Meta’s AI leader claims that the hype surrounding DeepSeek, a new Chinese AI competitor, is unjustified, despite Silicon Valley’s outrage.
Last week, DeepSeek alarmed US AI businesses by releasing a model that beat models from OpenAI, Meta, and other top developers on third-party benchmarks. It claimed to have made a lot less money and used subpar chips to do so.
According to Bernstein Research, DeepSeek’s current reasoning model, R1, costs $0.55 for every 1 million tokens entered, whereas OpenAI’s o1 reasoning model costs $15 for the same quantity of tokens. This indicates that DeepSeek’s models are priced much lower than OpenAI’s comparable models. Tokens are the smallest data units that an AI model can process.
The news caused a tech sell-off that wiped $1 trillion in market capitalization when it hit the markets on Monday. The company that makes high-end semiconductors that may cost at least $30,000, Nvidia, lost about $600 billion.
But according to Yann LeCun, Facebook AI Research’s chief AI scientist, there is a “major misunderstanding” over the intended application of the hundreds of billions of dollars spent on AI. LeCun said in a Threads post that the massive investments being made in US AI firms were mostly required for inference rather than AI training.
AI models apply their training knowledge to fresh data through a process called inference. Popular generative AI chatbots like ChatGPT react to user inquiries in this way. Processing costs rise as more inference is needed in response to user requests.
LeCun predicted that the cost of inference will increase as AI tools advance. “Once you put video understanding, reasoning, large-scale memory, and other capabilities in AI systems, inference costs are going to increase,” LeCun stated. He added, “So, the market reactions to DeepSeek are woefully unjustified.”
Thomas Sohmers, a founder of Positron, a hardware startup for transformer model inference, told that he agreed with LeCun that a greater portion of the price of AI infrastructure would be attributed to inference.
He predicted that the demand for inference and the infrastructure needed to support it would grow quickly. You’re missing the bigger picture when you look at DeepSeek’s training cost improvements and fail to understand that they will drastically increase inference demand, cost, and spend.
Accordingly, it is anticipated that DeepSeek will process more queries and invest a large amount of resources on inference as its popularity increases.
With the goal of making output creation simpler, an increasing number of firms are joining the AI inference sector. Some people in the AI business believe that the cost of inference will eventually decrease because there are so many providers.
However, only small-scale inference systems are affected by this. Inference costs are likely to be significantly higher for models such as DeepSeek V3, which offer free answers to a wide user base, according to Wharton professor Ethan Mollick.
As Mollick noted on X in May, “Frontier model AI inference is only costly at the scale of large-scale free B2C services (like customer service bots).” For internal business purposes, such as delivering a preliminary draft of an analysis or action items following a meeting, the cost of a query is frequently low.
Prominent tech companies have increased their spending in AI infrastructure throughout the last two weeks.
As the firm expands its own AI infrastructure, Meta CEO Mark Zuckerberg revealed more than $60 billion in projected capital expenditures until 2025. Zuckerberg stated in a Threads post that the business will be “growing our AI teams significantly” and had “the capital to continue investing in the years ahead.” How much of that would be allocated to inference was not specified by him.
President Donald Trump revealed last week that Stargate, a partnership of OpenAI, Oracle, and SoftBank, will invest up to $500 billion in AI infrastructure across the United States.