DeepSeek is the latest multimodal AI.
Technically, DeepSeek is the name of the Chinese company releasing the models. The current models themselves are called “R1” and “V1.” Both are massively shaking up the entire AI industry following R1’s January 20 release in the US.
Why? It comes with very low development costs, it’s open-source for commercial use, and it undercuts rivals like OpenAI — right at a time when the US government has bet more heavily than ever on its own home-grown AI advancements.
What Do I Need to Know About DeepSeek?
Deepseek offers a couple different models — R1 and V3 — in addition to an image generator. The key thing to know is that they’re cheaper, more efficient, and more freely available than the top competitors, which means that OpenAI’s ChatGPT may have lost its crown as the queen bee of AI models.
Here’s what to know about all of them.
DeepSeek R1
The DeepSeek model that everyone is using right now is R1.
This just in! View
the top business tech deals for 2025 👨💻
It’s at the top of the App Store — beating out ChatGPT — and it’s the version that is currently available on the web and open-source, with a freely available API. Unlike some other China-based models aiming to compete with ChatGPT, AI experts are impressed with the capability that R1 offers.
As influential tech investor Marc Andreessen put it a few days back: “DeepSeek R1 is one of the most amazing and impressive breakthroughs I’ve ever seen — and as open source, a profound gift to the world.”
It’s way cheaper to operate than ChatGPT, too: Possibly 20 to 50 times cheaper.
There’s some murkiness surrounding the type of chip used to train DeepSeek’s models, with some unsubstantiated claims stating that the company used A100 chips, which are currently banned from US export to China.
DeepSeek V3
However, the company’s other big model is what’s scaring Silicon Valley: DeepSeek V3.
The V3 model was cheap to train, way cheaper than many AI experts had thought possible: According to DeepSeek, training took just 2,788 thousand H800 GPU hours, which adds up to just $5.576 million, assuming a $2 per GPU per hour cost.
V3 is a more efficient model, since it operates on a 671B-parameter MoE architecture with 37B activated parameters per token — cutting down on the computational overhead required by ChatGPT and its 1.8T-parameter design.
Text-to-image generation: Janus Pro
Plus, there’s Janus Pro, the company’s text-to-image generator.
DeepSeek has reported that its Janus-Pro-7B AI model has outperformed OpenAI’s DALL-E 3 and Stability AI’s Stable Diffusion, according to a leaderboard ranking for image generation using text prompts.
According to the company, this model was trained on “72 million high-quality synthetic images.”
DeepSeek can run locally
One last thing to know: DeepSeek can be run locally, with no need for an internet connection. This is part and parcel with the model’s open-source release: Since the code is available on GitHub, it can be downloaded.
That marks another improvement over popular AI models like OpenAI, and — at least for those who chose to run the AI locally — it means that there’s no possibility of the China-based company accessing user data.
In Response, NVIDIA’s Stock Is Way, Way Down
Deepseek marks a big shakeup to the popular approach to AI tech in the US: The Chinese company’s AI models were built with a fraction of the resources, but delivered the goods and are open-source, to boot. The initial response was a big drop in stock prices for the biggest US-based AI companies.
AI chip company NVIDIA saw the biggest stock drop in its history, losing nearly $600 billion in stock-market value when stocks dropped 16.86% in response to the DeepSeek news.
How bad is that? Well, it’s more than twice as much as any other single US company has ever dropped in just one day. In other words, it’s not great.
The company’s response so far: It’s admitted that DeepSeek’s R1 model is “an excellent AI advancement.”
Other US shares are down, too
Shares dropped at other chipmakers as well, such as the Dutch company ASML, while the S&P 500 dropped more than 2% and Nasdaq fell 3.5%.
Plenty of experts are predicting that the stock market volatility will settle down soon. However, it might mark the end of an era for the constant steep upward trend for companies like NVIDIA, which had seen incredibly high growth since early 2023.
But this is good news for some tech giants
So, how does the AI landscape change if DeepSeek is America’s next top model?
Meta is likely a big winner here: The company needs cheap AI models in order to succeed, and now the next money-saving advancement is here.
Microsoft will also be saving money on data centers, while Amazon can take advantage of the newly available open source models.
Google, on the other hand, would have stood to make the most money from all those data centers. Something tells us that the massive tech giant will stay afloat, however. In the long run, cheap open-source AI is still good for tech companies in general, even if it might not be great for the US overall.
How Does the US’s $500-Billion Stargate Program Factor in?
The DeepSeek disruption comes just a few days after a big announcement from President Trump: The US government will be sinking $500 billion into “Stargate,” a joint AI venture with OpenAI, Softbank, and Oracle that aims to solidify the US as the world leader in AI.
Among the initiative’s plans are the construction of 20 data centers across the US, as well as the creation of “hundreds of thousands” of jobs, although the latter claim seems dubious, based on the outcome of similar previous claims.
With that eye-watering investment, the US government certainly seems to be throwing its weight behind a strategy of excess: Pouring billions into solving its AI problems, under the assumption that paying more than any other country will deliver better AI than any other country.
Now, DeepSeek has emerged to poke a hole in that thesis. If DeepSeek can get the same results on less than a tenth of the development budget, all those billions don’t look like such a sure bet.
The stock market — for now, at least — seems to agree.
