Karina Hong

"We're also you know scaling inference we're seeing scaling inference almost has no wall... I think there will be compute clause [costs]." The next wave of AI development is moving beyond simple chatbots into formal mathematical verification and agentic reasoning. These processes require models to "think" step-by-step, which demands massive, continuous inference compute. Startups are raising hundreds of millions of dollars specifically to funnel into these compute costs, directly benefiting Nvidia's hardware monopoly. LONG. The shift from pre-training to inference-heavy reasoning models creates a virtually limitless ceiling for GPU compute demand. Hyperscalers successfully develop and deploy custom silicon (ASICs) that offload inference workloads from Nvidia GPUs at a significantly lower cost.