正文 · AI 翻译
Thanks to IT Home reader 乌蝇哥的左手 for the tip!
IT Home reported on October 28 that Oracle said on Tuesday local time it would expand its cloud AI chip partnership with Nvidia while further diversifying its supply by bringing in products from Nvidia rivals, including chips from Alphabet-owned Google and Cerebras Systems.
Oracle and Nvidia said Oracle would deploy Nvidia’s latest-generation Vera Rubin AI servers at large scale, with the deployment reaching “several gigawatts.” The two companies did not disclose the deal’s value. Nvidia CEO Jensen Huang has previously said that each gigawatt of Vera Rubin system deployment could generate up to $40 billion in revenue for Nvidia (IT Home note: approximately 267.35 billion yuan at current exchange rates).
Oracle is also working to diversify its supply. The company said that even in systems based on Nvidia chips, it would stop using some of Nvidia’s networking chips and instead use Oracle’s own Acceleron networking chips.
Oracle also announced that it would offer Google’s Tensor Processing Unit (TPU) chips to customers for the first time. Oracle will be the first hyperscale cloud provider outside Google to offer Google TPUs. The companies said the service is expected to launch in 2027 and scale up gradually over several years to support large-model training, inference and agent applications. Google’s eighth-generation TPUs include the TPU 8t for training and the TPU 8i for inference.
Mahesh Thiagarajan, president of Oracle Cloud Infrastructure, said in a statement: “Different AI workloads have different computing requirements.” He said offering more choices would allow customers to select the right infrastructure flexibly according to the specific performance, throughput and cost-effectiveness requirements of various AI workloads.
In addition to Google’s chips, Oracle also plans to offer chips from Intel-backed SambaNova Systems and Cerebras Systems. Both companies focus on AI inference chips, used in the computing stage when an AI system generates a response after receiving a request from a human user.
Nvidia currently dominates the AI model training market, but competition in the AI inference market is more varied, with more participants.
Discussing the partnership with Oracle in an interview with Reuters, Cerebras CEO Andrew Feldman said: “This will greatly expand our tier-one customer base.” He added: “Our products are entering other companies’ data centers through large-scale deployments.”
Oracle plans to hold an investor day this week, but has not disclosed how it will finance these chip purchases.
On last month’s earnings call, Oracle said customers were increasingly covering chip procurement costs. The company expects related spending to reach approximately $90 billion to $95 billion in fiscal 2027 (approximately 601.538 billion to 634.956 billion yuan at current exchange rates).