(Reuters) -SoftBank Group is in talks to lead a funding round of up to $40 billion in artificial intelligence developer OpenAI at a valuation of $300 billion, including the new funds, sources said, in what could be a record single funding round for a private company.

The funding comes as Chinese startup DeepSeek has launched an inexpensive AI model that has caught worldwide attention and challenged expectations of what it costs to develop and deploy AI.

Still, SoftBank has valued OpenAI, the maker of ChatGPT, at $260 billion going into the funding round, up from $150 billion a few months ago, the sources, who requested anonymity to discuss private matters, told Reuters.

The funding is expected to come in the form of convertible notes and, similar to OpenAI's last funding round, is conditioned on OpenAI restructuring its business to remove control of the company by its non-profit arm.

Leading the funding round in OpenAI would mean a bold bet for SoftBank, the global tech investor and its CEO Masayoshi Son, whose group has about $30 billion in cash to deploy, according to filings.

OpenAI and SoftBank declined to comment. The Wall Street Journal reported earlier on Thursday that OpenAI was in talks with SoftBank for an investment round to raise nearly $40 billion that would value the AI startup at up to $340 billion.

SoftBank could invest $15 billion to $25 billion directly into Microsoft-backed OpenAI, some of which may be used to pay for OpenAI's commitment to Stargate, one of the sources added.

Stargate, a joint venture of Oracle, OpenAI and SoftBank, plans to invest up to $500 billion to help the U.S. stay ahead of China and other rivals in the global AI race.

SoftBank's investment would be on top of the $15 billion it has already committed to Stargate, the person said, adding the talks are at an early stage.

In earnings calls this week, the CEOs of Microsoft and Meta Platforms defended their massive AI spending, saying it was crucial to staying competitive in the new field. Microsoft has earmarked $80 billion for AI in its current fiscal year, while Meta has pledged as much as $65 billion.

DeepSeek has attracted attention in global AI circles after writing in a paper last month that the training of its model DeepSeek-V3 required less than $6 million worth of computing power from Nvidia H800 chips.

DeepSeek has used a common technique called model distillation to build on top of frontier AI models including those from OpenAI, and raised concerns among investors on whether labs such as OpenAI can maintain their edge to generate revenue while facing competition from lower-cost rivals.

(Reporting by Krystal Hu, Anna Tong, Milana Vinn; Arsheeya Bajwa in Bengaluru and Juby Babu in Mexico City; Editing Peter Henderson; Bill Berkrot and Muralikumar Anantharaman)