Interview with an
$INTC employee on why agentic AI is creating a new layer of CPU demand (
$NVDA,
$AMD,
$TSM ):
- The expert sees agentic AI as a meaningful driver of CPU demand growth beyond that required by traditional LLM inference. Where standard deployments use CPUs primarily to manage GPU tasks, agentic architectures introduce an entirely new layer of CPU-intensive work, encompassing agent orchestration, tool calls, and API interactions. As agentic adoption scales, the expert expects this to shift data center configurations meaningfully toward more CPU capacity relative to GPUs.
- The expert explains that the lower pricing for B200 and B300 relative to H100 is primarily a supply-side dynamic. As the supply chain matures around newer architectures,
$TSM and the broader ecosystem can produce Blackwell chips at greater volumes than they could for H100, enabling hyperscalers to lower per-instance prices to stimulate adoption.
- The more important point the expert makes is that lower instance pricing does not mean lower bills. Running the latest models on newer hardware generates significantly more tokens and traffic per session, meaning overall customer spend actually increases even as the headline rental rate comes down. The expert sees this as a deliberate strategy across the supply chain, in which every player is positioning itself to grow the overall revenue pool rather than simply competing on price.
- The expert sees agentic AI driving demand for both high-end GPUs and CPUs rather than shifting away from one toward the other. The LLM component of agentic workflows still requires the most capable chips, while the orchestration and tool-calling layer is driving a clear and growing increase in CPU demand, something
$INTC has already flagged by noting it is capacity constrained on CPUs.
- On the competitive landscape for CPUs, the expert sees the outcome as still too early to call. x86 from
$INTC and
$AMD has deep roots in the kinds of orchestration workloads that agentic AI demands, having handled similar tasks for years.
$NVDA's Vera CPU and
$ARM's recent AGI chip are making a push into this space, but the expert expects the established ecosystem advantages of
$INTC and
$AMD to be difficult to displace quickly.
- The expert expects training workloads to continue growing in intensity, driven by the industry's broad consensus that larger models consistently produce more capable and intelligent outputs, even in ways that cannot always be fully explained. This emergent intelligence from scale is what continues to justify investment in increasingly dense and powerful cluster architectures like the
$NVDA's Kyber rack.