Chinese startup DeepSeek’s open-source LLM was first introduced in December, and a reasoning model called RI was released last week that outperformed OpenAI’s latest model in third-party tests.
As a result, DeepSeek’s advancements absolutely slayed Nvidia and its US AI accelerator competitors on Monday, sending Nvidia’s shares down 17%.
Nvidia and the entire sector had their worst day Monday since March 2020, partly because of algorithmic trading that could correct in coming days. In addition to Nvidia’s dive of 17%, Broadcom dropped17% or more, while Micron dropped 12% and Arm by 10%. AMD dropped 6%. DeepSeek is privately held and not publicly traded in the US.
Many analysts were still positive on AI investments and future prospects, however.
“I don’t think we can ever have enough compute,” said Stacy Rasgon, senior analyst at Bernstein Research on CNBC. “I still think we’ll need lots of chips.” That includes Nvidia’s latest Blackwell GPUs, which are sold mainly for use in servers in the largest data centers.
Some analysts surmised that the reduced share price for Nvidia shows that the market has now shifted to support for using accelerators for inference rather than for training where Nvidia has traditionally excelled. Nvidia issued a statement indicating otherwise, saying that inference requires “significant number of Nvidia GPUs and high-performance networking.”
The full statement from Nvidia also praised DeepSeek: “DeepSeek is an excellent AI advancement and a perfect example of Test Time Scaling. DeepSeek’s work illustrates how new models can be created using that technique, leveraging widely-available models and compute that is fully export control complaint. Inference requires significant number of Nvidia GPUs and high-performance networking. We now have three scaling laws: pre-training and post-training, which continue, and new test-time scaling.”
Patrick Moorhead, lead analyst at Moor Insights & Strategy, picked up that theme. “Few are talking about the incredible potential downstream benefits of model and method optimization related to DeepSeek,” he wrote on X. “Think about how much more AI this potentially puts in the hands of consumer and businesses via software, SaaS, and end compute hardware companies. Oh, and the data inference demand curve doesn’t decline; it could speed up.”
He said DeepSeek has implemented more efficiency in training and inference with less data and less memory required. And because it offers its model open source, developers can leverage it. It will be a boon to enterprise software and SaaS and data center and data center edge inference, he added. “Devices get a lot more interesting if I can run what were cloud-only reasoning capabilities right on my phone or PC,” he wrote.
Jack Gold, lead analyst at J. Gold Associates, urged investors not to panic, and said Nvidia will weather the Monday downturn “just fine” partly because the company has added software and services to their portfolio. “I don’t see the hyperscalers pulling back, which is Nvidia’s biggest market,” he said via email to Fierce Electronics.
Gold also wondered whether DeepSeek’s claims are accurate. “Until we see further proof and others replicating their stuff, I’d take the announcements with a grain of salt,” he said. Gold also has written extensively on the move to inference and inference at the edge over the next two years. Those needs “certainly don’t require the massive scale out of Blackwell chips and will run on lower power, more modest processors.” Those are areas where AMD, Intel and ARM do well. “If the DeepSeek stuff is real and the need for longer processing power is accurate, then the proprietary chips at the hyhperscalers will be a huge boon to their business.”
Gartner VP Analyst Chirag Dekate told Fierce Network that DeepSeek R1 uses memory more efficiently by tapping into an FP 8 data format rather than FP16, thereby consuming less memory. Memory compression and key value caching is also used. He said the model relies on what’s called ‘mixture of experts’ concept as well.
RELATED; DeepSeek didn’t burst the AI bubble—it made it bigger
Politics arise related to DeepSeek’s China connection
Because of DeepSeek’s connection to China, its impact on US companies came up in the political realm.
House Speaker Mike Johnson in a news briefing on Monday alluded to DeepSeek as another means for China to gain an advantage over the US in AI. “ China is a terrible trading partner,” he said. “They abuse the system.They steal our intellectual property. They are now trying to get a leg up on us in AI as you’ve seen in the last day or so.”
Asked about Johnson’s comments, Moorhead said in a CBNC interview that US attempts to control China’s use of AI are complex. Because DeepSeek is an open source model, “people will find a way around” restrictions, he said.