10 Myths About DeepSeek AI That Everyone Gets Wrong

Separating Fact from Fiction in the AI Arms Race

Is DeepSeek AI the game-changer it’s made out to be, or is it just clever marketing and strategic hype? 👀

While some hail it as a revolutionary leap in AI efficiency, others argue that its success is built on borrowed (or even stolen) innovation and questionable practices. Rumour has it DeepSeek’s CEO started hoarding Nvidia chips like they were toilet paper during a pandemic — and that’s just the tip of the iceberg.

From its alleged $5.5 million training budget to its use of OpenAI’s ChatGPT for distillation, DeepSeek’s rise has sparked intense debate. And its story is far from simple — its impact on the AI landscape is anything but settled.

Having been deep in AI research for years, I’m here to tell you: this rabbit hole goes way deeper than anyone imagines.

In this article, I will debunk 10 myths about DeepSeek AI, so grab your popcorn (and maybe a VPN?) because this is one tech drama you won’t want to miss. 🍿

BTW, if you want to learn about AI and AI investments before anyone else, check out our Be Limitless community. We’re not just talking about the future of AI — we’re actively profiting from it.

Myth #1: DeepSeek Only Cost $5.5 Million

One of the biggest myths is that DeepSeek only spent $5.5 million to train their AI model, which is peanuts compared to the $10 billion spent by OpenAI or the $7 billion spent by Anthropic.

The myth comes from DeepSeek’s December 2024 paper, which reports $5.5 million spent on training costs. However, they explicitly note: “The aforementioned costs include only the official training of DeepSeek-V3, excluding the costs associated with prior research and ablation experiments on architectures, algorithms, or data.”

According to some estimates, DeepSeek’s actual spending is likely way, way higher: between $500 million and $1.3 Billion. These figures come from the SemiAnalysis researchers that show the real GPU requirements to train an AI like DeepSeek:

• 10,000 H100 GPUs — the gold standard for AI high-performance
• 10,000 H800s GPUs — same computational power as H100s, but lower network bandwidth
• 30,000 H20s — compliant to export controls
• Total: 50,000 Nvidia GPUs

These numbers far exceed their officially reported numbers.

It is also important to add that DeepSeek has used existing AI models, such as ChatGPT, for AI distillation. In other words, it used established AIs to accelerate training more cost-effectively. While DeepSeek does employ more efficient algorithms and synthetic data, the notion of a mere $5.5 million total investment simply doesn’t add up.

Myth #2: DeepSeek Stole From OpenAI

OpenAI blocked traffic from China after detecting billions of API calls and suspicious activity — which is why I now need a VPN to use ChatGPT in Hong Kong.

Maybe “stealing” is not the most accurate term, but DeepSeek did use OpenAI’s ChaGPT as an integral part of its training process through a method called distillation.

Distillation in AI is when a larger, more complex “teacher” model transfers its knowledge to a smaller “student” model. This technique helps create more efficient models at a lower cost. It’s not inherently “stealing”, as it’s a well-documented practice in AI research used to make models more accessible and efficient.

This is why, a while ago, if you asked DeepSeek, “What AI model are you?” it would respond with, “I’m ChatGPT.”

Myth #3: DeepSeek is the Best AI Model Available

Labelling DeepSeek as the “best” AI model available is a huge oversimplification.

First, a highly censored AI model cannot be considered the “best,” as its answers must primarily comply with censorship requirements — in this case, the Chinese government’s regulations.

According to benchmarks, DeepSeek R1 scores second after OpenAI o1:

According to benchmarks, DeepSeek R1 is not the best and definitely not the fastest (if you have tried it, you know it’s quite slow). Still, considering the architectural innovations and efficiency, they’ve made some cool improvements.

DeepSeek does introduce notable improvements that set it apart:

• Chain of Thought (CoT): When you prompt DeepSeek, you can see its reasoning while it answers the question. The AI “shows its work” instead of just giving the final answers. It breaks down problems into smaller steps and explains its thinking process, making it better at solving complex problems and helping you understand how it reached its conclusions.
• Mixture of Experts (MoE): DeepSeek breaks down into smaller specialized brains instead of one huge generalist brain. MoE uses a “router” to send tasks to specialized AI experts who are best suited to handle them. This makes it more efficient than having one big system trying to do everything and using less computing power.

Myth #4: DeepSeek Cannot Generate Images

DeepSeek dropped their AI image generator, Janus Pro, in late January 2025.

Janus-Pro is like an AI artist with two brains—one for understanding pictures and text and another for drawing new pictures. This is called dual-encoder architecture. While it’s not as good as DALL-E 3 or MidJourney AI (it can only generate 384 × 384 resolution images), it is open-source, free to use, and available on Hugging Face.

Janus Pro is like an AI artist with two brains: one for understanding pictures and text, and another for drawing new pictures. This dual-encoder architecture, while its capability is not as good as MidJourney AI or DALL-E 3 (it’s limited to 384 × 384 resolution images), offers distinct advantages: it’s open-source, free to use and available on Hugging Face.

