Barclays Predicts AI Compute Power Surge: A Billion-Agent Boon or Economic Bust?

Barclays recently indicated that by 2025, the current AI computational capabilities are expected to sufficiently support between 1.5 billion and 22 billion AI Agents. This revelation suggests that AI compute power may cover the majority of demand in the U.S. and EU, marking a potential turning point in the industry's focus from "meaningless benchmarks" to practical deployment of AI Agent products.

The Barclays report highlights several key findings about the supply and demand balance for AI compute power. By 2025, there will be approximately 15.7 million AI accelerators—such as GPUs, TPUs, and ASICs—available globally. Of these, 40% are expected to be used for inference underlining a shift towards more economically feasible open-source models. This cost-effective approach is exemplified by Salesforce with its Agentforce, which opts to utilize Mistral open-source models rather than costly proprietary ones.

Despite the apparently ample supply of compute power, Barclays points out structural challenges remain. For AI Agents to truly thrive and provide substantial value to both consumer and enterprise users, future advancements may require smaller, high-performance foundational models, increased inference chip installations, and potential repurposing of existing training GPUs for inference tasks. This signals that while overall compute capability appears to suffice, the dedicated compute needed for efficient, low-cost AI Agent products still faces a significant shortfall.

As the industry stands on the precipice of a burgeoning AI Agent marketplace, the economics of inference costs emerge as a critical factor. Barclays notes the vast number of tokens generated by AI Agents compared to traditional chatbots significantly raises inference costs. The disparate cost-effectiveness across models is stark; for instance, annual subscription costs for an AI Agent based on the OpenAI o1 model can reach $2,400 compared to just $88 for one using the DeepSeekR1 model, which also supports 15 times more users.

Furthermore, the appetite for "super agents" is on the rise, with OpenAI's upcoming premium product expected to process up to 35.6 million tokens monthly from as many as 44 daily queries, dramatically surpassing the capabilities of standard agents. However, the elevated costs associated with these high-demand products could potentially limit their widespread adoption, a point investors must consider seriously when evaluating their economic viability.