DeepSeek: the new AI front-runner and culprit behind crypto’s sell-off

On January 20, the Chinese AI startup DeepSeek released the open-source reasoning model DeepSeek-R1. Over the weekend of January 25–26, the system drew intense attention, prompting sell-offs across equities and cryptocurrencies.

What is DeepSeek?

DeepSeek is an artificial-intelligence startup founded in 2023 in Hangzhou, China. The company specialises in building open-source large language models.

In November, DeepSeek unveiled a “super-powerful” reasoning AI model, DeepSeek-R1-Lite-Preview. According to published tests, it performs on a par with OpenAI’s o1-preview.

At the end of December, the firm showcased its own LLM V3, which outperformed rivals from Meta and OpenAI in testing.

DeepSeek V3 has 671bn parameters. For comparison, Llama 3.1 405B has 405bn. The count reflects an AI system’s ability to adapt to more complex use cases and deliver more precise answers.

The system took two months to build, cost $5.58m and used far fewer computational resources than larger tech firms, running on Nvidia H800 chips.

Thanks to the new DeepSeek-R1 model, the company’s chatbot shot to the top of the free apps chart in the US App Store, overtaking ChatGPT.

R1: the o1 killer

DeepSeek has delivered top-tier performance for far less money. In a number of third-party tests, the Chinese model outperformed Llama 3.1, GPT-4o and Claude Sonnet 3.5. Specialists examined accuracy, the ability to solve complex tasks, maths and programming.

“[…] The developers really did manage to build an open-source neural network that performs computation at inference and is extremely efficient. We must take China’s developments very, very seriously,” commented Microsoft CEO Satya Nadella at the World Economic Forum in Davos (Switzerland). 

DeepSeek also surprised by managing to sidestep US export controls.

China’s #DeepSeek could represent the biggest threat to US equity markets as the company seems to have built a groundbreaking AI model at an extremely low price and w/o having access to cutting-edge chips, calling into question the utility of the hundreds of billions worth of… pic.twitter.com/wMVyczpRgT

— Holger Zschaepitz (@Schuldensuehner) January 24, 2025

“The Chinese company DeepSeek may represent the biggest threat to US equity markets, as it appears to have built a revolutionary AI model at an extremely low price and without access to cutting-edge chips, calling into question the utility of the hundreds of billions of capital investments being poured into this industry,” commented journalist Holger Zschaepitz.

DeepSeek introduced “distilled” R1 variants ranging from 1.5bn to 70bn parameters. The smallest can run on a laptop.

In one example, DeepSeek R1 was even run on a smartphone.

Wow DeepSeek R1 version 1.5B runs perfectly locally on my phone ?

So you can have a model that outperforms GPT-4o and Claude 3.5 Sonnet on math in your pocket.

Mind-blowing pic.twitter.com/0J3uLXI6ZR

— Paul Couvert (@itsPaulAi) January 22, 2025

The larger version requires powerful hardware but is available via API at prices 90–95% below OpenAI’s o1—$0.14 per million tokens versus $7.5 at the American rival.

To achieve high performance at lower cost, the Chinese developers “rethought everything from scratch”.

A paradigm shift

Dropbox’s VP of Product, Morgan Brown, explained DeepSeek’s approach and technical choices:

“Traditional AI is like writing down every number with 32 digits after the decimal point. At DeepSeek they thought: ‘What if we use only 8? It will still be accurate enough!’ Boom—75% less memory.”

DeepSeek also applied a “multi-token” system. A standard AI “reads like a first-grader”: “The cat… sat…”. The Chinese model reads whole phrases at once, twice as fast and 90% more accurate.

“But here’s what’s really smart: they built an ‘expert system’. Instead of one massive AI trying to know everything (as if one person were a doctor, a lawyer and an engineer), they have specialised experts that wake up only when needed,” Brown noted.

In traditional models, all 1.8trn parameters are active all the time. DeepSeek has 671bn parameters, but only 37bn are active simultaneously.

“It’s like having a huge team, but engaging only the specialists you actually need for each task,” added the Dropbox VP of Product.

The results are “mind-blowing”, the expert said:

• training cost: $100m -> $5m;
• required graphics processors: 100,000 -> 2,000;
• API costs: 95% cheaper;
• can run on gaming GPU.

“But wait, you say, there must be a catch! The thing is, it’s all open source. Anyone can verify their work. The code is public. The technical papers explain everything. This isn’t magic—just incredibly smart engineering,” Brown concluded.

DeepSeek achieved these results with a team of fewer than 200 people.

R1 has a downside—censorship. As a Chinese model, it is subject to government oversight. Its answers will not touch on Tiananmen Square or Taiwan’s autonomy.

DeepSeek r1 takeaways for policy:
1. Chinese labs will likely continue to be fast followers in terms of reaching similar benchmark performance to US models.
2. The impressive performance of DeepSeek’s distilled models (smaller versions of r1) means that very capable reasoners…

— Dean W. Ball (@deanwball) January 20, 2025

“The impressive performance of DeepSeek’s distilled models means that very capable reasoning systems will continue to spread widely and run on local hardware, out of sight of any oversight,” noted Dean Ball, an AI researcher at George Mason University.

Sell-off

The rapid rise in DeepSeek’s popularity led to a sell-off in stocks and crypto. Investors grew concerned about a bubble in the AI sector.

US AI startups spend billions of dollars training models, and their valuations reach into the hundreds of billions. DeepSeek has shown this is not inevitable.

On January 27, shares of Japanese firms tied to chip production fell.

US stocks also saw steep declines. Shares of Nvidia, the chief beneficiary of the AI boom, were hit hardest.

The sell-off in TradFi spilled into crypto, with AI-linked tokens hit first. AI agents were especially hard hit.

Crypto investors are evidently digesting what DeepSeek means for the future of AI in digital assets.

On January 27, the price of the leading cryptocurrency fell below $100,000. Major altcoins dropped even more.