Myth #5: DeepSeek’s Success Means Nvidia’s GPU Dominance is Over

This is clearly not true, as DeepSeek had to use 50,000 Nvidia H100 and H800 GPUs to train its AI models.

DeepSeek brings efficiency, but efficiency will increase demand for AI usage, increasing the demand for computing power, including Nvidia GPUs.

Nvidia stock dropped by 17% on Jan. 27th, eliminating a staggering $600 billion in market capitalization from the company. I see this as an overreaction, but I have to admit that Nvidia’s stock does look overpriced, with or without DeepSeek efficiencies.

Myth #6: DeepSeek Uses Illegal GPUs

If the US banned the export of Nvidia GPUs to China, how did a Chinese company acquire tens of thousands of Nvidia GPUs?

DeepSeek CEO Liang Wenfeng is a visionary entrepreneur. He reportedly started buying the type of GPUs that were banned long before they were banned.

On the other hand, since DeepSeek was founded and the ban on chip exports to China started, chip sales to Singapore have multiplied by 8. It’s not likely that these chips are all being used inside the tiny country of Singapore.

Singapore, with its significant Chinese population, serves as a trade bridge between the world and China, similar to Hong Kong’s role.

I’m a free market guy, and I believe that the US and China (and the world) would benefit from not having these export bans. I do not see DeepSeek buying Nvidia chips through Singapore as a crime.

Myth #7: DeepSeek’s Success Means US Investments Were Wasted

Briefly after the DeepSeek R1 model’s release, US stocks like Nvidia crashed hard, due to fears that demand for Nvidia chips would decline, as DeepSeek introduced an AI model that is 10 times more computationally efficient.

US companies have invested billions in AI data centers and AI chips — so… are these investments now wasted?

No. Far from it. More efficient models will increase AI usage, which will drive up demand for computing power and even further investment in computer chips.

This phenomenon is known as the Jevons Paradox: technological progress or efficiency improvements in resource usage lead to increased consumption rather than decreased.

We see examples of the Jevons Paradox everywhere:

Vehicle Fuel Efficiency

Cars now use less fuel per mile, encouraging fuel-saving behavior.

👉🏻 Paradox: This often results in more driving or people buying SUVs, increasing overall fuel consumption.

LED Lighting

LEDs use significantly less electricity than incandescent bulbs.

👉🏻 Paradox: The lower cost of light led to a huge increase in lighting, increasing worldwide energy consumption.

Computing Power Efficiency

Computer chips become 2x more efficient every 2 years, according to Moore’s Law.

👉🏻 Paradox: Improving computing efficiency often results in more complex software requiring greater processing power, leading companies to expand data centers and enhance data-intensive applications.

So, the answer is no, US investments are not wasted, and the demand for AI computing power will only increase.

Myth #8: DeepSeek Models Are a Direct Challenge to OpenAI’s Capabilities, Suggesting U.S. AI is Inefficient

AI development is not done in isolation.

Every new AI model stands on the shoulders of giants. DeepSeek would not be possible without OpenAI’s efforts and the billions of dollars invested.

Chinese companies are good at copying (and improving!) American technology, and this is another example.

Myth #9: DeepSeek Has Created an Unbreakable Moat Around Itself

DeepSeek’s success is more of a catalyst for industry-wide innovation rather than a sign of an immutable lead. Competitors will continue pushing boundaries, potentially closing any gaps through their own advancements.

It’s important to note that DeepSeek is open-source, with its advancements shared with everyone.

Using the MIT open-source license means users can freely use, copy, modify, and even sell software based on DeepSeek.

Myth #10: All the Data Goes Back to the Chinese Government

China has laws like the Cybersecurity Law, Data Security Law, and Personal Information Protection Law, which demand that Chinese companies store data in a way that can be shared with the Chinese government.

However, this doesn’t happen automatically. Theoretically, the Chinese government will access the data only if there’s a legal justification.

It’s important to note that the actual implementation and enforcement of these data protection laws remains a matter of ongoing discussion and observation.

DeepSeek AI: The Final Verdict

DeepSeek AI sparked loads of controversy, but the truth lies somewhere between innovation and ingenuity.

In the end, the truth is that AI models are being commoditized, and soon we will not be able to distinguish between DeepSeek, ChatGPT, Claude, or Grok.

So, is it the future of AI or just a clever contender? The answer isn’t black and white — it’s as complex as the algorithms it’s built on. And that’s exactly why DeepSeek is here to disrupt and accelerate us toward AGI and ASI.

The AI landscape changes daily, and tomorrow’s opportunities belong to those who stay ahead today.

Are you ready to take control of your AI future? Join my Be Limitless community today and gain access to cutting-edge AI strategies, investment opportunities, and expert insights that will put you miles ahead. 🦾

— Henrique Centieiro 🕺🏻

If you found any value in this article, throw me some love!

🥰 What you can do to support my work:Clap up to 50, leave a message to share your thoughts &be sure to follow. 💌

🌞 Stay in touch: