
Chinese artificial intelligence firm DiPesic has shaken the markets this week, claiming that its new AI model improves Openi and is part of its cost.
This claim – in particular that DiPsic’s large language model costs only $ 5.6 million – has raised concerns over the amount of iwattering that tech giants are currently training advanced AI workload Spending on computing infrastructure needed to give and run.
Investors are feared that the impact of DipoC’s disruption was erased near NVIDIA’s market capitalization on Monday to $ 600 billion-which is the largest single day for any company in US history.
But not everyone is convinced of the claims of DPC.
The CNBC asked industry experts for their thoughts about Depsek, and how does this in fact compare the viral chat -boot -chatting creator to Openi, which gave rise to the AI revolution.
What is the Deep Soccer?
Last week, DiPsic released R1, it The new argument model that rivals Openi’s O1 rivals. An reasoning model is a large language model that erupts into small pieces and considers many perspectives before reacting. It is designed to deal with similar complex issues for humans.
Deep Sak was based in 2023 by AI Foxed Constitutional Hedge Fund co -founder Liang Wenfering, co -founder of the High Flyer, to focus on large language models and reach artificial general intelligence, or AGI.
AGI refers to the idea of an AI easily as a concept that goes beyond or beyond the human intellect on widespread tasks.
Most of the technology behind the R1 is not new. However, it is noteworthy that DiPsic has first deployed it to the high-performing AI model-according to the company-there is a significant reduction in the need for electricity.
“It is a lot of possibilities for the industry to develop the industry,” said Xiang Lu, director of the Geo -Technology Practice, Eurasia Group.
“But Dipic proved that we are still in the newborn phase of AI development, and that the path established by the Open cannot be the only way to AI.”
How is it different from Openi?
DPCAC has two main systems that have collected Buzz by the AI community: V3, a large language model that opens its product, and R1, its argument model.
Both models are open sources, which means their basic code is available free and publicly available to other developers to customize and re -distribute.
Deep Sak models are much smaller than many other large language models. The V3 contains a total of 671 billion parameters, or variables that learn during model training. And while Openi does not reveal the parameters, experts estimate the latest model that is at least one trillion.
In terms of performance, depressic Says Its R1 model receives the performance of Openi’s O1 on reasoning tasks, including AIME 2024, Kod Force, GPQA Diamond, Mathematics -500, MMLU and SWE Bench confirmation.
In a technical report, the company said its V3 model training costs only $ 5.6 million – a portion of billions of dollars that notable Western AI Labs such as Openi and Entropic have trained and operated their basic AI models. Spend for However, it is not yet clear how much Deep Sak has to spend.
If training costs are correct, although, this means that the model was developed in a part of the cost of rival models through Openi, Entropic, Google And the other.
Daniel Newman, CEO of the Tech Insight Firm The Photography Group, said the developments proposed “a massive progress”, though it had doubted the exact data.
“I am sure that Deep Sak’s advances indicate a meaningful infection for scaling laws and this is a real need,” he said. “Saying that, there are still many questions and uncertainty around the full cost of the cost as it is related to the development of Deep Sak.”
Meanwhile, Paul Troilio, senior VP, China and Technology Policy, in the Advisory Firm DGA Group, noted that it is difficult to compare directly between Dipic’s model costs and major US developers.
He said, “5.6 million data for Dipic V3 was for only one training, and the company emphasized that the overall cost of R&D has not been represented to prepare this model “” “After that the overall cost was significantly high, but still less than the amount of US AI companies’ large companies.”
Deep Sak was not available for comment immediately when contacted by CNBC.
To compare DPCESK, Openai to Price
Both Deep Sak and Openi disclose prices on their websites to count their models.
DPCAC says the R1 costs 55 cents in a 1 million token input – referring to each individual unit of the text processed by the “token” model – and $ 1 million token output.
In comparison, the Openi Prices page for O1 charges $ 150 in an out token of 1 million input token and $ 60 in an output token of $ 1 million. Openi’s small, low -cost language model GPT -4 O and Money, the firm receives 15 cents per 1 million input token.
Doubts on chips
The disclosure of DiPsic KR One has already led to the heat of the public debate on the truth of his claim – at least because its models were created despite exports from the United States to China, which is high to China. The use of level AI chips is banned.
DiPsic claims that its progress is using adult NVIDIA clips, which include H800 and A100 chips, which are less developed than the modern H100s of the chipmaker, which cannot be exported to China.
After that Nvidia has come out and said that the GPUSek used was fully exported.
Is the original deal or not?
Industry experts seem to agree that what DiPsic has achieved is impressive, though some have emphasized doubts about some Chinese company’s claims.
“Dipsec is legally impressive, but the level of hysteria is a crime for many people,” “American businessman Palm Loki, who founded Okolas and Andorral, wrote on X.
“The number of $ 5 million is fake. It has been pushed by the Chinese Hedge Fund to slow down investment in the US AI startups, serve its shorts against US Titans such as NVIDIA and hide the theft of approval. “
London Headquarters Startup Chief Commercial Officer, Seva Regal, who offers access to DPCAC’s AI model through a distributed GPU network, said no reason to believe them in Deep Civic. No
“Even if it is closed to a particular element, it is still very effective.” “The logic of what they have explained is very sensible.”
However, some people have claimed that Dupic technology may not have been developed from the beginning.
Billionaire investor Vinod Khosla, without giving further details, said on X, “DiPsic has made the same mistakes that make O1 mistakes, a strong indication that the technology has been eliminated.”
It is a claim that the Openi himself told CNBC in a statement that he said in a statement on Wednesday that he was reviewing reports that DiPsic had to develop his AI model For their models have used “inappropriate” output data, a method called “Oson”.
“We take aggressive, active response measures to protect our technology and will continue to work with the US government to protect the most capable models built here,” the Open spokesman told CNBC.
AI’s commodities
However, Deep See is shaking around, AI scientists are largely agreeing that this is a positive step for the industry.
Yan Likone, Chief AI scientist MethodSaid the success of Depsek represented victory for open source AI models, not necessarily behind a famous open source AI model named Lilama for China than the American meta.
“For those who look at Deep Sak’s performance and think: ‘China is surpassing the United States in AI.’ You are reading this wrong: ‘Open source model is surpassing proprietary ideas’.
“Deptic has benefited from open research and open source (such as Meta Petrich and Lama). They brought new ideas and made them at the top of other people’s work. Because their work has been published and open source. , Everyone can take advantage of it.
See: Why DiPsic is endangering USA’s AI lead

– CNBC’s Katrina Bishop and Hayden Field contributed to this report