The AI industry’s dog-eat-dog race for compute power will see financial institutions launch ‘GPU debt funds’ in the next 12 months, VC firm predicts

• Generative AI startups have bagged billions in VC funding in an otherwise tepid tech market.
• They rely on compute power — harnessed through GPUs — which is becoming a costly commodity.
• A new report from Air Street Capital made 10 predictions for AI in 2024, including the launch of “GPU debt funds.”

According to a new report, financial institutions are likely to launch “GPU debt funds” in the next 12 months to meet the massive flood of demand from startups seeking to finance their access to graphics processing units.

Nearly a year after the release of ChatGPT-3, the use and variety of applications of artificial intelligence has skyrocketed.

According to Air Street Capital’s State of AI report, released on Thursday, generative AI startups have benefited the most from the boom, attracting more than $18 billion in VC funding by 2023. (This is nearly five times the $3.9 billion invested in the sector by venture capitalists in 2022.)

One of the main challenges that AI startups are currently facing is the need to process large amounts of data in order to generate high-quality output. With rising demand for these services, startups are scrambling to obtain a new form of currency — compute power, according to the report.

Startups can gain access to compute power by purchasing GPUs from chipmakers or renting it from cloud providers. Competition for hardware is fierce, and market leader Nvidia, which has a stranglehold on the chip market, is dealing with a chip shortage. GPU access is not cheap, and it can affect startup fundraising requirements.

“As we’ve firmly entered the era of large-scale AI, where both dataset and model size scaling drives performance improvements, startups developing large AI models are doing what they can to shore up Nvidia chips,” said Nathan Benaich, co-author of the report and partner at Air Street Capital. “This includes raising mega rounds to finance the acquisition of private clusters.”

According to Benaich, some startups with large compute requirements are raising funds by selling equity to venture capital firms in order to rent access to GPUs from their cloud providers.

“This is bad for founders and investors,” said Benaich. “It drives up valuations prematurely and sets unrealistic commercial expectations for the company to hit in order to justify its valuation in the future.”

The report’s authors predicted that financial institutions, such as banks, could launch GPU debt funds to replace VC equity dollars that would otherwise be spent on compute funding.

According to the report, GPUs have long shelf lives and can be used well into the next decade. GPU debt funds are also appealing to regulators, who are “keen to encourage responsible non-dilutive funding,” which typically “carries fewer regulatory requirements than equity financing,” according to Benaich.

It’s not a simple solution, but a few startups have begun to experiment with the idea of using chips as collateral for loans. Coreweave is using its H100 chips as collateral on a $2.3 billion debt facility, which the report’s authors described as risky.

“I don’t think GPU debt funds are going to happen overnight, but with interest rates still high, private credit is becoming increasingly appealing,” Benaich said in an interview with Insider.

According to Air Street Capital, AI companies will enter the mainstream consciousness even more this year. The report’s authors predicted that AI-focused analytics firm Databricks, for example, would file for an IPO. In other news, they predict that an AI-generated song will enter the Billboard 100 Top 10 next year.

Here is Air Street Capital’s complete list of 10 AI predictions for the coming year:

• A Hollywood-quality production employs generative AI for visual effects.
• A generative AI media company is being investigated for misusing its technology during the 2024 US election cycle.
• In a complex environment (e.g., AAA game, tool use, science), self-improving AI agents crush SOTA.
• The tech IPO market is thawing, and at least one major listing for an AI-focused company (e.g. Databricks) is expected.
• As part of the GenAI scaling craze, a group spends more than $1 billion to train a single large-scale model.
• On competition grounds, the FTC in the United States and the CMA in the United Kingdom are looking into the Microsoft/OpenAI deal.
• Beyond high-level voluntary commitments, we see little progress on global AI governance.
• To replace VC equity dollars for compute funding, financial institutions launch GPU debt funds.
• A song generated by AI enters the Billboard Hot 100 Top 10 or the Spotify Top Hits 2024 Top 10.
• Due to issues with enforcement and interpretation, the EU AI Act is not widely adopted as a model of AI regulation.