You cannot buy an NVIDIA H100 as an individual or small startup — allocation goes to hyperscalers first

You are a 3-person AI startup that needs 8 H100 GPUs to fine-tune a model. You contact NVIDIA's sales team. They do not return your call — minimum order quantities are 1,000+ GPUs for direct purchase. You try to buy from a reseller. They have a 6-12 month waitlist and mark up prices 50-100% ($40K-60K per GPU). You try cloud (AWS, Azure, GCP). H100 instances are available but cost $3-5/hour per GPU — $2,200-3,600 per GPU per month. For 8 GPUs, that is $17,600-28,800/month in cloud costs. After 12-14 months of renting, you have paid more than the purchase price of the GPUs and own nothing. So what? GPU access determines who can build AI. NVIDIA allocates production first to hyperscalers (Microsoft, Google, Amazon, Meta), then to large enterprises, then to mid-size companies. Startups and researchers get whatever is left — at inflated prices with long wait times. This creates a structural advantage for incumbents: Big Tech can pre-order 100,000 GPUs at volume discounts while a startup cannot buy 8 at list price. The entire AI startup ecosystem depends on cloud GPU rental, which means their #1 cost is a perpetual operational expense, not a one-time capital investment. Why does this persist? NVIDIA's production is constrained by TSMC's fab capacity (4nm process). Total H100/H200 production is estimated at 2-4 million units per year. Hyperscalers pre-order years in advance. There is no GPU spot market with transparent pricing — allocation is relationship-based. CoreWeave and Lambda Labs provide GPU cloud for AI startups but their pricing is only 20-30% cheaper than AWS.