The Chinese artificial intelligence company Deepseek has shaken the markets this week with states that its new model of AI surpasses the openings and costs a fraction of the price to build.
Affirmations – in particular that the Grand Language Model of Deepseek costs only $ 5.6 million to train – have aroused concerns about the sums of eyewitness that the technology giants are currently spending in computer information required for Train and execute workloads advanced.
The investor fears that the disruptive impact of Deepseek has erased nearly $ 600 billion in Nvidia market capitalization on Monday – the biggest drop in a day for any business in the history of the United States.
But not everyone is convinced by Deepseek’s claims.
CNBC asked industry experts their point of view on Deepseek, and how he really compares himself to Openai, creator of Viral Chatgpt Chatbot who triggered the AI revolution.
What is Deepseek?
Last week, Deepseek released R1, his New reasoning model that rivals the O1 of Openai. A reasoning model is an important language model that breaks down the invitation into smaller pieces and considers several approaches before generating an answer. It is designed to deal with complex problems in the same way as humans.
Deepseek was founded in 2023 by Liang Wenfeng, co-founder of the quantitative High-Flyer focused on AI, to focus on important language models and reaching artificial general intelligence, or AG.
Act as a concept refers to the idea of an AI that is equivalent or exceeds human intellect on a wide range of tasks.
Much of the technology behind R1 is not new. What is notable, however, is that Deepseek is the first to deploy it in a very efficient AI model with – according to the company – considerable reductions in energy needs.
“The point to remember is that there are many possibilities to develop this industry. The high-end chip / capital route is a technological approach,” said Xiaomeng Lu, director of geo-technology practice of the Eurasia group.
“But Deepseek proves that we are still at the nascent stage of the development of the AI and the path established by Openai is perhaps not the only route to a very competent AI.”
How is it different from Openai?
Deepseek has two main systems that have collected the Buzz of the AI community: V3, the large language model that imposes its products, and R1, its reasoning model.
The two models are open-source, which means that their underlying code is available and accessible to the public so that other developers can personalize and redistribute.
Deepseek models are much smaller than many other large -language models. V3 has a total of 671 billion parameters or variables that the model learns during training. And while Openai does not disclose the parameters, the experts consider its latest model to have at least one Billion.
In terms of performance, Deepseek said Its R1 model reaches performance comparable to the O1 of OPENAI on reasoning tasks, citing benchmarks, notably likes 2024, Codés, GPQA Diamond, Math-500, MMLU and Swe-Bench verified.
In a technical report, the company said that its V3 model had a training cost of only $ 5.6 million – a fraction of billions of dollars than notable Western IA laboratories such as Optaai and Anthropic spent on Train and manage their fundamental AI models. However, it is not yet clear how many deep costs to operate.
If the training costs are accurate, however, this means that the model has been developed at a fraction of the cost of rival models by Openai, Anthropic, Google and others.
Daniel Newman, CEO of Tech Insight Firm, The Futurum Group, said that these developments suggest “a massive breakthrough”, although he has lost doubt about the exact figures.
“I believe that the breakthroughs of Deepseek indicate a significant inflection for the scaling of laws and are a real necessity,” he said. “That said, there are still a lot of questions and uncertainties around the full image of costs with regard to the development of Deepseek.”
Meanwhile, Paul Triolio, senior vice-president for China and the technology policy manager of the DGA Group consulting firm, noted that it was difficult to make a direct comparison between the cost of the Deepseek model and that of the main American developers.
“The figure of 5.6 million for Deepseek V3 was just for a single training, and the company stressed that this did not represent the overall cost of R&D to develop the model,” he said. “The overall cost was then probably considerably higher, but still lower than the amount spent by large companies in the American AI.”
Deepseek was not immediately available to comment when contacted by CNBC.
Deepseek comparison, Openai on the price
Deepseek and Openai both disclose prices for the calculations of their models on their websites.
Deepseek says that R1 costs 55 cents per 1 million input tokens – “tokens” referring to each unit of individual text dealt with by the model – and $ 2.19 per 1 million production tokens.
In comparison, the Openai pricing page for O1 shows that the company invoices $ 15 per million entry tokens and $ 60 per million production tokens. For GPT-4O Mini, the smaller and low-cost language model, the company charges 15 cents per million entry tokens.
Skepticism on chips
The revelation of Deepseek de R1 has already led to a public debate animated on the veracity of its claim – in particular because its models have been built despite the export controls of the United States restricting the use of advanced ia chips towards China.
Deepseek claims to have had its breakthrough using mature Nvidia clips, including the H800 and A100 chips, which are less advanced than the cutting -edge H100 of the flea manufacturer, which cannot be exported to China.
Nvidia has since released and said that the GPUs used by Deepseek had fully in line with exports.
The real deal or not?
Industry experts seem to be suitable that what Deepseek has achieved is impressive, although some have urged skepticism with regard to some of the claims of Chinese society.
“Deepseek is legitimately impressive, but the level of hysteria is an indictment of so many people,” wrote the American entrepreneur Luckey, who founded Oculus and Andundil on X.
“The number of $ 5 million is a false. It is pushed by a Chinese hedge fund to slow investments in AI American startups, serve their own shorts against American titans like Nvidia and hide the escape from sanctions . “
Seeda Rejal, commercial director of Netmind, a startup based in London which offers access to Deepseek AI models via a distributed GPU network, said that he saw no reason not to believe Deepseek.
“Even if it is extinct by a certain factor, it is still as very effective,” rejal at CNBC during a telephone interview earlier this week. “The logic of what they explained is very sensible.”
However, some have said that Deepseek technology might not have been built from zero.
“Deepseek makes the same mistakes made by mistakes, a strong indication that technology has been torn off,” said billionaire investor Vinod Khosla on X, without giving more details.
This is an assertion to which Openai himself alluded, telling CNBC in a press release on Wednesday that he examined Deepseek reports could “inappropriately” have “used output data from his models to develop their model of AI, a method called “distillation”.
“We take aggressive and proactive countermeasures to protect our technology and will continue to work closely with the US government to protect the most competent models here,” an OpenAI spokesperson told CNBC.
Commodification of AI
However, the meticulous examination surrounding Deepseek is shaking, the scientists of the AI are largely agree that he marks a positive step for the industry.
Yann Lecun, chief scientist AI Metasaid Deepseek’s success was a victory for Open Source models, not necessarily a victory for China over the American meta is behind a popular Open Source model called Llama.
“To people who see Deepseek’s performance and think:” China exceeds the United States in AI. “You read this badly.
“Deepseek took advantage of open research and open source (for example Pytorch and Llama de Meta). They proposed new ideas and built them in addition to the work of others. Because their work is published and open source , everyone can take advantage of it.
WATCH: Why Deepseek endangers the head of America’s AI in Jeopardy
– Katrina Bishop of CNBC and Hayden Field contributed to this report
