Shares of Oracle Corp. closed 4% higher today following reports that it’s negotiating a $20 billion deal to provide Meta Platforms Inc. with cloud infrastructure.

Sources told Reuters that the Facebook parent would use the hardware for artificial intelligence training and inference. Separately, Bloomberg reported that the value of the deal could exceed $20 billion. It’s believed other terms may change as well before the contract is finalized.

The reports come two months after Oracle inked an agreement to build 4.5 gigawatts’ worth of data center capacity for OpenAI. One gigawatt corresponds to the power usage of several hundred thousand homes. Shortly after the companies announced their partnership, sources told the Wall Street Journal that the project will be worth $300 billion over five years.

AI infrastructure demand from OpenAI, Meta and other customers has buoyed Oracle’s stock, which is up more than 80% this year. Much of that gain materialized after the company’s earnings report last week. Oracle disclosed that its total remaining performance obligations, a measure of future sales, jumped 359% year-over-year to $455 billion.

The database maker is investing heavily in new infrastructure to meet user demand. Oracle expects to increase its capital expenditures by 65% during the current fiscal year, to $35 billion.

The company’s cloud platform, OCI, offers access to AI clusters with upwards of 100,000 Nvidia Corp. graphics processing units. The clusters also include the chipmaker’s switches and SHARP technology. SHARP reduces the amount of data that GPUs must exchange over the network to coordinate their work, which leaves more bandwidth for other workloads.

Meta may commission AI clusters with a different architecture if it finalizes its $20 billion cloud contract with Oracle. In particular, the Facebook parent might seek to replace some Nvidia products with its internally developed silicon.

In March, sources told Reuters that Meta had begun testing its first custom AI training chip. Earlier, the company detailed an accelerator called MTIA that’s optimized for inference workloads. It’s believed Meta will initially use the chips to power recommendation algorithms.

OpenAI may take a similar approach in its partnership with Oracle. According to the Financial Times, the ChatGPT developer will start mass-producing a custom AI accelerator that it has developed in collaboration with Broadcom Inc. next year. OpenAI has stated that the 4.5 gigawatts’ worth of infrastructure it’s commissioning from Oracle will include more than 2 million chips.

Meta’s cloud deal with Oracle, meanwhile, is likely part of the broad data center initiative it detailed in July. The project will see the Facebook parent invest hundreds of billions of dollars in new AI infrastructure. The first two data center clusters it plans to build, Prometheus and Hyperion, will each require several gigawatts of electricity.

Photo: Oracle