OpenAI has unveiled its latest AI model, GPT-4.5, but the firm鈥檚 boss says it is running out of hardware to power it. If ever-larger AI can no longer be run at scale, then are we looking at the end of the technology鈥檚 rapid progress, and perhaps even the bursting of a bubble?
There are certainly signs that things aren鈥檛 going as planned within OpenAI. As recently as 12 February, CEO Sam Altman that the company鈥檚 product offering had created a confusing picture 鈥� at the time of writing, OpenAI offered 鈥� and expressed a desire to return to a 鈥渕agic unified intelligence鈥� instead. That unified model was intended to be GPT-5, and it was to be offered at a limited level to even non-paying customers of OpenAI.
But at a launch event yesterday, OpenAI instead offered an incrementally updated version of GPT-4. A company called GPT-4.5 its 鈥渓argest and best model for chat yet鈥�, but Altman said a lack of computing capacity meant it could only offer the product to a small number of customers. 鈥淚t is a giant, expensive model,鈥� . 鈥淲e鈥檝e been growing a lot and are out of GPUs [processors that provide the computing power for AI].鈥�
Advertisement
As a result of this, its new model high. GPT-4.5 costs $75 per 1,000,000 tokens of input, and $150 per output. Its cheapest model costs $0.15 and $0.60 respectively.
Altman鈥檚 comments suggest GPT-4.5 is far larger than previous models, despite it being more than 10 times more efficient than GPT-4, according to . OpenAI didn鈥檛 respond to a request for comment or clarification.
The constant scaling-up that has delivered rapid progress in AI cannot go on forever, says at AI company Hugging Face. 鈥�The current way of training and deploying LLMs [large language models] is grossly inefficient 鈥� it鈥檚 essentially brute-forcing intelligence. Of course that鈥檚 bound to hit a wall,鈥� she says.
While Altman鈥檚 claimed that GPT-4.5 has 鈥渁 magic to it I haven鈥檛 felt before鈥�, Luccioni is unconvinced. 鈥淯sing terms like 鈥榤agic鈥� and 鈥楢GI鈥� [artificial general intelligence] makes the people making these models seem all-powerful,鈥� says Luccioni. 鈥淏ut I would argue more that Altman is the Wizard of Oz, distracting us so that we don鈥檛 look behind the curtain.鈥�
Indeed, AI companies are reluctant to open up their models to scientific study, partly due to protection of corporate secrets but perhaps also because they don鈥檛 want to expose the sources of their training data.
They are similarly cagey when it comes to revealing exact hardware requirements, energy use or cost. When details are released, such as for DeepSeek 鈥� the chinese model that was claimed to match the performance of cutting-edge models at a fraction of the cost and computational power 鈥� they are hard to verify. In truth, the industry is impenetrable to objective analysis.
at the University of Surrey, UK, says the industry鈥檚 approach over the past five years, to grow ever-larger, consume more energy and feed in more training data, was inevitably going to bump into constraints at some point, but there are efforts underway to overcome or side-step them. 鈥淚f the cost is too high [and] the compute requirement is too high, then it makes it non-viable as a business,鈥� says Rogoyski. 鈥淪o it鈥檚 in everyone鈥檚 interest to bring that down.鈥�
Rogoyski doesn鈥檛 see current LLMs as the long-term future of AI. Techniques such as distillation, which slim down AI models while retaining functionality, may make future models more efficient and cheaper to run. But there are also new architectures on the horizon that could run even existing models faster, including neuromorphic computing, and even quantum computers.
Whether or not companies can become profitable fast enough to remain in business is the 鈥�64-trillion-dollar question鈥�, says Rogoyski. 鈥淚t鈥檚 a bit of a Darwinian soup of ideas at the moment, and there鈥檒l be those that survive and thrive and those that die away.鈥�
