Training AI for Pennies on the Dollar: Are DeepSeek’s Costs Being Undersold?DeepSeek is allegedly running on a budget so low it makes OpenAI and Google look like they’ve been burning money for fun. However, we’re not seeing OpenAI, Google DeepMind, or even smaller AI startups rushing to implement DeepSeek’s methods.
China just claimed they’ve cracked the code on AI training costs, and honestly, that should make everyone a little suspicious. DeepSeek, their latest homegrown AI model, is allegedly running on a budget so low it makes OpenAI and Google look like they’ve been burning money for fun. The claim? DeepSeek trains AI models for “pennies on the dollar” compared to Western labs. Sounds amazing—if it’s true. But if there’s one thing history has taught us, it’s that China’s tech announcements should always come with an asterisk.
Remember when they said they’d built a quantum computer that left Google’s Sycamore in the dust? That turned out to be… let’s say, “optimistic.” So is DeepSeek really a breakthrough in cost efficiency, or are we just looking at another example of creative accounting?
The Math Problem
Let’s start with the numbers. DeepSeek’s researchers claim they’re slashing AI training costs by as much as 90%. That’s not a small improvement—it’s a seismic shift. If true, every AI company in the world should be knocking on their door, desperate to learn their secret. But there’s a problem: AI training is expensive for fundamental reasons.
It requires massive GPU clusters, insane amounts of electricity, and constant fine-tuning by highly paid engineers. NVIDIA isn’t handing out A100s in cereal boxes, and power companies don’t do charity work. So how, exactly, is DeepSeek running a large-scale LLM on a budget that sounds like it came from a middle school science fair?
One theory is that they’ve found an ultra-efficient training method that Western labs somehow missed. Another, slightly more realistic theory? They’re cutting corners. Maybe they’re using lower-quality data, skipping expensive fine-tuning, or just outright exaggerating their cost savings. It wouldn’t be the first time a Chinese tech company made big claims that didn’t hold up under scrutiny (like Evergrande). And given how secretive DeepSeek’s research has been, it’s not like we can just peek at their training logs to double-check.
The Hardware Problem
AI training isn’t just about really smart AI algorithms, it’s also about raw computing power. All the top labs in the world use thousands of high-end GPUs to crunch numbers in order to train their models, and those things aren’t cheap, no matter how rich you are. China, however, is currently dealing with major restrictions on advanced chips thanks to U.S. export bans. That means they don’t have easy access to the latest NVIDIA hardware, which raises a huge question: What are they even using to train DeepSeek?
If they’ve somehow managed to train a GPT-4 competitor using outdated chips or homegrown alternatives, that’s either the biggest efficiency breakthrough in AI history or proof that they’re just throwing out numbers without receipts. China does have some AI chips of its own, like Huawei’s Ascend series, but these aren’t exactly known for outperforming NVIDIA’s best. In fact, if you look online, it says they’re basically using outdated NVIDIA chips purchased legally in 2023. So either they’ve made a hardware leap nobody saw coming, or their cost savings come at the expense of quality. Wouldn’t be surprising if DeepSeek turns out to be less of a “ChatGPT killer” and more of a chatbot that doesn’t retain any information from past conversations.
The Catch?
So the big question here again would obviously be that if DeepSeek really does train AI models at a fraction of the cost, why isn’t everyone else copying them already? To be honest, while there is a lot of black box technology that no one fully understands, AI research isn’t some shadowy, closed-off world and major breakthroughs spread pretty fast.
Yet, we’re not seeing OpenAI, Google DeepMind, or even smaller AI startups rushing to implement DeepSeek’s methods. That suggests one of two things: either DeepSeek’s cost-saving approach isn’t as good as advertised, or it comes with major trade-offs.
Maybe their AI models are simply weaker. Training AI cheaply often means making compromises—less data, fewer refinements, lower precision. That’s fine if you’re making a chatbot for customer service, but not if you’re aiming for the kind of high-end AI that can rival OpenAI or Anthropic. Another possibility? The numbers are just being cooked.
China has a long track record of overstating its technological achievements, whether it’s GDP figures, scientific breakthroughs, or military capabilities. If DeepSeek’s cost savings don’t hold up under third-party evaluation, it wouldn’t be the first time a flashy tech announcement turned out to be a mirage.
In case you missed:
- Nvidia Project GROOT for humanoid robots
- Mainstream AI workloads too resource-hungry? Try Hala Point, Intel’s largest Neuromorphic computer
- What’s Nvidia doing in the restaurant business?
- NVIDIA just dropped “ACE” at CES 2025: Truly intelligent NPCs coming soon!
- CES 2025: NVIDIA’s Cosmos Just Gave Robots a ‘ChatGPT Moment’!
- These AI powered devices add smells to virtual worlds
- Meet Einstein Copilot, Salesforce’s new conversational AI
- This computer uses human brain cells and runs on Dopamine!
- How AI Is Helping Restore the World’s Coral Reefs
- Samsung’s new Android XR Headset all set to crush Apple’s Vision Pro
