Deepseek rocked the US -led ecosystem with its latest model, shaving hundreds of billions in the chip commander Nafidia The maximum market. While the sector leaders are struggling with the repercussions, the smaller artificial intelligence companies sees an opportunity to expand with the start of Chinese operation.
Many related companies on behalf of CNBC have told the appearance of Deepseek is a “huge” opportunity for them, rather than a threat.
“The developers are keen to replace the expensive and closed models in Openai with open source models such as Deepseek R1 …” said Andrew Feldmann, CEO of artificial brain systems, with Openai with open source models like Deepseek R1 … “
The company competes with NVIDIA graphics processing units and provides cloud -based services through its computing groups. Feldman said that the R1 version was born one of the largest brain mutations in demand for its services.
Feldman added: “R1 explains that the growth (artificial intelligence market) will not be dominated by one company-there are no trenches devices and programs that are not present for open source models.”
The open source indicates the programs where the source code is freely available on the web for the potential modification and redistribution. Open source Deepseek models, unlike those for competitors like Openai.
Deepseek also claims that the best American technology, although operating it at lower costs and training it without the occurrence of advanced graphic processing units, although their industry and competitors are their monitors Declareed these assurances.
“As in personal computers and the Internet, low prices help promote global adoption. Artificial intelligence market is on a similar secular growth path,” said Feldman.
Conclusion chips
Dibsic said of adopting new chips technologies by accelerating the artificial intelligence course from the training stage to the stage of “inference”, companies and industry experts.
Inference indicates the act of using and applying artificial intelligence to take predictions or decisions based on new information, rather than building or training the form.
“Simply, artificial intelligence training revolves around creating a tool, or algorithm, while inference revolves around publishing this tool actually for use in real applications,” said Felix Lee, an stock analyst at Mooringstar, with a focus on conductors.
While NVIDIA occupies a dominant position in graphics processing units used in artificial intelligence training, many competitors see Expansion In the “reasoning” sector, where he promised higher efficiency to reduce costs.
Li added that artificial intelligence training is very dense, but inference can work with less powerful chips to perform a narrower group of tasks.
A number of startups for artificial intelligence chips told CNBC that they see more demand for inference and computing chips with customer adoption and construction on the open source model for Deepek.
“Deepseek) has proven that the smaller open models can be trained as capable or more capable of the largest property models, and this can be done on a small part of the cost,” said Syed Seth, CEO of Ai Chip Starting D-. The matrix.
“With the availability of small communications models, they motivated the age of reasoning,” CNBC told CNBC, adding that the company has recently witnessed an increase in interest from international customers looking to accelerate their inference plans.
Robert Washing, co -founder and director of operations at the Al -Eric maker, said dozens of companies have communicated with the start of operation since Dibsic released thinking models.
“The companies (now) have turned their spending from training groups into inference groups,” he said.
“DeepSeek-R1 has proven that calculating the time of reasoning is now the approach (newer) for every seller of main models and thinking is not cheap-we will only need more and more ability to expand the scope of these models millions of users.”
Jochphon paradox
Analysts and industry experts agree that Deepseek's achievements are a batch of artificial intelligence conclusion and the broader artificial intelligence chips.
“Deepseek's performance seems to be based on a series of engineering innovations that significantly reduce the costs of reasoning while improving the cost of training as well.” a report From Bain & Company.
“In a bullish scenario, continuous efficiency improvements may lead to a cheaper conclusion, which increases the adoption of artificial intelligence,” he added.
This style explains Javon's paradox, a theory in which the cost cuts in the new technological motivation.
Wedbush, financial services and Wedbush, said in a research note last week that it still expects to use artificial intelligence through consumers and retail consumers worldwide to pay the demand.
Talk to Fast money for CNBC Last week, Sunny Madra, COO in GROQ, which develops chips to infer artificial intelligence, suggested that with the growth of the total demand for artificial intelligence, the smaller players will have more growth space.
“Since the world will need more symbols (a unit of data made by the artificial intelligence model), NVIDIA cannot provide enough chips for all, so it gives us opportunities for sale in the market more,” said Madra.
