If building an AI model from scratch is like baking a cake, then distillation is like reverse-engineering a top bakery’s best-selling recipe.
Think of a world-class pastry chef who spends years perfecting a cake recipe using the finest ingredients, experimenting with different methods, and finally creating something truly exceptional. Then, along comes a talented home baker who doesn’t have access to those rare ingredients but is clever enough to taste the cake, break down its flavors, and recreate it using simpler, more affordable ingredients—without sacrificing much of the taste.
That’s exactly what’s happening in AI today. The below NVIDIA chart catched eye for everyone on Jan 27, 2025.
A small Chinese startup, DeepSeek, has shown just how disruptive this can be. Using distillation, they managed to create an AI model that rivals OpenAI’s most advanced systems—in just two months, for less than $6 million. That’s a fraction of what OpenAI or Google would spend on similar technology.
This shift is reshaping the AI industry.
- Does this mean Big Tech is losing its edge?
- Is DeepSeek actually innovating, or just cutting costs?
- How will this affect the price and accessibility of AI for businesses and developers?
To understand where AI is heading, we first need to understand how distillation works—and why it’s such a game-changer.
The Power of Distillation: A Shortcut to AI Supremacy
AI development has traditionally been an arms race. The biggest players—OpenAI, Google, and Meta—compete to build larger and more sophisticated models, pouring billions into compute power, research, and fine-tuning. But distillation changes the rules of the game.
How Does It Work?
Instead of spending years training a brand-new model from scratch, distillation allows a smaller AI model to “learn” from a larger one by mimicking its responses. This works in three steps:
1. Big Tech builds a giant AI model. OpenAI, Google, and Meta pour millions into training state-of-the-art AI models using vast amounts of computing power.
Source of Pic: https://www.youtube.com/watch?v=BzgUOKFrHcA&ab_channel=CNBC
2. Smaller players “distill” that model. Instead of starting from scratch, smaller teams take a powerful AI model and train a smaller version by mimicking its responses
3. The smaller model learns to think like the big one. Through a process called knowledge distillation, the smaller AI absorbs the reasoning of the larger model, allowing it to function almost as well while using far less computing power
Source of Pic: https://www.youtube.com/watch?v=BzgUOKFrHcA&ab_channel=CNBC
This means that, for a fraction of the cost, a distilled model can deliver nearly the same performance.
Why Is This a Game-Changer?
Instead of spending years training a brand-new model from scratch, distillation allows a smaller AI model to “learn” from a larger one by mimicking its responses. This works in three steps:
- AI is no longer just for billion-dollar companies. Any startup or research team can now create a powerful AI model without breaking the bank.
- The cost of using AI is dropping fast. Businesses that rely on AI no longer need to pay sky-high prices to OpenAI or Google.
- More competition means better AI for everyone. With more players in the game, expect AI models to improve faster while becoming more affordable.
This isn’t just about DeepSeek—it’s about a new era of AI where expensive models can be distilled, copied, and improved upon. That changes everything.
3. DeepSeek’s Breakthrough: More Than Just Economics?
At first glance, it’s easy to assume that DeepSeek’s success is just about cost efficiency—a clever way to cut corners and still get great results. But is it really just about saving money, or is there real innovation behind what they’re doing?
The Numbers Don’t Lie
DeepSeek built a model that rivaled OpenAI’s cutting-edge AI in just two months, spending only $6 million. Meanwhile, OpenAI’s latest advancements likely cost them hundreds of millions in compute power, research, and engineering efforts.
But if distillation was the only factor, why haven’t Microsoft or OpenAI done the same thing? After all, these tech giants have access to massive AI infrastructure, talented researchers, and unlimited computing power.
Did DeepSeek Introduce Real Innovation?
Distillation alone doesn’t explain everything. There are a few key areas where DeepSeek seems to have done more than just copy existing models:
- Optimizing Distillation Itself
- DeepSeek didn’t just distill an AI model; they refined the process to make it even more efficient.
- They combined distillation with reinforcement learning to enhance reasoning abilities.
- Faster Training with Less Compute
- Researchers at Berkeley and Stanford demonstrated that they could train high-quality AI models in hours, using as little as $50 in compute power.
- DeepSeek applied similar efficiencies to scale its own models.
- Beyond Distillation: Tweaks to Existing Models
- DeepSeek took an older AI model (Quinn, from Alibaba) and enhanced its reasoning skills using both distillation and fine-tuned training techniques.
The result? A model that outperformed far more expensive alternatives.
Innovation or Just Smart Strategy?
It’s clear that DeepSeek isn’t just copying OpenAI—it’s pushing the limits of what distillation can do. But at the same time, their success wouldn’t be possible without the groundwork laid by AI giants like OpenAI, Google, and Meta.
Instead of reinventing the wheel, DeepSeek took the best models available and made them cheaper, faster, and more efficient. That’s not just good business—it’s a wake-up call for the entire AI industry.
4. The Open-Source Shift: The AI Moat is Shrinking
For years, AI companies like OpenAI and Google have guarded their models like a secret family recipe—keeping their tech locked behind closed doors to maintain a competitive advantage.
But thanks to distillation, those walls are starting to crumble.
