The signs that large-language models are starting to hit a ceiling in capabilities regardless of how much money and computing resources are thrown at them are piling up. Late last year, the founders of venture capitalist firm Andreessen Horowitz noted in an interview that the gains delivered by each successive generation of AI models were getting smaller. "They're sort of hitting the same ceiling on capabilities," Andreessen said about the various companies working on advanced AI models.

Part of the problem is data. At this point, the best AI models are already trained on essentially all the digital data available. Without more data, improvements in capabilities will have to be driven by new training methods or other innovations.

OpenAI kicked off the AI boom in late 2022 with ChatGPT, which was powered by the company's GPT-3.5 model. GPT-4 soon followed, which marked a major improvement in capabilities, and the company later launched a series of other models as part of the GPT-4 family.

GPT-4 launched nearly two years ago, and GPT-5 is expected to finally be released soon. The Wall Street Journal reported late last year that GPT-5 was behind schedule and costing the company a fortune. Given the explosion of competition since GPT-3.5 was released, OpenAI needs to prove that major improvements are still possible.

A bizarre product launch

As the world waits for GPT-5, OpenAI launched GPT-4.5 on Feb. 27. The company was quick to note that GPT-4.5 was not meant to be a replacement for GPT-4o, currently its most capable model, but rather an alternative for certain tasks like writing and brainstorming.

There are two things that make this launch strange. First, GPT-4.5 is at best a minor upgrade over GPT-4o at certain tasks. Second, the pricing is so high that there are unlikely to be many viable commercial applications. For those using OpenAI's APIs, GPT-4.5 is 30 times more expensive than GPT-4o for input tokens and 15 times more expensive for output tokens. "GPT‑4.5 is a very large and compute-intensive model, making it more expensive⁠ than and not a replacement for GPT‑4o," reads OpenAI's blog post announcing the new model.

OpenAI is limiting access to GPT-4.5 for now because it doesn't have enough Nvidia (NVDA 1.13%) GPUs to power the new model at a larger scale. The company is working on getting more GPUs and eventually making the model more broadly available.

Awful news for Nvidia

While OpenAI's need for more GPUs to power its latest model could be construed as a positive for AI accelerator champion Nvidia, the fact that this model is so expensive to run that it's effectively dead on arrival for any real-world applications is an enormous negative.

Nvidia's growth story hinges on a few assumptions. First, AI models will need an ever-increasing amount of computing power for training and inference. Second, AI models will meaningfully improve in capabilities as more computing power is thrown at them. And third, companies investing in AI will achieve acceptable returns on investments.

GPT-4.5 provides more evidence that large-language models are hitting a ceiling, and that no amount of additional computing power is going to unlock world-changing improvements. The reported delays for GPT-5 only add to this argument. If OpenAI is having trouble, you can bet other AI companies are as well.

As it stands today, AI models are useful and have plenty of real-world applications. Making AI models more efficient, like what Chinese start-up DeepSeek accomplished, could expand the market. But that's probably not enough to keep the AI boom going. Nvidia needs companies training new AI models to flock to each new generation of AI accelerator and continually invest in new AI datacenter capacity. In a world where incremental improvements in AI models involve dramatically higher costs, like with GPT-4.5, the math starts to fall apart.

I could be wrong, but it looks like GPT-4.5 could mark the beginning of the end of the AI bubble.