拐点已到，并提及公司预期到 2026 年（即便不含中国与部分亚洲市场）仍可凭 Blackwell 与 Rubin 早期放量累计进帐约 5,000 亿美元。

Nvidia used CES 2026 to detail its next-gen Rubin AI chips, saying Rubin is already in “full production” and will ship in volume in the second half of 2026; early cloud availability is expected via Microsoft Azure and CoreWeave. The company noted it typically updates its AI roadmap at GTC, scheduled this year for March 16–19.

Key quantified claims versus Blackwell: Rubin targets ~5× inference compute and ~3.5× training compute, and aims to cut inference token cost by up to 10× (down to one-tenth), lowering overall training and inference costs. Rubin also introduces integrated HBM4 with throughput up to 22 TB/s.

Despite the specs, the immediate market move was muted: the stock slipped 0.13% in extended trading after closing at $188.12. Huang emphasized a pivot toward “physical AI,” speaking to about 3,000 attendees, and reiterated expectations of roughly $0.5T in revenue through 2026 from advanced Blackwell sales plus early Rubin ramps, even excluding China and other Asian markets.