As a shortage of AI chips deepens, big cloud companies are prioritising GPU inventory for internal demand and large customers. That is increasing the burden on AI startups. There are also reports that startups are paying relatively higher costs to cloud companies to access GPU servers.

Even AI startups that have raised large amounts of funding are not immune to the AI chip shortage. The Information recently reported that AI startups backed by major venture capital firms such as Sequoia Capital, Founders Fund, General Catalyst and Andreessen Horowitz have also come under the direct impact of the AI chip supply crisis.

It echoes what happened in early 2023, not long after OpenAI released ChatGPT, when cloud companies focused computing capacity on internal demand and key customers such as OpenAI.

Unlike the early days of AI services, demand for AI coding is now surging, amplifying the fallout from the shortage of computing resources. It is reported that AI startups are finding it harder to access computing power. The Information cited cloud companies and startup founders as saying cloud service providers are limiting GPU capacity for smaller customers as demand from large AI developers such as Anthropic rises sharply.

In the process, cloud companies have been able to raise fees for using Nvidia GPU servers, which has significantly improved cloud providers' profitability, The Information reported.

Krea, an image-generation AI model startup, signed a contract six months ago to lease several hundred Nvidia Blackwell chips for six months at $2.8 per chip per hour. At the time, several cloud companies appeared to be competing to secure Krea as a customer, but the situation has recently changed completely.

Krea sought to buy more AI servers, but some cloud vendors' sales representatives were initially unreachable, it said. Those representatives later got back in touch and proposed terms that included sharply higher prices and a requirement to sign a three-year commitment to do business.

Krea ultimately signed a lease for several hundred Blackwell chips at a price 32 percent higher than before, The Information reported, citing Krea co-founder and CEO Victor Perez (빅터 파레즈).

Even if an AI startup wants to lease 1,000 GPUs, it is not easy to secure space from cloud companies. That is because it is not easy to find customers seeking to use GPU infrastructure on a larger scale than that.

Many companies are also reaching the time to renew past cloud lease contracts, which puts startups at a disadvantage in the current environment. As many AI startups' previously signed two- to three-year cloud contracts expire, cloud companies have gained opportunities to offer them more expensive plans or reallocate that capacity to other customers, The Information reported.

Amid shortages and rising costs, AI startups are also moving to find their own countermeasures.

General Catalyst is pursuing steps such as activating a shared pool of GPU capacity or negotiating directly with cloud companies to help its portfolio firms access GPUs. Some companies are also opting for an on-premises approach, building their own GPU systems by buying GPUs rather than renting them from the cloud.

Collide, an AI agent developer, is one such company. Building GPUs directly costs more in the short term than renting from the cloud, but Collide sees on-premises as a viable tactic given uncertainties around cloud leasing. Collide also expects that over the next few years, building its own system will cost less than renting from the cloud, The Information reported, citing CEO Colin McClelland (콜린 맥클렐랜드).