Silicon Valley is melting down over DeepSeek, an emerging Chinese competitor in the AI landscape, but Meta’s AI chief says the hysteria is unwarranted.
DeepSeek caused alarm among US AI companies when it released a model last week that, on third-party benchmarks, outperformed ones from OpenAI, Meta and other leading developers. It did so with subpar chips and, it said, vastly less money.
Data from Bernstein Research shows that DeepSeek priced its models 20 to 40 times cheaper than equivalent models from OpenAI. Its latest reasoning model, R1, costs $0.55 for every million tokens inputted, while OpenAI’s reasoning model, o1, charges $15 for the same number of tokens. A token is the smallest unit of data that an AI model processes.
The news hit the markets Monday, triggering a tech sell-off that wiped out $1 trillion in market cap. Chipmaker Nvidia — known for its premium chips, which can cost at least $30,000 — saw its market value cut by almost $600 billion.
Yann LeCun, chief AI scientist for Facebook AI Research, however, says there is a “major misunderstanding” about how the hundreds of billions of dollars invested in AI will be used. In a Threads post, LeCun said the huge sums of money going into US AI companies are needed primarily for inference, not training AI.
Inference is the process in which AI models apply their training knowledge to new data. It’s how popular generative AI chatbots like ChatGPT respond to user requests. So the more user requests, the more inference is required, and processing costs increase.
LeCun said that as AI tools become more sophisticated, the cost of inference will rise. “Once you put video understanding, reasoning, large-scale memory, and other capabilities in AI systems, inference costs are going to increase,” LeCun said. “So, the market reactions to DeepSeek are woefully unjustified.”
Thomas Sohmers, founder and CTO of Positron, a hardware startup for transformer model inference, told Business Insider he agreed with LeCun that inference would account for a larger share of the AI infrastructure costs.
“Inference demand and the infrastructure spend for it is going to rise rapidly,” he said. “Everyone looking at DeepSeek’s training cost improvements and not seeing that is going to insanely drive inference demand, cost, and spend is missing the forest for the trees.”
This means that, as its popularity grows, DeepSeek is expected to handle an increased volume of requests and so will have to spend a significant amount on inference.
A growing number of startups are entering the AI inference market, aiming to simplify output generation. With so many providers, some in the AI industry expect the cost of inference to drop eventually.
However, this only applies to systems handling inference on a small scale. Wharton professor Ethan Mollick said that for models like DeepSeek V3, which provide free answers to a large user base, inference costs are likely to be much higher.”Frontier model AI inference is only expensive at the scale of large-scale free B2C services (like customer service bots),” Mollick wrote on X in May. “For internal business use, like giving action items after a meeting or providing a first draft of an analysis, the cost of a query is often extremely cheap.”
In the last two weeks, leading tech firms have stepped up their investments in AI infrastructure.
Meta CEO Mark Zuckerberg announced over $60 billion in planned capital expenditures for 2025 as the company ramps up its own AI infrastructure. In a post on Threads, Zuckerberg said the company would be “growing our AI teams significantly” and has “the capital to continue investing in the years ahead.” He did not say how much of that would be devoted to inference.
Last week, President Donald Trump also announced Stargate, a joint venture between OpenAI, Oracle, and SoftBank that will funnel up to $500 billion in AI infrastructure across the United States.
The post Meta’s chief AI scientist says market reaction to DeepSeek was ‘woefully unjustified.’ Here’s why. appeared first on Business Insider.
