AI startups eagerly looking to adopt DeepSeek’s low-cost large language models are facing a rocky transition, Business Insider reports.
Despite the Chinese AI company’s recent surge into mainstream markets with models touted as affordable and powerful alternatives to Western options, US-based businesses are encountering significant hurdles, including slow speeds, inconsistent cloud access, and security concerns.
Startups like Counterforce Health have encountered difficulties finding reliable access to DeepSeek models through API providers. Counterforce’s CEO, Neal Shah, shared that his company has cycled through six providers before finding relative stability with Fireworks AI. Other providers were either too slow or unreliable.
Artificial Analysis, a platform monitoring AI model availability, reported that seven cloud providers were running DeepSeek models this week, but most operated at a fraction of DeepSeek’s original speed. Fireworks AI was a rare exception, delivering performance at about half the speed of the Chinese company’s native service.
Many US companies are wary of routing data through a China-hosted API and prefer using domestic providers. However, some US-based cloud services are struggling to provide consistent and fast access to DeepSeek models, largely due to a lack of adequate computing power and the surge in demand.
Even for companies that manage to secure reliable access, concerns persist over data security. Pukar Hamal, CEO of Security Pal, voiced apprehensions about using Chinese AI models, citing potential liabilities when dealing with enterprise clients.
“An enterprise wants to know your exact data architecture. If they see heavy reliance on a Chinese-made large language model, that’s going to be a problem,” Hamal said.
He added that DeepSeek’s popularity may wane in a few weeks.
Some companies, like Counterforce, are mitigating security risks by anonymizing user information before sending data to models, regardless of whether they use DeepSeek or Western alternatives.
“We don’t fully trust any model provider. You never really know where the data is going,” Shah remarked.
Despite the challenges, DeepSeek’s price advantage remains appealing for many startups. Shah noted that his company’s AI-enabled insurance tool can handle slight delays, but future applications requiring real-time interactions will require faster and more reliable service.
Several cloud providers, including Hyperbolic and Fireworks AI, are racing to improve access to DeepSeek models. Hyperbolic CEO Jasper Zhang reported a 150% increase in inference users after launching the new DeepSeek model, while Fireworks saw a 400% month-over-month spike in new users in January.
Together AI, another cloud service, said it is developing a fix to enhance speed in the coming days. Zhang highlighted the rising competition among open-source models and emphasized the need to democratize access to AI technology.
