Agentic traffic has surpassed human traffic. This is perhaps a tell that agentic AI token consumption is at least 10x that of chat AI. This explains why Grok AI, mostly used in chat mode, had enough idle GPU resources to rent out to Anthropic and Google. This also explains why memory vendors' capacity is a very long way from ever meeting agentic AI demand.
1/
The AI bottleneck is not the model.
It’s not even the GPU.
It’s memory.
And agentic AI is about to make the shortage much worse.
2/
The key signal:
Agentic/bot traffic has now surpassed human traffic on the internet.
That sounds like a web stat.
It’s actually an AI infrastructure alarm bell.
3/
Why?
Because a human “chat” is usually one prompt, one answer.
An AI agent does not behave like that.
It searches.
Clicks. Reads. Calls tools. Writes. Checks. Retries. Loops.
One task can consume 10x+ the tokens of a normal chatbot session.
4/
This is the difference between chat AI and agentic AI:
Chat AI answers questions. Agentic AI does work. That means the demand curve does not look like consumer software. It looks like labor automation.
5/
And labor automation runs all day.
Sales ops. Finance. Compliance. Support. Coding. Procurement. Research. Security. DevOps. Back office.
Every workflow becomes tokenized.
6/
Most companies have barely adopted agentic AI yet.
That’s the scary part.
We are not late cycle.
We are early.
And memory supply is already getting consumed.
7/
This explains something that looked strange:
xAI/Grok had enough GPU capacity to rent compute to Anthropic and Google.
Why?
Because chat demand is spiky. Agentic enterprise demand is persistent.
8/
Chatbots are products. Agents are infrastructure. Once companies trust them, they do not use them occasionally. They put them inside operations. And operations do not sleep.
9/
Everyone talks about GPUs.
But the real chokepoint is increasingly memory:
HBM. - $MU, $DRAM DRAM. - $MU, $DRAM Advanced packaging. $ASX Bandwidth. $MRVL Capacity. $GOOG, $AMZN
AI is not only compute-bound.
It is memory-bound.
10/
GPUs need high-bandwidth memory to be useful.
More AI accelerators means more HBM.
More HBM means more pressure on DRAM wafer capacity.
More DRAM pressure means tighter supply across the entire memory stack.
11/
This is not something you fix with a software patch.
Memory fabs take years to build.
Capacity is planned years ahead.
Supply cannot magically appear because demand inflected.
12/
So here is the uncomfortable reality:
We are still in the early stages of enterprise agentic AI adoption…
…and we have already stressed the memory supply chain.
That should make people nervous.
13/
The next phase of AI will not be won only by the company with the best chatbot.
It will be won by whoever controls:
Compute. Memory. Power. Datacenters. Supply contracts. Inference efficiency.
14/ The market keeps asking:
“Which model is best?”
Wrong question.
The better question is:
“Who has enough memory to run the agents?”
15/
My view:
Agentic AI demand is structurally different from chat AI demand.
It is heavier, more persistent, and more valuable.
That makes memory manufacturing one of the most important bottlenecks in the entire AI economy.
16/
The punchline:
The AI boom is not waiting on intelligence.
It is waiting on infrastructure.
And the scarcest layer may not be GPUs.
It may be memory.