The GPU shortage is back, and this time it's hitting AI startups where it hurts: their wallets and their timelines.
Multiple AI startups are reporting that cloud providers like Microsoft and Amazon are hoarding GPU compute for their internal teams and biggest customers — namely OpenAI and Anthropic — leaving smaller companies facing price hikes, long wait times, and increasingly draconian contract terms. Microsoft Azure sales management recently told employees that GPU wait times for cloud customers are expected to stretch into the end of 2026.
Take Krea, an image generation startup that raised $83 million from investors including Andreessen Horowitz and Bain Capital Ventures. Six months ago, it rented hundreds of Blackwell chips at $2.80 per card per hour on a six-month contract. When renewal time came, multiple cloud providers stopped answering calls. Krea eventually locked in at $3.70 per hour — a 32% increase — and had to commit to a full year. CEO Victor Perez says some vendors simply went silent, while others wouldn't even talk without a three-year commitment.
It's not just Krea. Will Falcon, CEO of GPU cloud provider Lightning AI, says his company has 40,000 GPUs online, but roughly 40 customers are queuing for a combined demand of 400,000 GPUs. Rental prices have jumped over 25% in six months.
Microsoft has implemented a tiered system for GPU access. Its top 1,000 customers (Tier 1) get priority. Smaller clients looking for Blackwell chips must commit to at least 1,000 GPUs for a minimum of one year — contracts starting in the tens of millions. Pay-as-you-go customers who let their GPUs sit idle for a few hours risk having access revoked. Even startups participating in the "Microsoft for Startups" free credit program have been warned: underutilize and you'll lose your GPU allocation.
The squeeze has caught the attention of venture capital firms. General Catalyst partner Hemant Taneja has sent a survey to portfolio companies about compute bottlenecks, and the firm is exploring shared compute pools or collective bargaining arrangements. Some startups are already considering buying their own hardware. Collide, an oil industry AI startup, plans to spend around $500,000 to purchase NVIDIA GPUs outright and rent data center space, bypassing the cloud queue and uncertainty altogether.
In a similar vein, Chinese AI firm Zhipu AI recently ended its unlimited coding plan as demand surged.