AI agents are using 5x more tokens than humans according to latest data from OpenRouter, with this ratio expected to accelerate to 10x and higher as agent adoption continues. As reported by Futurum Group CEO Daniel Newman, agents are using 14x more tokens while human usage is up only 2.8x since the February crossover point. However, the data reveals that more than 85% of agent tokens come from cached prompts, meaning agents are mostly rereading what they've already seen rather than generating new content. This pattern is consistent across platforms, with McKinsey's 2026 State of AI survey showing 40% of large organizations reporting scaled AI agents, up from 27% a year earlier. The trend shows agent usage growing 14x more than human usage while human usage increases only 2.8x, highlighting the fundamental shift in AI infrastructure requirements.
The scale of AI agent adoption is accelerating rapidly across major platforms. Meta's Muse passed 3 million weekly users in less than a month, while OpenAI launched Dots this week, always-on agents that work in the background. Most significantly, developer use of OpenAI's agent tools grew 100-fold in a year. Considering even 10% of Meta's roughly 3.6 billion daily users adopt agent technology, the upside potential for the entire value chain could be enormous. Recent analysis suggests that memory shipped through 2027 could run 33–171 million concurrent frontier-model agents, or billions using efficient open models, highlighting the massive scale of potential agent deployment.
The networking requirements for AI agents are driving significant changes in data center hardware specifications. In Nvidia's GB300 NVL72, GPUs represent 60%+ of the rack's value while networking accounts for only 10%. However, in the Rubin Ultra NVL576, GPUs drop to about 45% while networking increases to 20%. This shift reflects the growing importance of networking infrastructure as AI agents generate constant traffic that cannot be cached like traditional video content. The cost efficiency of this transition is evident in Epoch's recent analysis estimating a 47% per quarter decline in AI performance costs since 2023, which cuts both ways for the buildout - lower hardware ownership and operating costs can reduce investment justification, while more efficient models and serving systems reduce hardware requirements.
The networking bottleneck presents substantial opportunities across multiple sectors. Credo, best known for active electrical cables (AECs), grew revenue 115% last quarter with their CEO noting that optics is growing faster. Lumentum's revenue more than doubled last quarter, while Coherent's datacom and communications sales grew 59%. The competitive landscape includes major players developing agent purchasing capabilities, with Stripe, Visa, Google and OpenAI among companies developing agent payment systems, making Wood's "follow the agents" advice increasingly relevant for investors tracking this infrastructure evolution. Recent analysis suggests that for comparison, leading model developers already generate over $100 billion in combined annualized revenue, with continuing fivefold annual growth potentially reaching roughly $1 trillion by the end of 2027, still significantly below the spending implied by current capacity scenarios.
BlackRock has identified a potential role for cryptocurrency in this emerging ecosystem, arguing that AI agents could create new demand for payment systems built for machines. As reported in their September paper examining AI-digital asset overlap, agents might need to pay for API calls, buy data from other services, or rent computing power without human approval. Stablecoins can move around the clock, while blockchain-based payment protocols allow software to send small payments directly to other software, with Coinbase's x402 designed for machine payments. The economics of this ecosystem are evolving rapidly, with near-term shortages of serving capacity coexisting with potential supply scenarios, while deployment delays give demand more time to grow alongside accumulated hardware capacity.
