Yazılar

Baidu CEO Stresses Continued Investment in Cloud Infrastructure Despite DeepSeek’s Success

/in /tarafından

Baidu CEO Robin Li emphasized the ongoing need for investment in cloud infrastructure and data centers despite the recent success of DeepSeek, a Chinese AI startup. Speaking at the World Government Summit in Dubai on Tuesday, Li highlighted that while DeepSeek has demonstrated impressive efficiency in developing language models using less computing power, the demand for greater computational resources remains crucial.

“The investment in cloud infrastructure is still very much required. To develop models smarter than others, you need more compute,” Li stated, referring to the hardware that enables AI models to train, process data, and generate predictions.

DeepSeek has captured global attention for creating language models that rival leading systems like OpenAI’s GPT while utilizing significantly lower computational resources. This innovation has raised questions about the necessity for massive infrastructure investments, which has been a focal point of many AI companies’ strategies.

Baidu, one of the first Chinese companies to introduce AI products after the launch of OpenAI’s ChatGPT in late 2022, has developed its large language model, Ernie, which the company claims matches the capabilities of GPT-4. However, Ernie has seen limited adoption in the market.

Li, known for making bold statements about China’s AI industry, acknowledged the unexpected nature of innovation, citing DeepSeek’s rapid success as an example. “You just don’t know when and where innovations come from,” Li remarked.

The CEO also noted that the pressure to reduce costs and overcome compute constraints has driven Chinese companies like Baidu to innovate more efficiently. In an apparent shift from his earlier stance, Li recognized the potential of open-source approaches to accelerate AI development. “If you open things up, many people will be curious enough to try it. This will help spread the technology much faster,” he added.

Google Introduces New Class of Cheap AI Models as Cost Concerns Intensify

/in /tarafından

Google has introduced new, cost-effective AI models under its Gemini family, responding to increasing competition and concerns over the escalating costs of artificial intelligence. The new offerings, including the “Flash-Lite” model, are designed to compete with cheaper AI models like DeepSeek’s, a Chinese rival that has drawn attention for its low-cost AI training.

The company unveiled several versions of its Gemini 2.0 models, which offer varying levels of performance and pricing. Among these is the “Gemini 2.0 Flash,” which was released to the general public after being previewed to developers in December. Flash-Lite, a more affordable variant, has been developed in response to positive feedback on the earlier Flash 1.5 model. However, the cost of Gemini 2.0 Flash is higher than its predecessor.

Google’s new pricing strategy comes amid growing scrutiny from investors over the rising expenses of AI model development. Recently, DeepSeek revealed it spent just $6 million on the final training run of one of its models, prompting comparisons to the significantly higher costs cited by major U.S. AI firms, including Alphabet, Microsoft, and Meta. Despite this, DeepSeek’s low-cost model has spurred competitors to accelerate their AI spending, leading to concerns about the long-term profitability of such investments.

Pricing for Gemini Flash-Lite is competitive, with certain inputs costing as little as $0.019 per 1 million tokens. This is cheaper than OpenAI’s flagship model, which costs $0.075 per million tokens, and slightly higher than DeepSeek’s $0.014 model (though DeepSeek’s pricing will rise fivefold on February 8).

These updates reflect Alphabet’s response to the growing pressure to provide affordable AI models while maintaining a competitive edge in the rapidly evolving AI space. However, despite these advancements, investor concerns remain about the sustainability of high capital expenditures in AI development.

 

Chinese Chip Makers and Cloud Providers Rush to Support DeepSeek’s AI Models

/in /tarafından

Chinese chip makers and cloud service providers are quickly integrating DeepSeek’s artificial intelligence models into their systems, marking a significant turning point for the nation’s AI industry. Companies like Moore Threads and Hygon Information Technology, both AI chip manufacturers, announced that their computing clusters and accelerators are now compatible with DeepSeek’s R1 and V3 models. Moore Threads even shared a celebratory post on WeChat, praising DeepSeek’s ability to drive China’s AI sector forward by utilizing domestically produced graphic processing units (GPUs).

Huawei Technologies, which also develops AI chips, revealed it is collaborating with AI infrastructure startup SiliconFlow to integrate DeepSeek’s models into its Ascend cloud service. This integration has been described as a “watershed moment” by Bernstein analysts, highlighting the growing independence of China’s AI sector from advanced U.S. hardware.

Cloud giants like Alibaba, Baidu, and Tencent have also jumped on board, offering DeepSeek’s models through their respective services. DeepSeek’s AI assistant, launched last month, quickly gained popularity by offering a more data-efficient alternative at a fraction of the cost of global competitors, surpassing ChatGPT in app downloads from Apple’s App Store within days.

The company has drawn attention globally with its groundbreaking approach. DeepSeek’s research, published in December, claimed that its V3 model’s training cost less than $6 million in Nvidia’s H800 chips—significantly lower than the billions spent by companies like Meta and Microsoft. This has been a major factor in DeepSeek’s rising prominence, with its founder, Liang Wenfeng, becoming a cultural figure in China.

While Microsoft and Amazon have started offering DeepSeek’s models, some countries, including Italy and the Netherlands, have raised concerns over privacy, leading them to either block or investigate the AI app.