Global token demand is 31.7 billion per 10 seconds in 2026. Qualcomm CEO Cristiano Amon projects 1.27 trillion per 10 seconds by 2030, a 40x increase.
Key facts
- 31.7B tokens per 10 seconds in 2026.
- 1.27T tokens per 10 seconds by 2030.
- 40x increase projected by Qualcomm CEO.
- Agent-paced activity cited as primary driver.
- 127M H100-equivalent GPUs needed at current efficiency.
The projection, shared via X by AI commentator @rohanpaul_ai According to @rohanpaul_ai, frames the token explosion as a shift in AI's economic structure. Amon's core argument: the surge is not about smarter answers but about AI moving from human-paced interaction to Agent-paced activity.
Once agents become persistent, the economy of AI inference becomes background infrastructure. Every useful action carries a hidden bill: context must be carried, memory must be updated, sensors may need to be interpreted, and mistakes must be caught before they become expensive. This implies that token consumption scales not with user queries but with the number of autonomous loops running in parallel.
To put the numbers in context: 31.7 billion tokens per 10 seconds translates to roughly 3.17 billion tokens per second in 2026. By 2030, that reaches 127 billion tokens per second. For reference, GPT-4-level models consume roughly 1-2 tokens per inference step per user; a persistent agent running at 10 steps per second would burn 10-20 tokens per second per agent. Achieving 127B tokens per second implies billions of concurrent agent instances globally.
Key Takeaways
- Qualcomm CEO projects token demand rising 40x by 2030, driven by persistent AI agents.
- Infrastructure implications are massive.
The infrastructure implication
If Amon's trajectory holds, the hardware required to serve this demand dwarfs current data-center buildout. A single H100 GPU processes roughly 1,000 tokens per second for inference. Serving 127 billion tokens per second would require 127 million H100-equivalent GPUs — roughly 100x the current global installed base of AI accelerators. That suggests either radical efficiency gains (e.g., 100x improvement in tokens-per-watt by 2030) or a hardware buildout of unprecedented scale.
Qualcomm, as a mobile-first chip designer, has a clear incentive to frame the future as edge-heavy: on-device inference that reduces the data-center burden. Amon's statement does not specify the split between cloud and edge token processing, but the 40x multiplier implicitly argues that the marginal cost of a token must fall dramatically for the economics to work.
The agent-driven demand curve
The unique take here is that token demand is decoupling from human attention span. In 2024, most token consumption correlated with human reading speed — roughly 5-10 tokens per second per user. Agent-paced activity removes that ceiling. A single agent monitoring a supply chain could generate millions of tokens per hour without any human in the loop. The 40x multiplier may prove conservative if enterprise agent adoption accelerates beyond current projections.
The source is a single X post quoting a Qualcomm CEO remark, not a published report. No training data, benchmark results, or architectural details accompany the claim. Take the numbers as directional rather than precise.
What to watch
Watch for Qualcomm's next earnings call (expected late April 2026) where Amon may disclose internal token-demand projections or edge-inference chip roadmap updates. Also track NVIDIA's data-center revenue growth rate as a proxy for actual token consumption scaling.
