The large language model (LLMs) war is getting more intense, but two are seeing their limits.
When all began after the moment OpenAI's ChatGPT arrived and rewired public expectations for what AI could feel like, its rapid rise forced Google into its most aggressive response in years. Google responded with Bard, and then Gemini, kicking off a cycle of escalation that pushed both companies to release bigger, faster and more multimodal systems.
By early 2025, the battle had moved far beyond text.
OpenAI that had launched Sora, introduced Sora 2, pushing video synthesis into a new class of realism. Google countered the popularity with Nano Banana, an image generator that became viral in no time. Then, Google released Nano Banana Pro, a model so capable that social media users joked it was "the end of photography."
These two systems quickly became two of the most popular AI products of the year, and that popularity has now become a problem.
With Thanksgiving weekend and the start of the Holiday Season is just above the horizon, millions of people, free from work and experimenting with making surreal videos or hyper-realistic images, pushed Sora 2 and Nano Banana Pro harder than ever.
The trend of people using both Sora 2 and Nano Banana Pro is straining the AI infrastructure behind the scenes.
we’re setting usage limits for free users to 6 gens/day. chatgpt plus and pro users have unchanged limits, and everybody can purchase additional gens as needed. our gpus are melting, and we want to let as many people access sora as possible!
— Bill Peebles (@billpeeb) November 28, 2025
In a quietly but decisive response, both OpenAI and Google began to dial things back.
Bill Peebles, who leads Sora at OpenAI, announced that free users would now be limited to six video generations a day. His explanation: "our GPUs are melting" captured both the technical reality and the humor of the moment. Unlike previous temporary restrictions, OpenAI did not suggest these limits would disappear anytime soon.
Paid tiers remain unchanged, and users can buy additional generations as needed, signaling a firmer shift toward monetization.
As for Google, free users of its Gemini 3-powered Nano Banana Pro now receive only two image generations a day, down from three, and Google’s support pages no longer specify how many prompts Gemini 3 Pro can process for free. Instead, they simply warn that limits may change frequently.
"Limits may change without notice, including due to capacity constraints. When there’s a large increase in activity in Gemini Apps, we may change limits to maintain a high standard of quality. If capacity changes, limits for users without a Pro or Ultra plan may be limited before users with a plan," Google said in a support page.
Meanwhile, the older Nano Banana model still offers 100 images per day, underscoring how sharply Google is separating legacy tools from its newest, highest-demand systems.
Much like OpenAI, Google is preserving generous limits for paying customers while tightening access to its cutting-edge models for everyone else.
What’s happening behind these policy shifts is simple: the AI industry is running into the physical and financial costs of its own success.
High-fidelity video generation consumes orders of magnitude more compute than text, and photorealistic image creation isn’t far behind. The combination of skyrocketing demand, global GPU scarcity, and investor pressure for sustainable revenue models has forced companies to rethink the era of unlimited free experimentation.
For two years, tech giants happily subsidized the cost of mass adoption. Now those bills are coming due.
In the end, nothing stays free forever, especially when it’s being cooked on thousands of power-hungry (and expensive) GPUs.
Meanwhile, the models themselves have reached a level of realism with serious societal implications. Comparisons of Nano Banana versus Nano Banana Pro have gone viral, with many users claiming they can no longer tell the difference between AI-generated portraits and real photographs.
The shock and unease circulating online underline how quickly these tools are reshaping our sense of visual reality.
In just a few days, the AI boom experienced its first real capacity crisis. OpenAI restricted Sora. Google restricted Gemini 3 Pro and Nano Banana Pro. The timing may have been coincidental, but the message is clear: the age of generous, consequence-free free access is fading. As AI becomes more capable, more realistic, and more widely used, the cost of powering it is rising just as fast.
The LLM war is still accelerating, but the companies driving it are learning that scale has limits, and that the next phase of AI may depend as much on economics as on innovation.