Let me start with a COTD (Confession Of The Day): "I just canceled my OpenAI subscription." Why? Because I wanted to save £16/month - yes, but no, because of the new open-source model DeepSeek, which happened to flip the whole AI industry on its face — and it's free and open-source (if you missed it). DeepSeek is a Chinese open-source AI model that's about 90/95% cheaper than GPT-4 and just as smart (trust me, with its DeepReasoning R1 capabilities)
In all honesty I was sceptical at first, but with all the buzz around it, but in true developer style, I had to give it a try. And guess what? It's amazing! The responses are just as sharp, the reasoning more logical, and I think it's actually a better model with all that 'deep thought processing' transcript.
The DeepSeek Wall Street Panic
As you may have seen NVIDIA stocks rocketed with over 200% over the last 12 months, with their AI GPUs sales, partnerships with Cloud providers like AWS, etc. Until now, where it dropped by just over 50%. Why? DeepSeek costs just £4 million to train — compared to the hundreds of millions spent on models like GPT-4.
The DeepSeek workarounds:
- Smarter Hardware Choices: Due to US export restrictions on NVIDIA chips, the DeepSeek team turned to less expensive, older AI chips.
- Innovative Architecture: They reworked the traditional large language model architecture, removing unnecessary computations (thus many guardrails) to trim down costs.
- Mixture of Experts: Instead of relying on one giant model, they used a collection of smaller, specialized models. When you input a query, a “central brain” decides which expert is best suited to handle it.
These clever workarounds mean that DeepSeek offers performance comparable to models like GPT-4, but at a fraction of the cost. Imagine all of those £32k NVIDIA GPU Chips powering most AI data centers, now look like an overkill. And investors? They start to ask questions, and the stock market panics, meanwhile millions of users also found themselves an alternative to ChatGTP
AWS Adopts DeepSeek—Fast!
The momentum with this new model doesn't stop here with its low training cost. AWS is already all-in - as we would expect. Amazon added DeepSeek's models (R1, R1-Zero, and the vision-based Janus-Pro-7B) to Bedrock and SageMaker. Translation:
- Bedrock: Plug-and-play API access to models for users who want to play with AI.
- SageMaker: Build-Your-Own models for AI evangelists needing to tweak, train, or deploy models on their own.
- Cost Optimization Pillar: Since DeepSeek is open-source, you only pay for the AWS resources you provision (hello, Trainium and Inferentia chips!).
You can read more about the AWS DeepSeek availability here.
Turns out, OpenAI's literally just now launched their new "Deep Reasoning" model too - a pair of them "o3-mini" and "o3-mini-high". They're basically playing catch-up to DeepSeek's months-old tech - I mean who doesn't like competitive market. But here's the catch: they're still not open-source..
You can read more about the OpenAI's new "o3-mini" models here.
What's Next?
In my opinion the AI race just got a new perspective - it isn't just about the biggest financial backing - it's about who can innovate with least amount of resources. DeepSeek proved to us that doesn't need billions to build world-class AI. We need creativity, flexibility, and a willingness to break the rules.
So yeah, I canceled ChatGPT. Why pay £16 when I can tap into DeepSeek's power for free. At last we see democratization of AI — and it's about time.
If you're interested in experimenting with cutting-edge AI without breaking the bank, I highly recommend giving DeepSeek a try. Feel free to drop your thoughts on X I'm excited to see where this open source revolution takes us next!