A young upstart has the number-one spot, and we’re not talking about the Billboard Hot 100.
A free, open-source AI model recently launched by China-based startup DeepSeek is making investors take a LongLook at the companies supplying all that hardware to power today’s artificial intelligence efforts.
DeepSeek surpassed ChatGPT for the top spot in Apple’s App Store, leading to lower valuations on Monday for AI chipmakers like Nvidia and Broadcom.
“DeepSeek-R1 is AI’s Sputnik moment,” investor Marc Andreessen said in a Jan. 26 X post, referring to DeepSeek’s first reasoning model and the 20th century space race.
What is DeepSeek? Hedge-fund co-founder Liang Wenfeng created DeepSeek in 2023. The open-source DeepSeek-V3 model powering the technology relies on activating “only the most relevant portions of their model for each query, and that saves money and computation power,” Giuseppe Sette, president at AI market research firm Reflexivity, shared with CBS News.
About those costs. DeepSeek claimed in a Dec. 2024 paper that its total training costs added up to $5.576 million, assuming the cost is $2 per hour per GPU.
“That is a fraction of what Western hyperscalers are throwing at their LLM training (e.g., it’s 9% of the compute used for Meta’s Llama 3.1 405B model),” Richard Clode, portfolio manager at Janus Henderson Investors, wrote on Jan. 27.
Both Meta and Microsoft announced multi-billion dollar investment plans for AI infrastructure. President Trump recently cheered on a $500 billion joint-sector AI investment project called Stargate.
Top insights for IT pros
From cybersecurity and big data to cloud computing, IT Brew covers the latest trends shaping business tech in our 4x weekly newsletter, virtual events with industry experts, and digital guides.
The rise of DeepSeek, which reportedly will limit user registrations following recent cyberattacks, occurs weeks after the Biden administration issued restrictions on the export of US-made AI chips.
Ya get what ya pay for? A Jan. 28 report from security platform provider Enkrypt AI concluded that DeepSeek-R1 was “highly biased” and vulnerable to generating insecure code and toxic content.
Stock shake-up. DeepSeek’s claims of doing more with less hardware spurred a dip in stock value for powermakers like Constellation Energy and chipmakers like Nvidia, Broadcom, and Marvell Technology on Jan. 27.
While a one-year-old startup and its R1 features will still have to compete with other available commercial models (and stand up to scrutinizing red teamers), the idea of an effective open-source AI system has true mass appeal, according to investors like Andreessen and Jason Trennert, chairman and CEO of Strategas Research Partners. Though Trennert said on CNBC’s The Exchange on Monday, that he was “somewhat skeptical” of the low-cost model’s chances of disrupting the AI universe, he also said this: If DeepSeek does what it says, the product hurts dominant “Mag 7” tech vendors like Microsoft, Alphabet, Nvidia, and Meta, and helps, well, humanity.
“If you were able to get AI on the cheap that has the same quality, it would greatly speed up the productivity enhancements that we’re hoping to get from AI in the future,” he told CNBC.