In their 30s, office worker Kim often uses generative artificial intelligence (AI) "ChatGPT" for work. They consult ChatGPT for advice when writing emails to foreign clients or searching for information. However, after the recent trend of changing photos into "Ghibli-style" illustrations, Kim feels that ChatGPT's responses have become slower.
Kim noted, "While a simple explanation or translation usually takes 10 seconds, it's now taking 30 seconds or even 1 minute to get a response. It feels like ChatGPT is overloaded."
As demand for OpenAI's image-generating AI model "ChatGPT-4o Image Generation" (Imagegen) has exploded, ChatGPT is experiencing severe system overload. Industry analysts suggest that OpenAI could lag behind competitors like Anthropic and Perplexity in the long term if they cannot secure graphics processing units (GPUs), essential components for running their AI models.
◇ Altman pleading for "100,000" GPUs
Sam Altman, CEO of OpenAI, stated on the 1st (local time) via X that there has been a surge in subscribers wanting to use ChatGPT Imagegen, which could cause some functions to slow down. Altman said, "We are managing the situation, but please consider that OpenAI's new services may be delayed, that some functions may not operate properly, or that service might occasionally slow down due to capacity issues."
Altman also mentioned, "If we can secure 100,000 GPUs, please contact us as soon as possible." Previously, he had remarked on the 27th of last month that "OpenAI's GPUs are melting down," highlighting the popularity of ChatGPT Imagegen.
GPUs are the core hardware determining the performance capabilities of large language models (LLMs). The better the GPU performance, the faster the training and response rates of generative AI, and the more efficiently complex tasks can be handled. For instance, Google’s AlphaGo, which competed against Lee Sedol in the Go championship, used 280 GPUs. However, GPT-4 is expected to be equipped with 25,000 NVIDIA A100 GPUs.
◇ Is the popularity of Imagegen actually a negative factor?
If OpenAI is unable to secure additional GPUs, the server overload situation is likely to persist for the foreseeable future. OpenAI also plans to open access to their deep research tool, which is currently available only to paid subscribers, to free users soon. Deep research is an AI agent that writes expert-level research reports through deep search and reasoning.
Meanwhile, according to OpenAI, as of the 31st of last month, 1 million new ChatGPT users have been added within one hour. Currently, ChatGPT has 500 million weekly users, with around 20 million paid subscribers.
한편 오픈AI에 따르면 지난달 31일 기준으로 1시간 만에 챗GPT 신규 사용자 100만명이 추가됐다. 현재 챗GPT의 주간 사용자는 5억명, 유료 구독자는 2000만명 수준이다.