(Credit: DeepSeek)
Chinese AI company DeepSeek released an AI model that is sending shockwaves through the US tech industry due to its low cost and high performance.
Tech investor Marc Andreessen calls it "AI's Sputnik moment," causing US tech stocks to tumble this morning, with Nvidia the hardest hit, down 11%.
The app surpassed ChatGPT on the Apple App Store as the most-downloaded free application. But after receiving a flood of registrations and "large-scale malicious attacks"—possibly a DDoS attack—DeepSeek has "temporarily" restricted access to those with a +86 mainland Chinese phone number. You may still be able to sign in with a Google account.
(Credit: DeepSeek)We downloaded it from the app store before this, and found the interface nearly identical to ChatGPT. It currently has a knowledge cutoff of October 2023, as shown in the screenshot below. But the big game-changer is not the chatbot experience but the open-source model that powers it on the backend. It appears to be the highest-performing available for an ultra-low price. Being open source, anyone with the right skills can download it and use it.
The model, dubbed R1, came out on Jan. 20, a few months after DeepSeek released its first model. The company says its performance is "on par with OpenAI-o1,” the model OpenAI debuted in September 2024 with the ability to follow a chain of thought. The company called the approach a "new paradigm" that would bring AI models closer to human-level intelligence.
Looking at the pricing specs for the DeepSeek R1 reasoner model and OpenAI's GPT-o1, you'll see R1 is a fraction of the cost. DeepSeek charges $2.19 per 1 million output tokens, compared to $60 for GPT-o1. Even OpenAI's mini version of GPT-o1, which is intended to be a lightweight, more affordable version, costs $12.
Like DeepSeek, Meta has made open source its central AI strategy with its Llama models, which are currently free. "To people who see the performance of DeepSeek and think, 'China is surpassing the US in AI.' You are reading this wrong," says Meta's chief AI scientist, Yann LeCun, on Threads. "The correct reading is: 'Open source models are surpassing proprietary ones.'"
Meta has set up four war rooms to analyze DeepSeek’s technology, focusing on how the company reduced training costs and what data it may have used, according to The Information.
Controversial Chips, Data Collection
There is some controversy over the chips that power R1. DeepSeek founder Liang Wenfeng, a 40-year-old businessman, reportedly stocked up on Nvidia A100 chips before they were banned from export to China in September 2022. DeepSeek claims R1 runs on lower-capacity Nvidia H800 chips, but it's difficult to know exactly what is going on behind closed doors, especially since A100 chips are still being snuck into China in violation of US sanctions.
"Rather than weakening China’s AI capabilities, the sanctions appear to be driving startups like DeepSeek to innovate in ways that prioritize efficiency, resource-pooling, and collaboration," says the MIT Technology Review. The constrained chip supply may have forced the company to place a premium on efficient computing from the start of development.
More efficient computing also means less energy use—a welcome innovation given AI's astronomical energy needs.
Why has Nvidia's stock price tumbled, even though DeepSeek is still using its chips? Because its AI model does more with less. US tech companies are now coming under fire for a potentially bloated approach to AI development, requiring more and more chips, while being less efficient with them than DeepSeek appears to be. Last week, President Donald Trump announced that Oracle, OpenAI, and Softbank would spearhead a $500 billion investment in AI data centers with the Stargate Project.
Data collection is another concern with DeepSeek. Like TikTok, the company is obligated to share user data with the Chinese government if it requests it. The privacy policy lists the data it collects, and how it will use it, including to "comply with our legal obligations." As an open-source platform, that means US companies may feed it their data in addition to everyday chatbot users.