The software manufacturer Snowflake $SNOW (+0.39%) decided on Monday to include DeepSeek's models in its marketplace for AI models after receiving a number of customer requests.
As employees also described the models from
DeepSeek's models as "amazing", the US software provider weighed up the
the potential risks of hosting AI technology developed in China before finally deci
before finally deciding to offer it to its customers, said
Christian Kleinerman, executive vice president of product at Snowflake.
"We decided that
we don't see any problems in supporting this technology as long as we are clear
clear to our customers," he said.
In the meantime, the
American AI developers are rushing to analyze DeepSeek's V3 model.
DeepSeek published a research paper in December on the model, which
the basis for its popular app, but many questions such as the total development cost
are not answered in the document.
China is now from 18 months to
six months behind the most advanced AI models developed in the US, a person said.
developed in the US, one person said. Since DeepSeek has caused such a stir with its strategy of
free release strategy is causing such a stir, the company may soon run out of
may soon run out of chips to meet demand, this person predicted.
person predicted.
DeepSeek's progress
is not only based on a modest budget of 6 million dollars - a
a tiny sum compared to the 250 billion dollars that analysts estimate large
that analysts estimate large US cloud companies will spend on AI infrastructure this year.
will spend this year. The study points out that these costs relate
specifically refer to chip usage for the final training run and not to the entire
to the total development costs.
The training run is just the tip of the iceberg
of the iceberg in terms of overall costs, executives from two top labs told Reuters.
top labs told Reuters. The cost of determining the design of the
training run can be many times higher, they said.
The paper states that the
training run for V3 was conducted using 2,048 H800 chips from Nvidia, which were
designed to comply with US export controls enacted in 2022
rules that experts told Reuters would do little to slow China's AI progress.
would hardly slow it down.
Sources at two AI labs said,
they had expected that earlier stages of development would have relied on a much larger
amount of chips. One of the people said that such an
such an investment could have cost more than 1 billion dollars.
Some American
AI executives praised DeepSeek's decision to release its models as open source, which
source, meaning that other companies or individuals can freely use or modify them.
individuals can freely use or modify them.
"DeepSeek R1 is one of the most
most amazing and impressive achievements I have ever seen, and as open source
- and a great gift to the world as open source," said venture capitalist
venture capitalist Marc Andreessen in an article on X am Sonntag.
The acclaim that the models from
DeepSeek's models underlines the viability of open source
open-source AI technology as an alternative to costly and tightly controlled
controlled technologies such as OpenAI's ChatGPT, according to industry observers.
The most valuable companies on
Wall Street have risen in recent years on the expectation that only they have
have access to the enormous capital and computing power needed to develop and scale new
required to develop and scale new AI technologies. These
assumptions will be further tested this week and next when many
when many American tech giants report their quarterly results.
will present their quarterly results.
