This is exactly why I remain bullish on AI infrastructure.
Look at GPU rental vs AI token pricing since July. Token costs collapsed. H100 rentals? Still elevated around $2.70/hour.
Classic Jevons Paradox. When something gets cheaper, people use way more of it. AI tasks that were too expensive at scale suddenly become viable. If running an AI agent costs less, companies deploy thousands instead of hundreds. Coding assistants, search, customer support — every inference-heavy use case scales up.
More usage = more demand for GPUs, memory, networking, data centers, power.
Even as token prices crashed, H100 rental pricing held firm. That tells you demand for compute is still strong despite efficiency gains. Yes, supply dynamics and newer GPU generations matter. But the resilience in compute pricing while tokens drop? That's a signal.
I think the market underestimates how much additional AI usage lower prices will unlock over the next few years. We're early. Most businesses have barely started integrating AI into daily ops.
This is why I favor infrastructure suppliers over trying to pick which model provider wins. Whether it's OpenAI, Anthropic, Google, or open source — someone still has to provide the compute. As long as AI usage grows faster than efficiency cuts compute needs, the buildout continues.
Bullish on compute, memory, networking, and power.
Look at GPU rental vs AI token pricing since July. Token costs collapsed. H100 rentals? Still elevated around $2.70/hour.
Classic Jevons Paradox. When something gets cheaper, people use way more of it. AI tasks that were too expensive at scale suddenly become viable. If running an AI agent costs less, companies deploy thousands instead of hundreds. Coding assistants, search, customer support — every inference-heavy use case scales up.
More usage = more demand for GPUs, memory, networking, data centers, power.
Even as token prices crashed, H100 rental pricing held firm. That tells you demand for compute is still strong despite efficiency gains. Yes, supply dynamics and newer GPU generations matter. But the resilience in compute pricing while tokens drop? That's a signal.
I think the market underestimates how much additional AI usage lower prices will unlock over the next few years. We're early. Most businesses have barely started integrating AI into daily ops.
This is why I favor infrastructure suppliers over trying to pick which model provider wins. Whether it's OpenAI, Anthropic, Google, or open source — someone still has to provide the compute. As long as AI usage grows faster than efficiency cuts compute needs, the buildout continues.
Bullish on compute, memory, networking, and power.