Microsoft to put more computing power behind in-house AI models

The world’s largest software maker has mostly relied on large language models from the ChatGPT maker to power its products

Microsoft plans to expand its physical infrastructure to train its own artificial intelligence (AI) models that it hopes will be competitive with OpenAI, Anthropic and other firms.

The company will make “significant investments” in its own clusters of computing power to train models, consumer AI chief Mustafa Suleyman told employees during an all-hands meeting on Thursday (Sep 11).

It’s critical for a company of Microsoft’s size to have the ability to be self-sufficient in AI if it chooses to, Suleyman told employees, according to attendees who asked not to be named discussing an internal meeting. Microsoft is taking the approach of simultaneously deepening its ties with OpenAI, partnering with other model makers and building its own, he added.

Microsoft declined to comment.

The world’s largest software maker has mostly relied on large language models from the ChatGPT maker to power its products. Still, there have been signs of tension between the companies as each launches competing products and seeks out additional partners. DeepMind co-founder Suleyman joined Microsoft last year to lead development of the company’s own models and create an AI-focused consumer franchise.

Last month, Microsoft released the first of its large language models developed under Suleyman. It was trained on 15,000 Nvidia H100 chips. Cutting-edge models from Meta Platforms, Alphabet’s Google and Elon Musk’s xAI used clusters six to 10 times larger, Suleyman added, suggesting that Microsoft created its model more efficiently.

Microsoft’s plan is to approach all of its products with a multimodel approach, using the AI model customers prefer, chief executive officer Satya Nadella added. Earlier this week, the Information reported that the company plans to use Anthropic models in some of its products. BLOOMBERG