While skeptics continue to argue that artificial intelligence will never become conscious, a small Chinese company DeepSeek made an incredible breakthrough. It created a model that rivals GPT-4, but is cheaper, open, and available to everyone.
How to do more than OpenAI for $5 million?
The story of DeepSeek is a true “nerd tale.” It all started at the hedge fund High-Flyer Quant, which in 2021 foresaw that GPUs could be useful for more than just trading. After the introduction of US sanctions, the company found itself with 10,000 video cards that sat idle most of the time. Then the fund’s head Liang Wenfeng thought: “Why not try creating AI? It’s cooler than just making money.”
Typically, training AI costs hundreds of millions of dollars and requires thousands of server GPUs. DeepSeek achieved similar results with a budget of $5 million, and their API costs only $0.14 per million tokens (17 times cheaper than OpenAI).
What's the secret? Laziness and mathematics.
1. Efficient architecture: Instead of the traditional approach with millions of GPUs, their new Multi-head Latent Attention (MLA) architecture saves 90% of computations by ignoring 95% of unnecessary data.
2. Expert systems: Instead of one massive model, specialized modules operate, activated only when needed. 236 billion parameters, but only 37 billion are active at a time. It's like hiring a huge team of experts, but calling only those who are needed.
3. Hardware optimization: DeepSeek has made it so that their models can run even on gaming GPUs, not just on expensive server ones.
Result:
1. The cost of AI development has fallen by 20 times.
2. API in 95% is cheaper.
3. GPT-4-level performance, but at an affordable price.
Why is this breaking the industry?
China is breaking the monopoly of Western corporations on AI. If OpenAI and Meta could previously demand any amount for their models, now DeepSeek with its price tags makes you wonder why you should pay more.
For Nvidia this is a problem: Their business is built on selling expensive server GPUs, but if AI starts running on cheap hardware, the market will be shaken up.
American corporations are already in a panic. Some are urgently calling meetings, some are cutting prices, and some are simply pretending that nothing has happened. But everyone understands: this is just the beginning, and they will have to react urgently.
Alibaba Strikes Again: Qwen2.5
If DeepSeek showed that AI can be cheap, Alibaba also decided that it can be powerful. Their new model Qwen2.5-1M can handle up to a million context tokens. For comparison, ChatGPT Pro offers only 128 thousand, and that for $200 per month.
What can Qwen2.5-1M do?
1. Work with huge texts, databases and books without restrictions.
2. Process information 7 times faster than ChatGPT.
3. Generate not only text, but also images and videos.
4. Work for free, without subscriptions and VPN.
5. And of course, it's all open source. Want to build your own startup? Take it and use it however you want.
The Accessible AI Revolution
DeepSeek and Qwen2.5 – it’s not just about technology. It’s about how a few smart decisions can break the monopoly of giants. Now AI is no longer a toy for those with billions, but a tool available to everyone.
China is no longer catching up. Now other countries are starting to rack their brains over how to keep up in this race. And we seem to be starting to see something very interesting.
Results for the day
1. Nvidia shares fell 8.5%.
2. ASML lost 9.5%.
The crypto market was not saved either – Bitcoin fell lower $100 thousand, and altcoins fell by 10–20%.
What do analysts think?
As the market rages, major players are trying to assess DeepSeek's long-term impact:
1. Bernstein: The panic around Nvidia and Broadcom is exaggerated. Their market positions will remain strong due to high demand.
2. Citi: Yes, China has thrown down the gauntlet, but the US maintains an edge thanks to access to Nvidia's advanced chips.
3. Raymond James: DeepSeek could disrupt the market by reducing the need for huge GPU clusters, making AI model development cheaper and more accessible.
4. Cantor: On the contrary, the hype around the Chinese startup will spur interest in computing technology.
5. JPMorgan: DeepSeek's technology proves that the cost of AI development can be reduced and that investments in this area are often overvalued.
6. Jefferies: DeepSeek's cost-saving potential could become apparent as early as 2026 - prepare for a new round of competition.
DeepSeek: Too Good to Be True?
Meanwhile, service DeepSeek could not cope with the influx of users. After rapidly growing in popularity and reaching the first place in the American AppStore, the platform is experiencing interruptions in its work.
This raised questions among experts: is DeepSeek's infrastructure ready for mass use, or is this just hype with a short-term effect?
What's next?
Although DeepSeek has made waves and shaken up markets, analysts are not rushing to bury Nvidia and other giants. The startup may accelerate the development of technology, but the global GPU infrastructure is still in demand.
What do you think? Is this the beginning of a new era or just a flash that will fade away without leaving a trace?
