After having been expected to launch in May, DeepSeek R2 still isn’t here.
DeepSeek R1 went viral in early 2025, and there was an expectation that R2 would bring major improvements and even lower costs. DeepSeek R1 stunned the world and tanked the US stock market because its performance was similar to ChatGPT o1, but it was far cheaper to train and run. DeepSeek didn’t have access to the same hardware as US AI firms, like the newest Nvidia chips. Instead, it used whatever AI chips it could buy, legally or through the black market, and relied on software optimizations to train a reasoning model as powerful as what ChatGPT could offer.
At the time, I pointed out that high-end hardware would still be needed to train frontier AI models, even if other AI firms adopted DeepSeek’s software optimizations. These companies would also need to invest heavily in data centers to ensure their services could serve millions of users without issues.
It turns out access to advanced hardware is one reason DeepSeek isn’t able to ship the R2 reasoning model.
A new report from The Information (via Reuters) says DeepSeek has not yet decided when to release the DeepSeek R2 model.
CEO Liang Wenfeng is reportedly not satisfied with the performance of what should be the firm’s best reasoning model. DeepSeek engineers are still working on the model, waiting for the CEO’s approval to release it.
The issue with DeepSeek’s R2 timeline comes down to hardware, which is ironic. Earlier this year, DeepSeek touted its software innovations that allowed it to train and deploy a reasoning model as good as the best ChatGPT version at the time.
It turns out access to enough AI chips from companies like Nvidia is still necessary to roll out a better model. The Trump administration banned certain Nvidia chips from being sold to China, the kind DeepSeek could have used to train AI models.
Most cloud customers using DeepSeek R1 rely on Nvidia H20 chips purchased before the ban. These H20 chips are no longer available to them and are currently the only Nvidia AI chips allowed for export to China.
A surge in demand could overwhelm Chinese cloud providers working with DeepSeek, now that they rely on these Nvidia server chips. That might affect the DeepSeek experience.
As we’ve seen over the years with ChatGPT, it’s not enough for an AI model to be advanced and offer better features than its predecessors. For example, OpenAI also needs the infrastructure in place to provide reliable access to ChatGPT. It must ensure uninterrupted service to hundreds of millions of users around the world. Downtimes must be rare, and ChatGPT has to respond quickly to user prompts.
The surge in popularity brought by the 4o image generation model a few months ago did affect ChatGPT’s performance. The service went down more than once.
The DeepSeek R2 experience could suffer in the same way if Chinese cloud providers can’t meet the demand R2 would likely generate.
While the world waits for DeepSeek R2 announcements, there’s still nothing official about the new reasoning model. Rumors say it will offer better coding capabilities and support reasoning in multiple languages beyond English and Chinese.
Reports from late April also claimed that DeepSeek R2 was 97.3% cheaper to train than GPT-4. The same rumors suggested the cost would drop to $0.07 per million input tokens and $0.27 per million output tokens.
The Information usually has solid insights into the tech industry, often revealing details about new products well before launch. Even if the blog is off about the DeepSeek R2 delays, the fact remains that DeepSeek hasn’t said anything. Something seems off, and the new Nvidia ban looks like a reasonable explanation.
The longer DeepSeek R2 is delayed, the more breathing room US AI firms get. OpenAI, Google, and others can promote their advanced AI products worldwide without pressure from Chinese competition. They also get more time to reduce costs, which is already happening. For example, ChatGPT o3 just became cheaper to use a few days ago. OpenAI will likely continue optimizing its frontier models before challengers like DeepSeek R2 hit the market.