Oracle expands AI chip options with Google TPUs and startup hardware

Diversifying the AI hardware stack

Oracle has announced a significant expansion of its semiconductor supply for Oracle Cloud Infrastructure (OCI). The company plans to integrate chips from Google and various startup semiconductor firms into its existing infrastructure. This shift marks a move toward a multi-vendor approach to provide more variety for its customers.

While Oracle has historically been heavily focused on Nvidia hardware, this new strategy aims to strengthen its market position. By offering a broader range of AI processors, Oracle hopes to give corporate clients more options for their specific computing needs. This variety is particularly important as the demand for diverse AI capabilities grows across different industries.

The expansion was revealed during the Oracle AI World event in Las Vegas. The company intends to make these new chips available to users who are looking for alternatives to the standard industry offerings. This strategy allows Oracle to position itself as a flexible provider in a rapidly evolving hardware landscape.

Integrating Google TPUs

One of the most notable additions to the Oracle cloud will be Google's tensor processing unit (TPU). This is a custom-designed chip optimized specifically for machine learning and AI workloads. This partnership represents the first time a major data center operator other than Google will offer TPU services to the public.

Oracle will begin introducing these chips next year and plans to increase the supply over several years. Unlike some cloud arrangements where the provider simply rents space, Oracle will install the TPUs in its own facilities. The company will be responsible for the critical backend infrastructure, including power, cooling, and networking for the hardware.

This collaboration is an evolution of a pre-existing relationship between the two tech giants. The companies previously worked together on database services and launched a cloud partnership in June 2024. This new move extends their cooperation from data management into the realm of AI compute infrastructure.

Startup chips and inference

In addition to Google's hardware, Oracle is bringing in equipment from the startup semiconductor company Cerebras. These chips are scheduled to be deployed starting in the first half of next year. The company is also planning to offer hardware from SambaNova Systems to its cloud customers.

The inclusion of these startups highlights a shift in the competitive landscape for AI inference. Inference is the process of using a trained AI model to make predictions or generate content. While Nvidia currently holds a strong lead in the AI training market, competition is intensifying among custom chips and smaller startups for inference tasks.

By providing these specific options, Oracle is catering to users who may find certain hardware more efficient for specific tasks. This variety allows for a more nuanced selection process where customers can balance performance against the rising costs of chip expenses. It provides a path for companies to optimize their AI spending across different types of workloads.

Scaling Nvidia and infrastructure

Despite the move toward diversification, Oracle is not moving away from Nvidia entirely. The company plans to continue expanding its cooperation with Nvidia to support high-demand training needs. Specifically, Oracle intends to scale up its existing AI server, known as the Vera Rubin NVL72, to the multi-gigawatt level.

Scaling to the multi-gigawatt level indicates a massive commitment to physical infrastructure. This requires significant investment in data center capacity to house the high-powered servers. Oracle aims to provide the scale necessary for the largest AI models while simultaneously offering the smaller, specialized chips from its new partners.

The company did not disclose the specific contract amount for these hardware expansions. However, the scale of the project underscores Oracle's ambition to remain a primary destination for large-scale AI computing. They are balancing the need for massive, high-performance clusters with the need for diverse, flexible options for smaller applications.

Financial hurdles and market pressure

The massive investment required to build out this infrastructure presents a significant financial challenge for Oracle. The company has kept its fiscal 2027 capital spending plan between $90 billion and $95 billion to manage these costs. This reflects the heavy burden of funding the physical data centers needed to support modern AI demands.

The market has reacted to these high costs with some volatility. As Chosunbiz reported, Oracle's share price fell 47% from its year-to-date high on June 1. This decline was attributed to the heavy financial weight of the ongoing data center investments required to stay competitive in the AI space.

Oracle has noted that customers are facing rising expenses for AI chips, which may drive the demand for its multi-chip strategy. By providing more options, Oracle hopes to help customers manage these costs while still achieving their AI goals. The company must now navigate the balance of high capital expenditure and the need to provide a profitable, diverse service to its clients.