$6M AI Disrupts the $100M Club

Author

Anshuman Pandey
January 31, 2025

This week's biggest news has been DeepSeek, the China-headquartered AI company that has launched its own version of an interface, like ChatGPT. Here’s a quick list of things we know so far:

Not born yesterday

Though it’s gained popularity now, DeepSeek has been launching models since November 2023. Last spring, it launched V2, which performed better than US-based companies’ LLMs, costing cheaper to run.

So, what now?

In January 2024, DeepSeek launched R1, a ‘reasoning’ model, which means it fact-checks itself and is generally more reliable. The company claims its performance matches OpenAI’s o1 model standards. We’ll have to watch and watch.

Why is it a big deal?

There are dozens of large and small models, so why is it a big deal? The primary reason is that it is open source.

Open-source models: While businesses like OpenAI keep their models proprietary, DeepSeek has adopted an open-source approach similar to Meta’s Llama.

Cheaper to build: Apparently, it cost only $6m to train, where as Sam Altman indicated they spent $100m on GPT-4.

Cheaper to use: DeepSeek R1 is said to be 20-50x cheaper to use than OpenAI’s o1 model.

“To people who see the performance of DeepSeek and think: ‘China is surpassing the U.S. in AI.’ You are reading this wrong. The correct reading is: ‘Open-source models are surpassing proprietary ones.’”

In essence, the widespread adoption of DeepSeek can fundamentally change the AI landscape. This means that:

  • Enterprises won’t pay for the model itself but focus on adapting, fine-tuning and maintaining it

  • Customized AI solutions and service providers will gain more value from the AI industry than model-makers

  • Startups with proprietary models have to rethink their competitive strategy

  • Chipmakers have to invest more in research and development as costs continue to fall

So, it’s no wonder that the Nasdaq fell 3% on 27 January and Nvidia (the most popular chipmaker for AI applications) lost $589B in market capitalization.

How’s it going?

Silicon Valley seems to have taken to it enthusiastically. Microsoft added DeepSeek R1 to Azure AI Foundry and GitHub. Amazon and Perplexity AI have done so, too. It has also become the most downloaded free app on the Apple App Store, dethroning ChatGPT!

What’s the catch?

Well, it’s been hardly a week since the model was released, so it has yet to be rigorously tested by users and third-party experts. However, some challenges are already emerging.

Censorship: DeepSeek’s training is limited by what the Chinese government allows. It can only provide responses that “embody core socialist agenda.” That can matter to some of us.

Security: Recently, DeepSeek had to pause new user registrations owing to large-scale malicious attacks. A million log lines and some secret keys are believed to have been exposed.

Privacy: In reality, DeepSeek collects pretty much the same information that most chatbots do—name, email ID, IP address, etc. Concerns have risen because the company stores its data in China, which ranks third among the most cyber crime-prone countries in the world.

Should we try DeepSeek?

The interest is definitely there. Goldman Sachs is interested in trying DeepSeek. If a bank is willing to experiment, it should be okay, I guess.

Anyway, in a controlled environment with minimal data access and strong security protocols, you might want to validate DeepSeek yourself. When you choose the use cases wisely, the cost advantages can be significant.

Wanna give it a go? Hit reply, and we’ll set it up for you.

P.S. DeepSeek R1 is now available on Tune Chat