Does U.S. have enough AI capacity for agents?

Investing.com -- The U.S. AI industry appears to have sufficient infrastructure to support the emerging wave of agentic products—software systems that perform complex tasks autonomously—but some questions remain around affordability and future scalability.

According to Barclays (LON:BARC), the industry is on track to have about 16 million accelerators online in 2025, with roughly 20% potentially available for agent inference workloads.

“Admittedly, most of this is running chatbots today, but we expect those products to be replaced eventually by much more capable agents,” Barclays analysts led by Ross Sandler said in a note.

Using a range of model assumptions, Barclays estimates the sector could support between 1.5 billion and 22 billion agents next year.

“Agents that run on expensive models (like OpenAI’s o1) result in lower capacity. More likely we expect most agentic products to run on lower cost models (like DeepSeek R1, Llama or Mistral),” the analysts said.

A move toward more efficient models like DeepSeek could enable the industry to support “15x more users” compared to the current state.

In terms of demand, enterprise adoption alone could be vast. According to Barclays, there is potential to replace over 1 billion enterprise software seat licenses and serve upwards of 10 billion enterprise tasks.

Consumer markets would add further pressure, with a global use base of more than 3 billion likely to “gobble up a huge percentage” of available inference compute.

“As the inference cost curve bends down further, and as more accelerators are installed, industry capacity would go up notably and be in a position to handle both consumer and enterprise agents,” analysts continued.

OpenAI’s anticipated “super agents”—with pricing ranging from $2,000 to $20,000 per month—may further stretch available capacity.

These agents are expected to consume significantly more tokens per query than chatbots, driving much higher compute needs. Agentic AI products have a notably higher token consumption trend compared to chatbots due to their autonomous nature.

The economics of agent subscriptions remain a key challenge. Products built on large proprietary models like OpenAI o1 may be “untenable” at consumer-friendly price points, whereas those using more efficient alternatives like DeepSeek R1 can operate at dramatically lower costs.

Salesforce’s (NYSE:CRM) Agentforce, for example, already runs on Mistral’s open-source models ranging from 7 billion to 141 billion parameters.

“As more inexpensive models become available (Llama 4 or DeepSeek R2) later this year, we should see this shift to lower-cost models powering agents accelerate,” analysts noted.

They expect the growing use of open-source models to ease cost pressures and extend scalability. But longer term, sustaining agent growth at scale will likely require “cheaper, smaller and similarly high performing foundation models,” more inference chips, and potentially a reallocation of training hardware toward inference workloads.