Prompts explode AI agents use 5x more tokens for 10x growth
The relationship between humans and artificial intelligence is no longer a simple comparison; it is a dramatic, accelerating divergence. A recent prediction from Futurum Group CEO Daniel Newman suggests this gap is set to widen exponentially, hinting at a future where AI agents will dominate the digital landscape far beyond human reach.
The data paints a striking picture of this shift. AI is currently being used by AI agents five times more frequently than humans, and this ratio is set to climb dramatically—predicted to reach 10x, 20x, and even 30x in the coming years. This isn’t just a statistical footnote; it represents an exponential scaling of cognitive activity, demanding a fundamental rethinking of how we measure value and adoption.
When examining the sheer scale of this usage, the numbers are staggering. As of August, AI agents are consuming 7.3 trillion tokens, dwarfing the 1.4 trillion tokens used by humans. This vast disparity highlights that the engine driving this new era is not just smarter tools, but systems operating at a scale unimaginable just a few short years ago.
But the true story lies not just in the volume of tokens, but in how they are utilized. A major factor fueling this agentic boom is the power of memory. It turns out that more than 85% of these agent tokens come from cached prompts—the agents are primarily rereading information they have already processed, demonstrating a pattern of highly efficient, though sometimes repetitive, self-correction.
This reliance on memory introduces a critical physical bottleneck. To store this context, models rely on what is known as the KV cache, and this rapidly growing demand for stored context is starting to strain the physical limits of hardware. The memory required for this context is now outpacing the capacity of GPU High-Bandwidth Memory (HBM), forcing a scramble for memory solutions.
This physical constraint translates directly into market forces. The need for high-bandwidth memory is intensifying as agents multiply. For instance, one consultancy testing advanced AI models found that 96% of the input they received was simply re-reading previous conversation—a perfect microcosm of the cached data problem.
As the agentic revolution accelerates, the demand for specialized memory is skyrocketing. Industry experts predict that if Newman’s projection holds true, buyers will be bidding fiercely for memory against even more agents, further pressuring supply chains. This dynamic links AI development directly to the scarcity of physical components like RAM and HBM, suggesting that the next great technological race may be fought over silicon and storage.
Credit