Distillation Means No Model is Truly Private Anymore
Once a model is publicly released, someone can distill it into a competitive alternative. Even if OpenAI or Google refuse to open-source their models, researchers can still:
- Interact with them,
- Collect their outputs,
- And train their own versions using those responses.
Ali Ghodsi, the CEO of Databricks, put it:
“You might think you haven’t open-sourced your model, but you actually have.”
The Rise of Open-Source AI
DeepSeek took things a step further by open-sourcing its V3 and R1 models, allowing anyone to use and modify them.
- Hugging Face, a major AI platform, is also helping accelerate open-source AI by hosting thousands of freely available models.
- University researchers are now training advanced models for as little as $450, proving that AI development is no longer limited to Big Tech.
Even OpenAI’s Sam Altman, who spent years defending a closed-source approach, recently admitted:
“We have been on the wrong side of history here and need to figure out a different open-source strategy.”
That’s a major shift in thinking—one that could change the AI industry forever.
Why Open-Source Always Wins
History has shown that open-source technology often outpaces closed, proprietary systems. Here’s why:
- Faster Innovation
- When AI research is open, thousands of developers can improve it—not just a single company.
- Lower Costs
- Businesses no longer have to rely on expensive AI APIs from OpenAI, Google, or Meta.
- More Competition
- Startups and independent researchers can now build cutting-edge AI models without needing billion-dollar budgets.
The Big Shift: AI Costs Are Plummeting
For businesses, this means the cost of running AI applications is dropping fast.
- DeepSeek R1 costs $2.19 per million tokens, while OpenAI’s equivalent model costs $60.
- That’s not just a small difference—it’s an order-of-magnitude drop in pricing.
For AI startups, this changes everything. A year ago, companies building AI-powered products had no choice but to pay OpenAI’s high fees. Now, they can choose cheaper, open-source alternatives—or even train their own models.
Where Does This Leave OpenAI and Google?
Companies like OpenAI and Google spent years building walled gardens around their AI models, believing that:
- The cost of training AI would keep competitors away.
- Proprietary models would always be the best.
- Open-source AI would never catch up.
But DeepSeek’s success proves otherwise. With distillation and open-source collaboration, AI is becoming a commodity—and Big Tech’s monopoly on AI might not last much longer.
5. The Economic Disruption: AI Costs Are Plummeting
For the past few years, AI companies have been locked in a race to build bigger, better models—spending billions on compute, talent, and data. But DeepSeek’s success is forcing everyone to ask a tough question:
If you can distill a powerful AI model for a fraction of the cost, why would you pay more?
The Cost of AI is Crashing
Not long ago, if you wanted to build an AI-powered app, you had to pay premium prices for access to cutting-edge models. That meant companies had little choice but to rely on OpenAI, Google, or Anthropic, even if the costs were high.
But now, with model distillation and open-source AI, the script has flipped.
- DeepSeek R1 costs just $2.19 per million tokens.
- OpenAI’s equivalent model costs $60.
- That’s nearly a 30x difference.
For businesses, this is a game-changer. Imagine running a company that relies on AI-powered chatbots, automation, or data analysis. If you could cut your AI costs by 90% or more, that would make a massive impact on your bottom line.
AI Startups Are No Longer at the Mercy of Big Tech
Just a year ago, many AI-powered startups were totally dependent on OpenAI, Google, or other major players. They were often called “ChatGPT wrappers”—meaning their entire business depended on API access to these models.
But with cheaper alternatives like DeepSeek, the balance of power is shifting:
- Startups can train their own distilled models instead of relying on expensive APIs.
- The cost of launching an AI-powered business is dropping from millions to thousands of dollars.
- AI isn’t just for Silicon Valley giants anymore—it’s for everyone.
Will OpenAI and Google Be Forced to Lower Prices?
The market is changing fast. If OpenAI wants to keep its dominance, it will have to:
- Lower its prices to stay competitive.
- Improve its models to justify the higher costs.
- Offer exclusive features that open-source alternatives can’t easily copy.
The early signs of this are already here. Right after DeepSeek launched its R1 model, OpenAI made one of its models free for consumers. That’s no coincidence—it’s a direct response to the new competition.
The Future of AI in a Distilled World
DeepSeek has proven that AI doesn’t have to be expensive. Distillation has democratized AI development, allowing smaller teams to compete with billion-dollar companies. The rise of open-source AI is forcing industry leaders to rethink their strategies.
This shift raises some big questions about the future:
- Will all AI models eventually be open-source?
- Will AI companies survive by offering unique features, rather than just powerful models?
- How will businesses adapt to this new AI economy?
Not Sure, but one thing is clear: the cost of AI will never be the same again.
For startups, developers, and businesses, this is an exciting time. AI is becoming cheaper, faster, and more accessible than ever. And thanks to distillation, the days of paying a fortune for AI might soon be over.
The AI industry has entered a new era—one where speed, efficiency, and accessibility matter just as much as raw power. The companies that adapt will thrive. The ones that don’t? They’ll be distilled and conquered.
Source: https://www.youtube.com/watch?v=BzgUOKFrHcA&ab_channel=CNBC
