The impact of reasoning AI models from DeepSeek and OpenAI is reportedly expected to shift the focus of artificial intelligence (AI) investments while also boosting AI spending overall.
While the debut of the DeepSeek models led observers to question the need for investment in AI infrastructure, it also led to a greater focus on reasoning models, which require greater spending on inference, Bloomberg reported Monday (March 17).
As a result, Bloomberg Intelligence now expects the investments in AI by hyperscale companies like Amazon, Meta and Microsoft to increase faster than it previously forecast, with more on that money being spent on running AI systems after they have been trained, rather than on data centers and chips, according to the report.
These companies are expected to spend $371 billion on data centers and computing resources in 2025 — 44% more than they spent in 2024 — and $525 billion a year by 2032, the report said.
By 2032, nearly half of all AI spending will be directed toward inference, as reasoning models enable companies to make more money from software, per the report. At the same time, the share of investment directed toward training is expected to drop from 40% to 14%.
DeepSeek shook up the AI world in late January when it released AI models that performed on par with OpenAI’s and Google’s top models but at a fraction of the cost and with far fewer of Nvidia’s GPUs.
Shortly after the release of the DeepSeek AI model that rocked the AI world, Meta CEO Mark Zuckerberg said during an earnings call that the U.S. AI industry is shifting toward AI processing, or inference, as reasoning AI models rise in popularity.
OpenAI released in February what it called its “most cost-efficient” reasoning AI model, the o3-mini, saying it is a “small” but “powerful and fast” model that outperforms earlier models especially in science, coding and math, and comes in three reasoning levels: low, medium and high for tougher tasks.
The model is part of OpenAI’s o1 series, which can reason through tasks but takes longer to respond than non-reasoning models. Reasoning models can also tackle tougher tasks and solve harder problems.
Source: https://www.pymnts.com/
