OpenAI Is Building Its Own AI Chips, So Why Is It Turning To Google?

The race to build artificial intelligence (AI) models is increasingly becoming a race to secure the computing power needed to train and run them. As AI services attract more users and handle increasingly complex tasks, companies need access to vast amounts of computing capacity. Securing that capacity has become as important as developing the technology itself.

Against this backdrop, OpenAI’s reported move to source AI chips from Google raises questions about its dependence on rival technology providers and the growing competition to secure computing power. The two companies compete in AI products, with OpenAI’s ChatGPT going up against Google’s Gemini.

However, according to The Information, OpenAI is planning to purchase Google’s Tensor Processing Units (TPUs), a move that could deepen their commercial relationship. The report, cited by Reuters, did not disclose the financial terms or other details. Neither company provided further information.

The development may appear contradictory because OpenAI is also working with semiconductor company Broadcom to develop its own AI accelerators. But designing custom chips and securing enough computing capacity to meet immediate demand are two different challenges. For OpenAI, using external suppliers while developing its own processors could be a way to address both.

OpenAI’s Google Move

OpenAI already relies heavily on Nvidia’s graphics processing units (GPUs) to train and run its AI models. GPUs are widely used for AI computing because they can handle large numbers of calculations simultaneously. Google’s TPUs, meanwhile, are processors designed specifically to handle machine-learning workloads.

Reuters reported that OpenAI had started renting Google's AI chips in 2025. The reported plan to purchase them would expand that relationship as the company looks for additional computing resources. Google’s TPUs have also gained traction as an alternative to Nvidia’s processors, with AI companies exploring different hardware options for their workloads.

“Building your own AI chip and having enough computing capacity are two very different problems,” says Bikash Barai, Co-founder of FireCompass. He adds that OpenAI’s custom processors will take time to manufacture, deploy and optimise at scale, while the demand for computing continues to grow.

Hastimal Jangid, Co-founder of RankRabbit AI, says Google's chips could help OpenAI add capacity and potentially access competitive pricing for suitable workloads. He believes using TPUs alongside Nvidia GPUs and its own processors could help the company meet immediate requirements while working towards greater control over hardware costs and performance.

Vikram Raichura, Founder and Managing Director of Helo.ai by VivaConnect, notes, “You cannot put today's customer needs on hold while building the perfect solution for tomorrow.” He says OpenAI needs to balance investment in its own technology with the practical requirement of meeting current computing demand.

Building Chips Isn’t Enough

OpenAI’s custom-chip initiative is part of a broader effort to gain greater control over the infrastructure powering its AI models. In October 2025, the company announced a collaboration with Broadcom to develop and deploy 10 gigawatts of custom AI accelerators and networking systems. OpenAI will design the accelerators, while Broadcom will help develop and deploy them. The rollout was targeted to begin in the second half of 2026 and continue through 2029.

The objective is to build hardware around the requirements of OpenAI’s own models and products. Custom processors could help the company optimise performance for particular workloads and reduce its dependence on commercially available chips over time. However, developing a processor is only the beginning of a much longer process.

Companies must secure manufacturing capacity, high-bandwidth memory and advanced packaging. They also need to ensure that thousands of processors can communicate efficiently and work with the software, models and data centres running their AI services.

“Designing a powerful AI chip is difficult, but making thousands of those chips work efficiently together is an even bigger challenge,” says Barai. He points to networking, power consumption, cooling and manufacturing yield as additional challenges that can affect the cost and performance of a large-scale deployment.

Jangid also highlights the investment required in engineering, testing and specialised infrastructure. Software compatibility is another concern because a chip must work efficiently with the company's AI models and systems. “Such difficulties make custom chips an ongoing expense and not a quick, cheap solution,” he says.

Why Google’s Chips Matter

For Google, making TPUs available to external customers creates an opportunity to generate revenue from years of investment in custom silicon. The company can supply computing infrastructure to businesses that need AI processing capacity, even if they compete with its own AI products.

The opportunity also extends beyond chip sales. Customers using TPUs through Google Cloud may require other services, including computing, storage and networking. This can strengthen Google Cloud's position in the market for AI infrastructure.

Google is expanding the reach of its processors beyond its own cloud platform. On October 27, Google and Oracle announced that TPUs would become available through Oracle Cloud Infrastructure from 2027. The arrangement will allow customers to access Google's processors alongside other AI computing options on Oracle's platform.

“In technology, a competitor in one market can be a very valuable customer in another,” says Barai. He says Google can earn revenue from supplying infrastructure to OpenAI even as the two companies compete in AI models.

Managing Multiple Chip Suppliers

The shift towards multiple chip suppliers is not limited to OpenAI. Anthropic, the company behind Claude, has said it uses Google's TPUs, Amazon's Trainium and Nvidia's GPUs to train and run its models. It has also expanded its computing partnership with Google and Broadcom, with additional TPU capacity expected from 2027.

Using different processors can help companies match hardware to specific workloads, reduce dependence on one supplier and manage supply-chain risks. However, it can also increase software adaptation, engineering and operational costs.

Jangid says companies must weigh the benefits of diversification against the complexity of running different chip platforms. Raichura similarly cautions that a lower hardware price does not automatically translate into savings once integration and management costs are included.

“The important metric is not the price of an individual chip, but the total cost of completing an AI task reliably,” says Barai. For AI companies, the decision is therefore not simply about buying more processors. It is about choosing the right combination of hardware to deliver computing capacity reliably and efficiently.

OpenAI’s reported interest in Google’s TPUs reflects this changing approach to AI infrastructure. Developing custom processors may give the company greater control in the future, but external suppliers remain important as it scales its services. For Google, selling TPUs offers a way to turn its chip investments into a wider infrastructure business. In the AI race, companies may compete over models and products, but their computing needs can create commercial relationships across the same industry.