Kunal's Blog

Why DeepSeek’s AI Innovations Are Shocking the World (and Making Nvidia Nervous)

Let’s talk about why DeepSeek’s latest breakthroughs in AI are turning heads—and why Nvidia, sitting atop a $3T market cap, might be feeling the heat.

AI Training Was Absurdly Expensive… Until Now

Right now, training the best AI models costs an eye-watering amount. Companies like OpenAI and Anthropic are spending upwards of $100M just on compute power. They need colossal data centers packed with thousands of $40K GPUs—basically, the AI equivalent of running an entire power plant just to keep one factory going.

Then DeepSeek arrived and said, “What if we did this for $5M instead?” And they didn’t just theorize—they actually pulled it off. Their models are going toe-to-toe with GPT-4 and Claude on many tasks. The AI industry is in shock.

The Game-Changing Innovations

How did they do it? By completely rethinking how AI should work. Traditional models operate like writing every number with 32 decimal places. DeepSeek asked, “What if we only used 8? It’s still precise enough!” That simple shift slashed memory needs by 75%.

Then there’s their radical “multi-token” approach. Most AI models process words like a first-grader reading aloud: “The… Apple… Ball…Cat..” DeepSeek’s model reads entire phrases at once. The result? It’s twice as fast while maintaining 90% of the accuracy. And when you’re dealing with billions of words, that’s a massive efficiency boost.

But here’s where they really flipped the script: Instead of building a single, monolithic AI that tries to know everything (like expecting one person to be a doctor, lawyer, and engineer all at once), they created an “expert system.”

Traditional models keep all 1.8 trillion parameters active all the time. DeepSeek’s model has 671B parameters in total—but only 37B are active at any given moment. Think of it like having a huge team but only calling in the experts you actually need. The efficiency gains are mind-blowing.

The Impact in Numbers

• Training cost: $100M → $5M
• GPUs needed: 100,000 → 2,000
• API costs: 95% lower
• Runs on gaming GPUs instead of high-end data center hardware

No Catch—Just Brilliant Engineering

Now, you might be thinking, “Okay, but what’s the downside?” The crazy part? There really isn’t one. DeepSeek’s work is completely open source. The code is out there. The technical papers lay everything bare. This isn’t magic—it’s just a smarter way of doing things.

And this changes everything. Until now, AI has been a game only the biggest tech companies could afford to play. If you didn’t have a billion-dollar data center, you were out of luck. But DeepSeek just rewrote the rules. Now, a handful of good GPUs might be enough.

A New Era for AI

This is a potential nightmare for Nvidia. Their entire business depends on selling ultra-expensive GPUs with fat profit margins. But if AI models can suddenly run on much cheaper, more accessible hardware, demand for those pricey GPUs could plummet. You can see why they’d be sweating.

And here’s the kicker: DeepSeek did this with fewer than 200 people. Meanwhile, Meta has entire teams where just their salaries exceed DeepSeek’s total training budget—and yet, their models aren’t as efficient.

A Classic Disruption Story

This is how industries get disrupted. Established players keep optimizing the old way of doing things, while newcomers throw out the rulebook and start fresh. DeepSeek asked, “What if we made AI smarter instead of just throwing more hardware at it?”

The implications are huge:

• AI becomes far more accessible
• Competition in the AI space skyrockets
• Big Tech’s protective “moats” start looking more like puddles
• Hardware costs and energy consumption drop dramatically

Of course, OpenAI and Anthropic won’t sit still. They’re almost certainly racing to adopt these techniques. But the efficiency genie is out of the bottle now. The old “just buy more GPUs” approach is officially outdated.

The Inflection Point

This feels like one of those moments we’ll look back on as a turning point. Like when personal computers made mainframes obsolete or when cloud computing changed everything.

AI is about to get a whole lot cheaper and a whole lot more accessible. The real question isn’t if this will shake up the industry—it’s how fast.