Table of Contents
ToggleThere are 40,000 GPUs online, with a waiting list of 400,000. According to The Information , Will Falcon, CEO of GPU cloud provider Lightning AI, provided figures that precisely describe the supply-demand ratio of the AI computing power market in 2026 as 1 to 10. His platform has approximately 40 customers waiting, and rental prices have increased by more than 25% in the past six months.
Six months ago it was $2.80, now it's $3.70 after renewal.
Image generation startup Krea (which has raised $83 million in funding, including from Andreessen Horowitz and Bain Capital Ventures) leased hundreds of NVIDIA Blackwell chips six months ago for $2.80 per chip per hour, signing a six-month contract. When it came time to renew, several cloud providers simply stopped answering their calls.
It ultimately sold for $3.70, a 32% increase, and the contract term was extended to one year. Krea CEO Victor Perez stated that some companies only agreed to negotiate after offering three-year contracts, while others simply did not respond.
This is not an isolated case. According to the Ornn Compute Price Index, hourly rental rates in Blackwell have soared to $4.08, a 48% increase in two months.
The price of a one-year lease on the H100 also jumped from $1.70 in October 2025 to $2.35 in March 2026, an increase of nearly 40%.
Microsoft's rating system
Microsoft implements a tiered management system for GPU access, prioritizing computing power allocation to approximately 1,000 of its largest customers (Tier 1); smaller customers who want to lease Blackwell must commit to at least 1,000 GPUs for at least one year, with contracts starting at tens of millions of dollars.
Microsoft may revoke the access rights of pay-as-you-go customers if they leave their GPUs idle for even a few hours. Startups participating in the "Microsoft for Startups" free credit program have also been informed that their GPU licenses could be revoked if they do not fully utilize them. Azure sales management recently informed employees that the GPU wait time for cloud customers is expected to continue until the end of 2026.
Microsoft is now concentrating its computing power on large clients like OpenAI and Anthropic, leaving small and medium-sized AI startups to wait in line.
Venture capitalists have started to find their own solutions.
Hemant Taneja, a partner at General Catalyst, has sent questionnaires to portfolio companies to investigate computing power bottlenecks. The firm is planning to share computing power pools or allow portfolio companies to collectively negotiate prices.
A more extreme example is Collide, an AI startup in the oil industry, which plans to spend approximately $500,000 to directly purchase NVIDIA GPUs and rent data center space to run them itself, in order to avoid waiting lists and price uncertainties. When an AI company in the oil industry decides to return to the old path of buying hardware and renting data centers, everyone knows that the chairs in the cloud are already too crowded.
Related reports
How amazing is Nvidia's H100 AI chip? Why is it so hard to find?
A First Look at GPU-EVM: A Rising Star in the Parallel EVM Track, Using EVM to Train AI Agents
