Cloud giants are hoarding $NVDA GPUs for their own AI products and biggest enterprise customers, leaving smaller AI startups with worse access and higher prices

The situation is becoming a structural compute squeeze:
GPU rental prices are rising sharply. H100 one-year contract pricing reportedly moved from around $1.70/hour in October 2025 to $2.35/hour in March 2026, a nearly 40% increase
Microsoft is reportedly forcing customers who want Blackwell GPUs to commit to at least 1,000 chips for one year, meaning contracts can easily reach tens of millions of dollars. Smaller startups simply cannot compete with that
Even older Nvidia GPUs now have weeks to months of waiting time, and Microsoft is enforcing a strict use it or lose it policy, removing GPU access if customers leave capacity idle
The biggest cloud customers are being prioritized. Microsoft reportedly has a top tier of around 1,000 major clients that get preferred access, while startups backed by major VCs are still struggling to secure compute
A lot of GPU capacity is still poorly used. A Cast AI report says 95% of provisioned GPU capacity sits idle, because many companies over-reserve GPUs out of fear of losing access
VCs are now trying to solve the problem directly. General Catalyst and others are exploring shared compute pools and direct negotiations to help startups access GPUs