Oracle Cloud adds Cerebras and Google TPU, expands NVIDIA partnership to multi-gigawatt scale

Oracle announced that it will bring Cerebras Systems' AI chips into its cloud services, a move that marks another heavyweight customer for the company seen as Nvidia's most formidable emerging rival, as competition in the AI inference chip market accelerates.

Oracle co-CEO Clay Magouyrk announced the partnership on Tuesday at the company's annual customer conference, with Cerebras systems expected to be officially deployed on Oracle Cloud Infrastructure in the first half of 2027.

English text is being prepared.

After the news was announced, Cerebras shares rose more than 10% at one point during intraday trading. However, since the company's listing in May this year, its shares had still fallen a cumulative 41% as of Monday's close.

Oracle aggressively expands its chip footprint

English text is being prepared.

On the Nvidia partnership, Oracle said it will expand deployment to "multi-gigawatt scale," using Nvidia's latest-generation "Vera Rubin" AI servers.

Nvidia CEO Jensen Huang previously said the Vera Rubin system could generate up to $40 billion in revenue per gigawatt for Nvidia, but neither side disclosed the specific value of this agreement.

Notably, even in systems deploying Nvidia hardware, Oracle chose to replace some of Nvidia's networking products with its own "Acceleron" networking chips, further strengthening its self-reliance at the infrastructure level.

In addition, Oracle announced the first introduction of Google's TPU (Tensor Processing Unit) chips.

"Different AI workloads have different computing needs, and diversified choices give customers flexibility to select the infrastructure best suited to each workload based on performance, throughput and economics," said Mahesh Thiagarajan, president of Oracle Cloud Infrastructure.

Cerebras: Scale expansion still fails to dispel investor doubts

English text is being prepared.

"This is a major expansion of our tier-one customer base," said Cerebras CEO Andrew Feldman. "We are being deployed in other companies' data centers at large scale."

The company had previously signed well-known customers including OpenAI and Amazon Web Services. According to the average analyst forecast compiled by Bloomberg, Cerebras' revenue is expected to reach about $3 billion next year, roughly tripling from this year, while full-year revenue this year is still expected to be less than $1 billion.

Even so, investor doubts have not dissipated. Nvidia's current quarterly AI chip sales are close to $100 billion, giving it an overwhelming advantage over Cerebras in market scale.

With Nvidia dominating the training chip market, both Cerebras and SambaNova have targeted the more fragmented inference market as their breakthrough point, but whether they can gain a firm foothold there remains a core question for investors.

Capex plans and financing arrangements remain to be clarified

On the capital expenditure front, Oracle has not yet disclosed the scale of Cerebras products it plans to purchase or the specific financing method for chip purchases.

On last month's earnings call, Oracle said customers are increasingly bearing chip procurement costs themselves, and it maintained its capital expenditure target of about $90 billion to $95 billion for fiscal 2027.

Oracle previously said it would invest $70 billion in capital expenditure in the fiscal year ending May 2027, aiming to join the ranks of top AI computing power suppliers.

Oracle also plans to hold an investor day this week, where it may disclose more details on its data center expansion strategy and capital allocation arrangements. Risk disclosure and disclaimer

Markets carry risks, and investment requires caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation or needs of individual users. Users should consider whether any opinions, views or conclusions in this article are suitable for their particular circumstances. Investment based on this is at the user's own risk.