The AI Cost Revolution: How a Small Company Disrupted the Industry's Economics

Thank you for being part of AI enthusiast community, where you receive weekly insights on Machine Learning & AI.

Enjoy the content? Support the newsletter by sharing it with your friends! They can join the newsletter for free via LinkedIn or at RiteshAI.com - just a note, at some point, this newsletter will be renamed to AI Guru.

Disclaimer: I work for Cerebras Systems but opinions in this article are my own.

Today, I want to dive into something that's creating waves in the AI industry - a breakthrough that could fundamentally change how we think about AI costs.

What Just Happened?

With recent news of DeepSeek, a Chinese company that's barely two years old with just 200 employees, has achieved something remarkable: they've created an AI model that performs similarly to or better than OpenAI's latest models, but at a fraction of the cost. The impact? Nearly $600 billion wiped off NVIDIA's market cap in a single blow.

Why This Matters

For those of you new to AI, let me break this down. When you use ChatGPT or any AI service, you're interacting with a trained model. Think of training like education - it takes time, resources, and lots of energy. Traditionally, training these models has required:

• Hundreds of thousands of GPUs
• Around $100 million in costs
• Months of computation time
• Massive infrastructure

DeepSeek's Breakthrough

Here's what makes DeepSeek's achievement extraordinary:

• They spent only $5M instead of $100M
• Used thousands rather than hundreds of thousands of GPUs
• Achieved similar or better performance than leading models

The Secret Sauce: Dual Innovation

1. Training Innovation (GRPO) DeepSeek's novel approach, called Group Relative Policy Optimization (GRPO), is fascinating. Imagine a study group where students learn not just from their own work, but by comparing notes and approaches with others. GRPO does something similar for AI models, creating a more efficient learning environment.
2. Inference Innovation What's equally impressive is how DeepSeek optimized their model's operational efficiency:

• Expert System: Out of 671B total parameters, only 37B are active at any time
• Multi-Token Reading: Processing text in chunks instead of word-by-word, doubling speed
• Precision Optimization: Reduced from 32 to 8 decimal places, cutting memory use by 75%

These optimizations translate to dramatic cost savings:

• API costs up to 90% less per million tokens
• Input costs of $0.14 (vs. competitor's $30.00)
• Output costs of $0.28 (vs. competitor's $60.00)

Historical Context: The Pattern of Disruption

We've seen this pattern before:

• Netflix disrupted Blockbuster by leveraging internet infrastructure
• Amazon transformed book retail, challenging traditional bookstores
• SpaceX and ISRO showed space exploration doesn't need billion-dollar budgets (ISRO's Mars mission cost just $75M compared to the usual $1B+)
• Now, DeepSeek might be doing the same for AI

Open Questions

However, some crucial questions remain:

1. How much research groundwork was needed before this breakthrough?
2. In their GRPO implementation, are they using other LLMs (like GPT-4 or Claude) as judges? If so, what are those hidden costs?
3. Can this approach scale beyond their current implementation?

Looking Forward

This breakthrough could mean:

• More accessible AI development for smaller companies
• Faster innovation cycles
• More competition in the AI space
• Potentially lower costs for AI services

My Take

While there are still uncertainties, I believe this represents a significant shift in AI economics. Just as ISRO showed us space exploration could be done for a fraction of the traditional cost, DeepSeek is showing us that advanced AI development might not need the massive resources we once thought necessary.

I'd love to hear your thoughts on this. How do you think this might affect your work or industry? Feel free to reply to this email with your perspectives.

Shameless Plug:

Outside of my day job, I try to share my AI knowledge and wisdom via live and on-demand training courses. Here are some ways through which you can leverage my learning:

-----

Starting Your AI Journey:

-------

?? Want to start free on learning chatGPT? https://lnkd.in/ejSxHx45

?? Looking to build your AI skills but don't know where to start? Perfect course: AI for Everyone (https://lnkd.in/eSSct3Eb)

?? How about mastering Claude AI? I go in very detail - https://lnkd.in/eh56Zbvg

Looking to train entire team? https://www.aiguru.in/corporate-training

-------

Role-Specific Coursera Programs:

-------

?? DevOps Professionals: https://lnkd.in/eEzm9e9n

?? Program Managers: https://lnkd.in/eQPuNXtm

?? Sales Professionals: https://lnkd.in/ee2z6wkc

?? Marketing Professionals: https://lnkd.in/e8hPRv6R

?? Telemedicine Data Security Professionals?: https://lnkd.in/eEp4PqBp

?? Risk Management Professionals? https://www.coursera.org/learn/genai-for-risk-management-essentials

?? Creating Sales Proposals? https://www.coursera.org/learn/genai-for-sales-proposals-and-presentations

-------

?? Prefer Interactive Learning? Join our Quarterly Live Sessions:

-------

1) AI for Leaders: https://lnkd.in/edehm85X

2) AI for Engineers: https://lnkd.in/edAu9t88