Until recently, much of the investment in AI has gone to data centres and chips that are used to train, or develop, massive new AI models. Now, tech firms are expected to move more spending to inference, or the process of running those systems after they’ve been trained.
The shift in investment has been accelerated by the release of new reasoning models from OpenAI and China’s DeepSeek, among other companies, the report said. These systems take more time to compute responses to user queries, mimicking the process of how humans think through problems.
The rise of DeepSeek, which claimed to develop a competitive model for a fraction of the cost of some leading US rivals, prompted questions in the US tech industry over heavy investments in developing AI. Some leading AI companies are now embracing more efficient AI systems that can run on fewer chops.
But reasoning models also offer new opportunities to make money from software, according to the report, and potentially transfer more of the cost from the development stage to after the model is rolled out. That’s likely to drive greater investment in this approach, and boost spending on AI overall, the report said.
See also: Musk's xAI seeks up to US$200 billion valuation in new funding
“Capital spending growth for AI training could be much slower than our prior expectations,” Mandeep Singh, an analyst with Bloomberg Intelligence, wrote in the report. But the immense amount of attention on DeepSeek, he wrote, will likely push tech firms to “increase investments” in inference, making it the fastest-growing segment in the generative AI market.
While training-related spending is expected to make up more than 40% of hyperscalers’ AI budgets this year, that segment is expected to drop to just 14% by 2032, according to the report. By contrast, inference-driven investments could make up nearly half of all AI spending that year.
Alphabet's Google appears best positioned to make this pivot quickly, thanks to its in-house chips that handle both training and inferencing, Singh writes. Other companies, such as Microsoft and Meta, have leaned heavily on Nvidia chips and might not have as much flexibility.
