Qualcomm CEO: AI Computing Supply-Demand Crunch May Persist Until H2 2027

Taylor Wilson
Published 2026-06-17About 6 min read

Qualcomm CEO Cristiano Amon warns that memory and wafer capacity cannot keep pace with surging AI demand, a bottleneck he expects to persist until the second half of 2027. This means compute is no longer a commodity you buy on demand — it is becoming the scarce resource that defines who wins in AI.

01

Why is compute suddenly not enough?

Demand for AI training and inference is growing faster than memory and wafer production can expand.
In plain terms = chip factories have a fixed output ceiling, but the number of AI companies lining up to use those chips has spiked far beyond it.
Qualcomm CEO Amon compares the current crunch to the pandemic-era chip shortage and expects no relief until the second half of 2027.
02

GPU rental prices doubled — what does that signal?

Baseten CEO Tuhin Srivastava says Nvidia B200 GPU cloud-rental rates jumped from $2.63 to $5.10 per hour at renewal — nearly double.
This means → it is not just new GPUs that cost more; even already-deployed hardware commands a rising scarcity premium at contract renewal.
This reflects a deeper signal: the market is not short of better GPUs — it is short of GPUs, period.
03

Why does "owning 1,000 GPUs" equal a competitive moat?

Srivastava puts it bluntly: the core bottleneck in AI is no longer model capability — it is compute supply.
In plain terms = the race is not about "whose AI is smarter" but about "who has enough GPUs to train and run AI at all."
Holding 1,000 GPUs effectively locks in more than a year's worth of future AI production capacity — compute access itself has shifted from a cost line to a competitive moat.
04

Can Qualcomm's "disruptive roadmap" change the equation?

Amon hints at a "disruptive technology roadmap" to be unveiled in late June, covering edge devices, data centers, and memory efficiency.
This means → Qualcomm is attacking the bottleneck from the demand side — not building more chips, but making each chip and each memory unit work harder.
Whether that roadmap delivers will be tested at a clear checkpoint: the second half of 2027, when supply is supposed to catch up with demand.

Content is for reference only, not financial advice.