GPU computing power shortage repeats: major companies like OpenAI and Anthropic are consuming the supply, AI startups are queued until the end of the year

Microsoft, Amazon, and other major cloud providers are consolidating NVIDIA GPU computing power for their internal teams and top clients like OpenAI and Anthropic, leaving small and medium AI startups facing triple pressures: rent increases of 32%, queues until the end of the year, and contract thresholds soaring to hundreds of millions of dollars. Some venture capital firms have begun planning shared compute pools for collective bargaining, while some startups are simply paying out of pocket to buy GPUs to avoid queues.
(Background: NVIDIA’s stock price hits a record high with a market value exceeding $40k; $5 billion invested in Intel has doubled in less than a year, with a book profit of $12.7 billion.)
(Additional background: Full transcript of Jensen Huang’s GTC 2026 speech: AI demand reaches trillions of dollars, computing power jumps 350 times, OpenClaw turns every company into AaaS.)

Table of Contents

Toggle

• Six months ago $2.80, renewal now $3.70
• Microsoft’s tiered system
• Venture capital firms are already finding their own solutions

40k GPUs are online, with demand for 400k in the queue. According to The Information, GPU cloud provider Lightning AI CEO Will Falcon provided a figure that precisely describes the supply-demand ratio of the AI compute market in 2026 as 1 to 10. About 40 customers are waiting on his platform, and rental prices have increased by over 25% in the past six months.

Six months ago $2.80, renewal now $3.70

Image generation startup Krea (total funding of $83 million, with investors including Andreessen Horowitz and Bain Capital Ventures) rented hundreds of NVIDIA Blackwell chips at $2.80 per card per hour six months ago, signing a six-month contract. When renewing, several cloud providers directly stopped answering calls.

It ultimately settled at $3.70, a 32% increase, with the contract extended to one year. Krea CEO Victor Perez said some providers quoted three-year terms to negotiate, while others simply did not respond.

This is not an isolated case; the Ornn Compute Price Index shows that Blackwell’s hourly rental rate has soared to $4.08, a 48% increase in two months.

The one-year lease price for H100 also jumped from $1.70 in October 2025 to $2.35 in March 2026, nearly a 40% increase.

Microsoft’s tiered system

Microsoft implements tiered management for GPU access, prioritizing about 1,000 largest customers (Tier 1) for compute allocation; smaller clients wanting to rent Blackwell must commit to at least 1,000 units for at least one year, with contracts starting at hundreds of millions of dollars.

Pay-as-you-go customers may have their access revoked if GPUs are left idle for a few hours. Startups participating in the free tier of “Microsoft for Startups” are also told that if they do not utilize the resources sufficiently, their GPU permissions will be withdrawn. Azure sales management recently informed staff that GPU wait times for cloud customers are expected to continue until the end of 2026.

Currently, Microsoft is concentrating compute resources for major clients like OpenAI and Anthropic, leaving small and medium AI startups in queues.

Venture capital firms are already finding their own solutions

General Catalyst partner Hemant Taneja has surveyed invested companies about compute bottlenecks, and the firm is planning shared compute pools or collective bargaining through investment consortiums.

An even more extreme case is oil industry AI startup Collide, which plans to spend about $500k to directly purchase NVIDIA GPUs and rent data center space to run their own operations, bypassing queues and price uncertainties. When an AI company in the oil sector decided to revert to buying hardware and renting server rooms, everyone knew that the cloud’s seat had become too crowded.